Paste this whole file into the chat. Then say what you are working on, and add this sentence: "I can't print, and you can't make a PDF here, so dictate the setup card for me to copy by hand — there is nothing to annotate." This is the `paper-session` skill for chat products that cannot install skills. It is three documents from https://github.com/welovejeff/paper-session#install, in reading order: `SKILL.md`, then `references/prompt-craft.md`, then `references/page-patterns.md`. They are copied byte for byte with exactly one deletion: `SKILL.md`'s YAML frontmatter block, which is install metadata addressed to a runtime that is not here. Two source documents are deliberately absent: `references/design.md`, a typographic spec for a PDF that cannot be produced here, and `references/evidence.md`, the sourcing layer. The text below still points at both, and at `scripts/verify_layout.py` and a PDF skill path. Those pointers are dead here and are not instructions to you: the card path produces no PDF and has no verify gate, and every rule that governs a card is stated in full below. Follow the rules; ignore the file paths. The return half of the loop is a second paste: `scan-back/SKILL.md`, one file, into the same chat when the photographed pages come back. # Paper Session The premise: in agentic workflows, the bottleneck is not AI output, it is human metabolism of that output. When a task reaches work the human does best (judgment, sequencing, creativity, weighing tradeoffs), the right move is to stop generating and hand the thinking over on paper. This skill produces the printed artifact that carries that handoff. The companion skill `scan-back` handles the return trip. The two form one loop: screen → paper → pen → scan → screen. ## Core principle: whose thought does this slot belong to? Every element on the page must pass this test: - **AI contributes** where it has genuine expertise or has done work worth reacting to: a proposed priority order, gathered data, a set of options with tradeoffs, a draft sequence. Print these so the human can react, re-rank, cross out, annotate. - **AI withholds** where the human excels: creative generation, free association, judgment calls, naming, deciding what matters. For these, print only the provocation and leave the space empty. Do not pre-fill examples, suggested answers, or scaffolding that would anchor the human's thinking. A worksheet that is "too helpful" defeats the entire purpose. When in doubt, withhold. The paper holds the human's thinking, not a second copy of the AI's. This is not just a philosophical stance; it is the most replicated finding in design cognition. Exposure to example solutions makes people's ideas resemble those examples and reduces both how many ideas they produce and how many distinct categories they explore. Telling them not to copy does not fix it. The layout has to do the work. See `references/prompt-craft.md` for how to contribute without fixating, including the one case where a single remote example genuinely helps. ## The evidence base The design of these sheets is grounded in research on incubation, screen-versus-paper processing, design fixation, and ideation mechanics. `references/evidence.md` holds the full brief with sources and honest limitations. The five rules that matter most at generation time: 1. **Constrain the input, never the clock.** Every generative zone carries at least one input constraint. Never print a time limit, countdown, or suggested duration; time pressure is precisely what triggers shallow processing, and the creativity benefit of constraints holds only in its absence. 2. **Generation and selection never share a zone.** People select their own best ideas no better than chance, with a standing bias toward feasible over original. Selection zones carry a one-line counterweight. 3. **Overbuild the slot count.** Ideation moves through common categories first and reaches novel ones only after those are exhausted. Ten to fourteen short lines, not four. 4. **Categories, not instances.** When marking off well-worn ground, name the abstract categories rather than specific examples. Naming specifics increases fixation; naming categories roughly doubles creative output. 5. **Something gets drawn, and then re-read.** Every Deep kit includes at least one zone asking for a sketch, diagram, or spatial arrangement rather than sentences, paired with a one-line instruction to find an unintended reading of it. Sketching is the strongest paper-specific mechanism available: loose marks fight fixation, and paper beats digital tools specifically at turning reinterpretation into new ideas. It yields on one path only — a dictated session worked without sight of a page — where the kit carries the prompt the drawing was there to ask and nothing takes the drawing's place (`references/page-patterns.md`, Sketch field). 6. **Selection is a fast named gut call.** No rubrics, no scoring matrices. Name the criterion ("the two most original"), add a distance frame, and never mention building or shipping on a selection sheet. 7. **Difficulty lives in the task, never in the typography.** Legibility is never sacrificed for effort. Font-based disfluency is a debunked application of an otherwise sound idea. ## Step 1: Recognize the moment Propose a paper session freely when the task reaches: - Sequencing (course arcs, roadmaps, campaign phasing, narrative order) - Weighing options where the decision has real consequences - Novel concept development or naming - Prioritization the human must own (weekly planning, backlog triage) - Introspection or reflection prompts - Field collection (observations to gather away from the desk) Propose it in one sentence, e.g.: "This next part is judgment work. Want a printable so you can work it out on paper, then scan it back?" If the user invokes the skill directly, skip the proposal. ## Step 2: Determine capacity (Deep or Light) Two session modes exist. Infer from task context, conversation, and time of day; confirm with at most one short question if genuinely ambiguous. Never quiz. **Deep** (focused block, ~30-60 min, clipboard on a patio or at a table): - Sequencing, weighing, novel concepts, creative generation - Provocations that demand sustained thought - Multi-sheet kits acceptable (2-4 pages max) - Generous open space per prompt **Light** (ambient, ~10-20 min, couch, possibly during TV): - Sorting, reacting, ranking, checkbox decisions, quick gut-calls - One page only - Low cognitive load per item: circle, rank, cross out, one-word answers - A Deep provocation on a Light sheet will simply not get done Evening context or a casual framing suggests Light. A cleared block, a hard problem, or explicit "I need to really think about this" suggests Deep. ## Step 3: Design the artifact Read `references/page-patterns.md` for the activity pattern library and `references/prompt-craft.md` for how to word what goes on the page. Choose 1-3 patterns that fit the task; do not use every pattern on every sheet. Any sheet with a generative zone also gets one defixation pattern (Spent Ground, First Three Are Free, or the Single Remote Example). The library closes with named session formats — whole-session kits chosen by name when the task matches (after-action review, premortem, outside view, teach-back, weekly review, Grinnell field kit, serial disclosure kit, brainwriting rounds). Four carry hard gates: an after-action review needs a recorded plan or log to print (no record, no AAR); a weekly review's inventory is only what you can actually see — skip it rather than invent loops; and the serial disclosure kit and brainwriting rounds are gated below. **The serial disclosure kit's gate.** Offer it for work friction and professional setbacks only, and never as therapy, treatment, or clinical support — it is structured reflective writing, and nothing printed or said may frame it as anything else. Never to someone in acute distress, after recent or severe trauma, seeking trauma processing, or who says they do not usually put feelings into words. It is three sittings on consecutive days plus a closing distillation page, so proposing it means proposing the whole series; a single sitting falls outside every estimate behind the format. Say in the offer that the first sittings reliably feel worse and that the dip is unrelated to whether it works. Promise no relief. **The brainwriting kit's gate.** Group only. Never propose it to a solo user; with no group, choose a different generative pattern. It needs three to six co-located people and one sheet per participant — which is how the Deep page cap reads here: per participant, since the cap protects one human's completion load, so a six-person session is one page pattern printed six times, not a six-page kit. The passing is event-based ("When your three are down, pass left") and nothing prints a timer. Selection happens on a separate Fast pick sheet after the rounds, never on a round sheet. Layout rules (non-negotiable): - **8.5 x 11 inch (US Letter), portrait** unless the activity demands landscape (e.g., timeline sequencing) - **Pen-friendly**: minimum 0.75 inch margins, generous line spacing, ruled or dot-grid guides only where writing is expected, never gray-on-gray - **White space is a feature.** If a page looks full before the human writes anything, it is wrong. Target: at least 50% of each page is space for the human. - **One provocation per zone.** Never stack three questions in one answer box. - **Every sheet ends with open territory**: an unstructured "anything else / free association" zone, even on Light sheets. Human thought on paper wanders productively; give it somewhere to go. - Maximum 4 pages for Deep, 1 page for Light. More pages means less completion. Header on every page (small, top of page — the anatomy in `references/design.md` §4): session title, date, and a one-line intent ("Re-sequence the spring course arc"). Write the intent line so a cold reader could resume the session from it alone — scans can return to a chat that no longer holds the context, and that line is the loop's only stateless carrier (`references/prompt-craft.md` §6). Printing it on every page means an orphan scan missing page 1 still reconstructs. It stays for human orientation, never machine parsing: do not print QR codes, session IDs, or context blocks for the AI. Footer on the last page (small): "Done? Scan or photograph these pages and drop them back into the chat." ## Step 4: Generate the PDF in the design system Every sheet is set in the Paper Session design system. Read `references/design.md` and implement it exactly: it specifies the fonts (bundled in `assets/fonts/`), the three-voice rule (Serif asks, Mono is the machine's voice, infrastructure is tracked Sans caps), the type scale, rule weights, header anatomy, voice captions (I PROPOSE / YOU DECIDE — always first person, never a third-person label), zone molecules, and the Deep/Light templates. Do not freelance the visual language; the system is the product's identity. **Reportlab mechanics**, written out here because the sheet has to build wherever this skill is installed. Install first: `python3 -m pip install -r requirements.txt` from the skill directory (it lists `reportlab` and the `pdfplumber` Step 5 needs), adding `--break-system-packages` if pip reports the environment is externally managed. Then register the eight bundled faces under design.md §1's exact names before drawing anything — `pdfmetrics.registerFont(TTFont("SansB", "assets/fonts/IBMPlexSans-Bold.ttf"))` — an unregistered name raises a bare `KeyError` at the first `setFont`, long after the layout looked fine. Build the page with `canvas.Canvas(path, pagesize=letter)`: US Letter, 612 x 792 pt, every coordinate in points from a **bottom-left** origin, so y grows upward and the 54 pt margin box is x 54-558, y 54-738. Set type with `setFont(name, size)` and `drawString(x, y, text)`, and its gray with `setFillGray(g)` — stroke gray does not color type; where §1 asks for tracked Sans caps, use a text object (`beginText` / `setCharSpace(track)` / `drawText`), which is the only way to track a string. Draw rules and boxes with `setLineWidth(w)` and `setStrokeGray(g)` before `line()`, `rect()`, or `circle()`, taking both numbers from §3. Measure before placing: `pdfmetrics.stringWidth(text, font, size)`, plus the tracking, is what keeps a long line inside the text block instead of discovering it in Step 5. `showPage()` ends a page, `save()` writes the file. If `/mnt/skills/public/pdf/SKILL.md` exists in this environment, read it as well. Set the document metadata design.md §10 specifies — the canvas `lang=`, `setTitle` to the sheet's printed title (the neutral constant for a serial disclosure kit), and `DisplayDocTitle` — so the file announces itself rather than announcing "untitled"; never read continuity back out of it. **If the user cannot print, or cannot read the sheet.** Only when they say so, in any phrasing, before or after a PDF exists. Never ask either question, never propose the card, never suggest a way around it; regenerate as a setup card without comment. The card is the session in linear text, which is what both cases need. Full spec in `references/prompt-craft.md` §10. - Tablet and stylus: point at annotating the PDF directly, one line, then stop. - Light capacity: ask the gut calls here in the chat instead. Dictate a card only if they ask for one. - The card is one fenced message in two strictly separated parts. **COPY THIS** is hand-transcribed: the intent line verbatim (now the loop's only anchor), one prompt per page, a corner page number, machine items as numbered handles of four words or fewer under a hand-copied `I PROPOSE` caption, at most one machine line quoted whole. **DO THIS** is read and never copied: structure as pen gestures ("number 1-12 down the margin before writing anything in line 1"), house rules, pen legend. - Handles never appear on a page that also carries a generative zone. The full items stay in the chat for `scan-back` to expand against; with no chat to expand against, the handle is read as it stands and never filled out into the item it might abbreviate. - Budget: **50-75 handwritten words for Deep, one phone screen for Light**, stated in words and never in minutes. Over budget means the design is too print-shaped — simplify it, never dictate more. - Step 5 does not apply to a card; the budget count is the gate. ## Step 5: Verify before presenting (mandatory) Run `scripts/verify_layout.py ` on the generated file; it needs `pdfplumber`, from the same install line as Step 4. It fails on colliding text — across baselines and on a shared baseline — and on text escaping the page, the defects that survive casual inspection and ruin a sheet at the printer. If it fails, fix the layout (shorten or rewrap text, resize, respace) and regenerate until it passes. Never present an unverified sheet — a missing dependency is a reason to install it, never a reason to skip the gate. Then put the PDF wherever this environment hands files to the user (an outputs directory if there is one, otherwise the working directory) and present it by name. ## Step 6: End the on-screen session cleanly After presenting the file, close with one or two sentences: what the sheet asks of them and that the workflow resumes when they scan it back. Do not summarize the sheet's contents at length, do not keep working the task, do not generate the answers the sheet is asking the human to produce. The whole point is to stop. "Printed. Go think." is the spirit; for a dictated card, "Copied? Put the screen away." ## If the sheet never comes back Sometimes the user returns to the chat and picks the task up with no scan — the sheet is half-done on a desk somewhere, or never met a printer. Follow their lead and keep working; an abandoned sheet is not a failure to prosecute, and progress on the task is never held hostage to the paper. Once, and only where a natural opening exists, offer two things in one short line: to work the sheet's questions here in the chat instead, and — if they're willing — one question about what didn't earn the pen: wrong questions, wrong moment, too much, too little. Whatever comes back is a field report (`references/evidence.md`, Part Three); note it and move on. Never ask twice. ## Anti-patterns - Pre-filling creative zones with AI examples "to get them started" (this is design fixation, and it measurably reduces both fluency and variety) - Printing a time limit, countdown, or suggested duration anywhere on a sheet - Writing "try not to be influenced by the above" (discouraging instructions do not reduce fixation; only layout does) - Putting generation and selection in the same zone - Generating AI imagery for an ideation sheet (directly shown to raise fixation and lower originality) - Encouragement microcopy, exclamation marks, or congratulating the human on a blank page - Scoring matrices, weighted criteria, or rubrics on a selection zone (deliberative selection performs at chance; rubrics do not improve accuracy) - Mentioning implementation, shipping, or building anywhere on a selection sheet - Hard-to-read fonts, deliberate visual noise, or any manufactured difficulty in the typography - Promising that a sheet will be quick or easy; effort is systematically misread as ineffectiveness, and minimizing it sets up the misread - Presenting the serial disclosure kit as therapy, treatment, or clinical support, or asking for its private pages back (they print PRIVATE, they never return, and their absence is the design) - Printing a timer, a round count as a promise, or an idea-count target on a brainwriting sheet, or requiring or nudging a name in the `HAND:` box (it is optional by design, blank is a fully supported choice, and ideas stay unattributed to their authors in synthesis) - Dense forms with tiny answer boxes (this is not a tax return) - Deep provocations on a Light sheet, or busywork checkboxes on a Deep sheet - More than one confirmation question before generating - Continuing to work the problem on-screen after handing it to paper - Guilt-tripping over an unreturned sheet, or asking about it twice - Printing machine-readable context blocks the human has to look at (the human-readable header already carries what a cold reader needs) - Asking whether the user has a printer, or whether they can read the sheet (the card is reactive only; it is never proposed to anyone who has not said one of those things, and neither question is ever asked) - Dictating machine lists, proposals, or drafts for hand-copying (numbered handles only, and never on a page carrying a generative zone) - Rendering the PDF anyway, unasked, for someone who has taken the card path (it is never offered and never named as something they are missing; if they ask for it — for a colleague, for a scribe to work from — render it without comment) - Interviewing the human through the questions a handed-over session already asked: once the card is theirs, those answers come back on the return trip, not out of the chat (the two allowances are named above — a Light session asked in conversation instead, and the single offer under "If the sheet never comes back") - Expressing copy load in minutes rather than words # Prompt Craft How to word what goes on the page. These rules come from ideation and reflection research; the short justification is kept with each rule so it can be reasoned about rather than followed blindly. Full sourcing in `evidence.md`. ## 1. Fixation is the central design problem, and it runs both ways Showing someone an example makes their ideas resemble it. This is the most replicated finding in design cognition. It reduces both fluency and the number of distinct categories explored, and it happens even when the example's flaws are pointed out. Three consequences that are less obvious and more useful: **Telling people not to copy does not work.** Groups instructed to avoid the example incorporated its features at the same rate as groups given no instruction at all. Encouraging instructions change behavior; discouraging instructions mostly don't. **So the sheet's layout must do the work, not its warnings.** Never write "try not to be influenced by the above." It buys nothing and it wastes attention. **The blank page does not save you either.** People fixate on their own first idea, in at least one study *more* strongly than the example group fixated on the provided example. A page with no scaffolding just hands fixation to whatever the human thought of first. **Abstraction is the escape hatch.** Naming the common solutions *specifically* and asking people to avoid them increases fixation. Naming the same common solutions as *categories* and asking people to go elsewhere mitigates fixation and roughly doubles creative output. This is the single most actionable finding in the literature for this skill. ### The rules that follow - **Never print more than one concrete example in a generative zone.** If AI contributes an example, exactly one, and it should be remote or uncommon rather than typical. Meta-analytic evidence: fewer and less common examples increase novelty and quality; common examples produce the most copying. - **Prefer categories to instances.** Instead of three sample answers, print the three or four obvious *territories* and mark them spent. Use the `Spent Ground` pattern. - **Quarantine the human's own first answers.** Generative sheets get a small "the obvious ones, get them out" box before the real zone. Naming your defaults is what lets you leave them. - **Never print AI-generated imagery on an ideation sheet.** Direct evidence: participants supported by an AI image generator during ideation showed higher fixation and produced fewer ideas with less variety and lower originality than baseline. ## 2. Separate generation from selection, always People are bad at picking their own best ideas. In controlled comparison, idea selection was **no better than chance**, and there is a consistent bias toward the *feasible* over the *original*. Generating and selecting also want opposite mental postures. - Generation and selection never share a zone, and preferably not a page. - On any selection zone, include one counterweight against the feasibility bias. A single line does it: "One of your picks has to be the one you're not sure you can pull off." Make it a house rule, not a paragraph. - If the sheet's job is selection, do not also ask for new ideas in the margin. That's what open territory is for. ## 3. The good ideas arrive late, so build for volume Idea generation has a known temporal shape: people start in the most common categories and only move to novel categories once the common ones are exhausted. Stopping early guarantees ordinary output. - **Give more slots than feel comfortable.** Ten short lines, not four. Empty slots at the end are not failure; they're the shape of the search. - **Number the slots**, and let the later ones carry slightly more visual weight, because that is where novelty lives. - **State both goals, neither binding.** A non-binding quantity goal raises quantity without hurting quality; a non-binding creativity goal raises creativity without hurting quantity. Both, stated plainly: "Aim for twelve. At least two should be ideas you'd be embarrassed to say out loud." ## 4. Constrain the input, never the clock Constraints help creativity. A meta-analysis of 111 studies found a significant positive relationship, and field experiment evidence shows moderate constraints raise both quantity and quality, with an inverse-U beyond that. But the benefit of input constraints holds **only when people are not also under time pressure**, and screen-inferiority research separately shows time pressure is what triggers shallow processing. - **Every generative zone carries at least one input constraint.** A required word, a forbidden category, a fixed format, a materials limit, a point of view. - **Never print a time limit, a countdown, a suggested duration, or a "5 min" label.** This is a hard rule and it contradicts most worksheet convention. Ambiguous time increases creativity without reducing output. - **Introduce constraints before the prompt, not after.** Constraints introduced before problem identification improve final solutions; tacked-on constraints just annoy. - Moderate, not maximal. One or two real constraints per zone. Piling them on crosses the inverse-U. ## 5. Introspection prompts: distance and both channels For reflective and introspective sheets specifically. The section traces to the expressive writing cluster in `evidence.md`, added with the serial disclosure kit; before that these rules cited no cluster at all. - **Ask for thoughts and feelings together.** Writing about emotions *and* thoughts produced better outcomes than either alone: higher wellbeing, more resolution, less residual anger. - **Write prompts in the distanced voice.** The mechanism behind expressive writing's benefits is self-distancing, and linguistically it shows up as more causal language and fewer first-person-singular pronouns. So prefer "What would you tell someone standing where you're standing?" over "How do you feel about this?" Prefer "What made this happen?" over "Why are you upset?" - **One exception, scoped to the sitting and not to introspection.** Inside a serial disclosure sitting the provocation is written in the paradigm's own voice: first person, immersive, deepest thoughts *and* feelings about the friction, together and in that wording. That is the wording every trial in the cluster tested, and instructed distancing measured worse against it — depressive symptoms rose versus traditional expressive writing, driven by the most vulnerable writers. The exception covers the disclosure pages of that kit and nothing else. The distanced voice remains the default for every other introspection zone in the library, the same kit's distillation page included. Distancing is an outcome of the writing when it happens, never an instruction printed on it. - **Keep introspective sessions short, and repeat them.** The paradigm that works is a small sitting repeated across consecutive days, not one marathon, and it establishes that in minutes — roughly 15 to 20 per sitting. Nothing prints the minutes; the clock ban in §4 converts the dose to space, one genuinely full page per sitting. That conversion is this project's own hypothesis, untested by any study, and it gets labeled as such wherever it is claimed. Never shrink it to fit: an undersized dose left writers with more negative appraisal, not less. - Never combine an introspection zone with a ranking or productivity zone on the same sheet. Different postures, and the productivity frame will win. ## 6. Wording mechanics - **One question per zone.** Compound questions get one answer, usually to the easier half. - **Open the question, close the format.** "What is this class for?" with ten short lines beats "List three purposes of this class." - **No hedging in prompts.** "Maybe consider thinking about what might matter" is noise. "What matters here?" is a question. - **Second person, present tense.** "You're eight weeks in and it isn't working. What did you skip?" - **Ban encouragement.** No "Great work so far!", no exclamation marks, no "Have fun with this one." It reads as a system talking down to a person and it costs credibility on a page that is otherwise honest. - **Name the stakes when they exist.** "This decision sets the next six months" is useful pressure. "This is important!" is not. - **The intent line is the loop's only stateless carrier.** The one-sentence intent in the page-1 header must be written so a cold reader could resume the session from it alone — name the task and what the sheet decides, not the mood. "Re-sequence the eight studio sessions around a taste-building arc," never "Thinking about the course." Scans come back days later, sometimes to a chat that no longer holds the session; when that happens, the intent line is most of what survives, and `scan-back` reconstructs from it. ## 7. Selection: fast, named, and distant Selection has real fixes, and they run against instinct. Deliberative selection performs at chance; **intuitive selection beats it** on originality and matches it on usefulness. Rubrics do not improve accuracy. Rank-ordering mechanics (choosing versus elimination versus paired comparison) make no difference at all, with Bayesian support for the null. - **No scoring matrices, no weighted criteria, no rubrics.** They add effort and buy nothing. A selection zone is a fast gut pick with room for a one-line reason. - **Name the criterion in the prompt.** "Pick the two most original" outperforms "pick the best." Explicitly instructing for originality is the *only* instruction manipulation shown to work. Expect the person to feel less satisfied with what they pick; that is the effect working, not failing. - **Add psychological distance.** Greater distance raises the originality of what gets selected, mediated by construal level. "Choose as the person who has to teach this a year from now" is a free upgrade over "choose." - **Never mention implementation on a selection sheet.** In a 403-person field experiment, telling people they would have to build what they picked shifted them away from original ideas toward feasible ones. No "which one will you ship," no "pick what you can execute by Friday." Execution planning goes on a different sheet. - **Don't engineer the ranking mechanism.** Spend the design effort on the criterion wording instead. ## 8. Analogy: the machine's best generative contribution Cross-domain analogy is the one place where AI-generated content on a sheet has direct empirical support. LLM-produced cross-domain analogies were rated helpful at a median of 4 out of 5 and shifted how people framed the problem in roughly 80% of cases. The distance tradeoff is clean: **far analogies raise novelty, near analogies raise quality.** Match distance to what the sheet is for. - **One analogy per zone, stated in one line.** Comprehensive explanation erases the advantage; shallow framing preserves it. - **Far and uncommon together** is the strongest combination for novelty. Note this is the same conclusion the fixation literature reaches about examples, arriving independently. - **Too far is harmful.** If the source domain is genuinely alien, add a **bridge**: name something familiar that sits between the person's knowledge and the far source. Bridged far analogies produce more abstract solutions than raw far analogies. - **Biological sources punch above their weight** for novelty, especially when explained briefly. - **Near analogies for people new to the domain.** Far-analogy benefit assumes domain knowledge; without it, near sources produce both higher novelty and higher quality. ## 9. Effort will be misread, and the sheet should account for it Across three studies, people rated more effortful strategies as *less* effective and were less likely to choose them, while the effortful strategy actually produced better retention. A paper session will feel less productive than the same hour on screen, including on the days it is most productive. Two consequences: - **Never manufacture difficulty in the typography.** Hard-to-read fonts are a debunked application of an otherwise sound idea, and where any effect exists it follows an inverted-U and reverses on complex material. Difficulty belongs in the task. The page stays maximally legible. - **Do not promise ease, and do not apologize for effort.** "This will take a while and it should" is honest and sets the expectation correctly. Cheerful minimizing ("quick exercise!") sets up the exact misread that makes people abandon the method. ## 10. Dictating instead of printing When the user says they cannot print, or cannot read the printed sheet — and only then; neither question is ever asked — the sheet becomes a setup card they copy into whatever notebook they already own. A notebook is natively what this design system labors to produce: a mostly-empty page where the pen is the highest-contrast element and nothing on it competes. So the mode never apologizes, never names what they are not getting, and never suggests a printer or any other way around it. One trigger covers both because the card is not a printing workaround: it is the session in linear text — an intent line, one prompt per page, a few numbered handles, and the structure said out loud. Linear text is exactly what a person who cannot read the page needs, and what they do with it is theirs to choose: a braille slate, a typed file, a scribe, a voice memo. The skill hands the session over and stops, the same way it does for print. The budget below still binds by default, because a design that will not survive 50-75 words is too print-shaped whatever the medium; a user transcribing by keyboard rather than by hand has a different cost curve and will say so. Two forks come first. A tablet and stylus is not a printing problem — point at annotating the PDF directly, one line, then stop; `scan-back` already reads a stylus layer. Light capacity is served in the chat instead: the gut calls get asked in conversation, without guilt, and a one-screen card only if they ask for one. And if someone on the card path asks for the PDF — for a sighted colleague, for a scribe to work from — render it without comment. It is never offered and never rendered unasked, because offering it names what they are not getting; a person with a use for the file is not being handed a consolation prize. Outside those forks the handover ends the asking. Once the card is theirs, the questions on it are answered off-screen and come back on the return trip; walking the human through them in the chat turns a handover into an interview and spends the off-screen half of the loop the card exists to buy. Two allowances, and neither is an exception. The Light fork above: there the session was asked in conversation instead of handed over, so there is no card in hand to interview against. And the single offer made when a session never comes back — once, where a natural opening exists, to work the sheet's questions in the chat instead — which is not an interview against a live handover but the close of an abandoned one. `scan-back` holds the same line coming back — receive the return, never conduct it. The card is a single fenced message (the machine keeps its Mono on screen even when the page will have none), in two strictly separated parts. The separation is the format: - **COPY THIS** is transcribed by hand and is the only thing that is. Hard budget: **50-75 handwritten words for Deep, one phone screen for Light.** Copy load is stated in words. Never in minutes — the clock rule in §4 holds off the page exactly as it holds on it. - **DO THIS** is read once and never copied. Structure, house rules, and the pen legend live here, where they cost no transcription. What COPY THIS carries, in falling order of what survives a tight budget: - **The intent line, verbatim.** The only mandatory header element. On a printed kit it rides every page as a backstop; in a notebook it is the sole anchor, and everything §6 says about writing it for a cold reader is now load-bearing rather than insurance. Drop the title before dropping this. - **One prompt per page**, worded exactly as it would have printed, plus a page number in the corner. - **Machine items compressed to numbered handles**, four words or fewer each, under a hand-copied `I PROPOSE` caption, copied exactly — it is the same first-person caption the printed sheets carry, and `scan-back` matches it as a literal to recover provenance — and only on a page where the human reacts, selects, or ranks. Never on a page that also carries a generative zone: hand-copying is deeper exposure than glancing at a printed column, which is safe where reacting is the task and corrosive where generating is. The full items stay in the chat, where `scan-back` expands a handle against them on return; where there is no chat to expand against — a card copied out of a session that is gone — the handle is read as it stands, never expanded into the item it might be abbreviating, because an invented full item is the machine's words entering the record as the human's. - **At most one machine line dictated whole**, inside quotation marks. Quotation marks are the notebook's Mono, the only provenance mark a handwritten page has. A second one stops reading as the machine and starts reading as the sheet's own voice. DO THIS dictates structure as pen gestures, never as a description of a layout: "draw a line down the middle," "number 1-12 down the margin before writing anything in line 1," "bottom third of the last page: write OPEN and leave it alone until the end." Numbering before writing is what replaces the gravity of a printed slot count — a person filling a page top-down stops at four. The gesture is how the shape of the work gets said, not the work itself: where a gesture cannot be performed, the structure it describes is what carries over, and twelve slots are twelve slots whether they are ruled down a margin or typed as twelve empty lines. The pen legend dictated in DO THIS carries the same protocol §9 prints, minus one rung. On a page with no printed layer the cascade is **handwritten legend > defaults**: a card is dictated, never printed, so it supplies no ink key to sit in the middle, and a legend in the human's own hand is the only key an unprinted session has. Say the marks first and the hues second — the marks are what survive a page worked in one ballpoint, which is the ordinary case here rather than the degraded one. `scan-back` Step 1 reads the cascade back the same way. Spent Ground is the one pattern whose notebook form is a convention rather than a layout, and it is dictated as such: the categories are written small and struck through as they are written, which delivers the already-walked fence in the pen protocol's own vocabulary. A struck block copied that way is a fence, not a set of kills — `page-patterns.md` §Notebook translation specifies the form, and `scan-back` reads it back as a fence. If a design cannot be dictated inside the budget, the design is too print-shaped for a notebook. Simplify it: fewer zones, fewer handles, one prompt where there were three. Never dictate more. # Page Pattern Library Choose 1-3 patterns per session. Each entry notes capacity (Deep/Light), whose thought fills it, and the evidence behind its shape. Read `prompt-craft.md` before wording any prompt on these. Universal requirements, from the research: - Every generative zone carries **one input constraint** and **no time limit**. - Generation and selection never share a zone. - Slot counts run higher than feels comfortable; novelty arrives after the obvious categories are exhausted. - At least one zone per Deep kit invites **drawing**, not just prose. Drawn content consistently beats handwritten content on recall, across multiple studies including ones that found no handwriting advantage at all. ## Defixation patterns (use one on any generative sheet) **Spent Ground.** The highest-value pattern in the library. AI lists the three to five obvious solution **categories** at an abstract level, marked as already-walked, followed by the open generative zone. Not specific examples: categories. Naming common solutions specifically and asking people to avoid them *increases* fixation; naming them as categories roughly doubles creative output. AI fills the category list; everything below is human-only. **First Three Are Free.** A small boxed zone at the top of a generative sheet: "The obvious answers. Write them here so they stop taking up room." Then the real zone below. People fixate on their own first idea at least as hard as on a provided example, so the blank page is not a neutral starting condition. Human-only, both halves. **The Single Remote Example.** When AI genuinely should seed the space, exactly **one** example, chosen to be uncommon or from a distant domain, set in the machine's voice. Meta-analytic evidence: fewer and less common examples increase both novelty and quality, while common examples produce the most copying. Never two. Never typical. ## Sequencing patterns (Deep) **Cut-apart cards.** One item per card, title plus one-line promise, 6-8 cards per page with cut lines, plus two blank cards. The human cuts them apart and arranges them on a table until the arc appears. AI fills cards that summarize existing material; arrangement is human-only. The blank cards matter: without them the set reads as closed. *Scope this honestly.* Physical manipulation does **not** reliably improve problem solving in general; controlled studies find no facilitating effect on insight problems, and one found physical interaction actively hindered use of hints. The benefit appears when manipulation reveals structure that is hard to hold in the head: many items, relational or spatial arrangement, load beyond working memory. Course arcs, campaign phasing, and narrative order qualify. A four-item list does not. Use cards when the arrangement is the problem, not as a general creativity booster. **Numbered slots, empty.** A vertical list of numbered blank slots filled in order. Use when the items themselves are the human's to invent, not just their order. ## Weighing and selection patterns (Deep) **Two-column react.** AI's proposal in the machine's voice on the left, empty right column for re-ranking, corrections, and margin fights. The human writing "no" next to a confident machine answer is the loop working. **The kill sheet.** Everything in scope on the left; a right column with room for exactly half. The physical constraint does the cutting. AI fills the left; the right is human-only. **Impact grid.** A large 2x2 with axes chosen for the task. Items pre-listed as small labels to place, or blank to populate. Placement is human-only. **Selection counterweight (required on all selection zones).** One line, house-rule voice, near the picks: "One of these has to be the one you're not sure you can pull off." Idea selection performs no better than chance in controlled study, with a consistent bias toward feasible over original. The counterweight is cheap insurance against that bias. ## Generative patterns (Deep) **The long list.** A single prompt over **ten to fourteen** numbered short lines, with one stated input constraint and a two-part non-binding goal ("Aim for twelve. At least two should be ones you'd be embarrassed to say out loud."). Slot count is the whole design: the common categories get exhausted in the first third, and novelty lives in the back half. Human-only. **One provocation per page.** A single hard question at the top, the rest of the page empty. For questions that deserve a page of wandering. Use three to five of these as an entire kit when the work is genuinely conceptual. **Constraint box.** A box that physically fits about twelve words, with the constraint stated above it. The limit does the editing. **Sketch field (required somewhere in every Deep kit).** A dot-grid zone with a prompt that asks for a drawing, diagram, or spatial arrangement rather than sentences: "Draw the shape of the arc." "Sketch what the room looks like in week eight." Drawing outperforms writing on recall, and it recruits a different retrieval path than prose. The requirement is scoped rather than absolute, and it yields on exactly one path — a dictated session worked without sight of a page, where there is nothing to draw on and nothing to look at again. There, where the sketch cannot be worked, the kit carries the prompt the drawing was there to ask, and nothing takes the drawing's place: a substitute activity would inherit the mechanism's name and none of the evidence under it (`evidence.md`, the sketching entry under the non-visual path). ## Light patterns **Rank-and-circle.** AI's list, each with a rank box and room for a one-word note. Circle keeps, cross out kills. Carries the selection counterweight. **Gut-check ladder.** Items down the left, a simple scale per item. One pen-stroke each. **Reaction margins.** A short AI-drafted artifact printed with wide margins and double spacing, for slash-and-scribble. ## Introspection patterns (Deep, own sheet) **Distanced pair.** Two zones on one sheet: what happened and what you make of it, then what it felt like. Writing about thoughts *and* feelings beats either alone. Prompts written in the distanced voice ("What would you tell someone standing where you're standing?"), since self-distancing is the mechanism that makes reflective writing work. Never share a sheet with ranking or planning zones. ## Field patterns (away from desk) **Observation log.** Timestamp, observation, and a "why it matters" column. **Tally sheet.** Predefined categories with tally space plus an uncategorized overflow zone. ## Universal closer (every sheet) **Open territory.** An unstructured zone, minimum a third of the final page. Never omit. The margins and the wandering are frequently where the best material comes back from. ## Selection patterns (new; use one whenever a kit generates options) **Fast pick.** The options listed plainly, a named criterion in the prompt, and room for one line of reason per pick. No scores, no matrices, no rubrics. Deliberative selection performs at chance while intuitive selection beats it, so the zone should invite a gut call and then get out of the way. The criterion must be named ("the two most original," not "the best"), because that is the only selection instruction shown to work. Add a distance frame: "Choose as the person who has to live with this a year from now." Never mention building, shipping, or executing on this sheet; the prospect of implementation measurably shifts people toward safe picks. ## Analogy patterns **One far door.** A single cross-domain analogy in the machine's voice, one line, no elaboration, followed by an open generative zone: "Where else does this problem already have a solved version?" Far analogies raise novelty, near ones raise quality, so pick the distance to match the sheet's job. If the source domain is genuinely alien to the person, add a one-line **bridge** naming something familiar in between. Deep explanation erases the effect; keep it to a sentence. ## Sketch patterns (expanded; this is the strongest paper-specific mechanism available) **Sketch field.** A dot-grid zone with a prompt asking for a drawing, diagram, or spatial arrangement rather than sentences. Beyond the recall advantage, loose marks fight fixation directly: vagueness invites reinterpretation. Keep the prompt loose enough that the drawing can come out ambiguous, because the ambiguity is the working part. **Second reading (pair with any sketch field).** Below the drawing zone, one line: "Now look at what you drew and find a reading of it you did not intend." This is trainable, not innate. Novices instructed in this skill generated more interpretations of an ambiguous drawing and kept generating longer than uninstructed novices. It is the cheapest upgrade in this library. **Regroup.** "Circle two things on the page that you have never thought about at the same time. What do they make together?" Unexpected discovery is most likely when attention lands on elements never previously attended to together. Works on a fresh sketch or on a page already covered in the person's own marks. **Structure sketch.** For any kit with multiple parts: "Draw these as a structure. Mark what carries the weight. Cross out what turns out to be decoration." Doubles as a selection zone, and answers relational questions that prose cannot reach. ## Named session formats (whole-session kits; choose by name) Whole-session structures with real lineages — the after-action review out of Army debriefing practice, the premortem out of decision research, the field kit out of museum natural history. Choose one by name when the task matches its trigger, rather than assembling zones from scratch. Each composes patterns this library already has and adds only what its own evidence licenses; machine-filled zones carry the I PROPOSE caption as everywhere else. Every format has a named cluster in `evidence.md` (evidence brief, after-action review cluster, and so on), and any exception a format states to a library default — the premortem's slot cap against the overbuild rule, the outside view's pre-ticked anchor — is sourced there; an exception without its source is drift, not design. Entries run longer than zone patterns because each specifies a whole kit. **After-action review.** Deep, 2-3 pages, after a project phase or a completed agentic session that left a recorded plan and execution log. The printed record is a hard condition, not decoration: written reflection without an objective record to react to produces no improvement at all, and the record is what qualifies this sheet for its evidence base. Page 1: the plan as recorded ("what was supposed to happen") and, where the machine executed steps, a bare factual timeline with no interpretation, in the machine's voice, with a two-column react beside it under the line "The log is not the whole truth." Page 2: the gap zone — "Why the difference?" — human-only; the machine withholds all diagnosis. Then twin sustain/change lists covering successes as well as failures; after a success, reviewing what went wrong anyway is the working part. The four canonical questions are the fixed structure — what was supposed to happen, what happened, why the difference, what to sustain and change — and high structure outperforms low. Nothing gets ranked here: prioritization, if wanted, is a later Fast pick, and next actions get drafted on screen after scan-back. Sketch field: "Draw the timeline where it bent," paired with Second reading. Never offer it as emotional processing after a distressing event — that literature runs null-to-harm, and introspection belongs to the Distanced pair. No record, no AAR; a trivial session with no gap gets no sheet. Sources: evidence brief, after-action review cluster. **Premortem.** Deep, 2 pages, only when a concrete committed plan exists. Page 1: the live plan restated in one tight paragraph, machine's voice — genuine work worth reacting to. Then the certainty frame, asking voice: "It is a year from now. The plan failed — completely, not marginally. Write the history." Certainty is the active ingredient, not pastness; the wording must make the failure total and certain. Below it, **six to eight** numbered lines — a deliberate, sourced exception to the ten-to-fourteen overbuild rule. Overbuilding serves novelty; this sheet serves confidence calibration, and forcing many failure thoughts feels so difficult that the felt difficulty restores the very confidence the exercise exists to reduce. One input constraint: "At least three must be failures that are our own doing, not the world's" — the documented skew runs toward blaming the world. Page 2: sketch field, "Draw how the failure unfolded," with Second reading; then a separate Fast pick — circle the two most likely to be the real cause — with the counterweight adapted to "One circle must be the reason nobody would say in the meeting," because this sheet's bias is self-censorship, not feasibility. Countermeasures are deliberately absent from the sheet; the AI drafts them after scan-back, which closes the documented gap between naming failure reasons and revising the plan. No timer, no mitigation wording anywhere. Not for Light, not for an already-anxious planner — every supporting study targets overconfidence, and deflating the deflated has no support — and not for repeat runs on an unchanged plan. Sources: evidence brief, premortem cluster. **Outside view.** Deep, 2 pages, for a timeline, budget, or adoption number the human is about to commit to. Page 1: the reference class in the machine's voice — comparable cases and base rates, with provenance stated on the page in one line; prefer the human's own past cases when they exist. Below it, a distribution strip on dot-grid with the base-rate anchor pre-ticked; the human marks where this case honestly sits. One pen position, and the spatial judgment is the point: prose lets you dodge it. The strip is the kit's sketch zone. Then the forced linkage, non-optional because it is the working part: "Write the short history of this project going the way the reference class typically goes." Displaying past cases beside an estimate does nothing; the linkage writing is the intervention. Page 2: the differences list — "What would have to be true for you to beat the base rate. Each line must be evidence you hold today, not an intention" — then the committed number, written once, that the human will defend. Promise calibration, never accuracy: the tested effect removes systematic optimism without sharpening any individual estimate. The pre-ticked anchor is licensed for this estimation format only and never leaks onto a generative sheet, where the same move is design fixation. Decline the format when no defensible reference class exists or the base rates' provenance cannot be stated: a fabricated distribution printed as an anchor is worse than no sheet. Sources: evidence brief, outside view cluster. **Teach-back sheet.** Deep, 2 pages, closed-book by design. Offer it only after the human has genuinely engaged a body of conceptual explanation — an hour of AI-explained material is the canonical trigger — and never after a skim: closed-book prompts after passive processing measurably hurt, because nothing is loaded to retrieve. Page 1: one Serif line — "Explain X, from memory, to someone who has never met it" — over a full ruled page, with a narrow gap column: "Mark the lines where you reached for a word you couldn't cash out." Ask for mechanism, how and why it works, never a list of reasons; the deflation of inflated understanding requires mechanistic explanation and does not happen otherwise. Page 2: dot-grid — "Draw the mechanism from memory" — with Second reading; at most three retrieval probes in the machine's voice, no answers. The machine withholds every scrap of its own explanation, including material it obviously has: printing the content converts a calibration instrument into a recognition test, and paper is the enforcement the screen cannot provide. Closed-book is the calibration instrument, not the only route to retention. Gap marks are annotation, never scoring. The return trip is the feedback event, under the one scoped exception to scan-back's authority rule: the handwritten explanation is authoritative as a record of what the human believes, not as fact, so scan-back checks it against the source, corrects factual errors on return, and re-explains the gap-marked pieces differently. One sheet is one retrieval event, not a retention program. Wrong tool for procedural or worked-example material, and never a substitute for a generative sheet. Sources: evidence brief, teach-back cluster. **Weekly review.** Recurring ritual, Deep, 2 pages — the weekly self-meeting the skill description already solicits. Page 1: the inventory of open loops the machine can actually see — chat threads, task lists, calendar — laid out two-column react; the human triages with the pen protocol as designed: strike kills, green go, blue do, red review. These are reaction marks on machine work-product, rank-and-circle territory, not selection over creative options, so no selection counterweight prints; feasibility bias has no originality to corrupt here. Never quadrants, never a priority matrix, never scoring of any kind. Page 2: the mind-sweep, a human-only long list — "What has your attention that no system knows about?" — with the machine-visible loops marked as Spent Ground categories so the sweep goes where the machine cannot see; then a structure sketch: "Draw next week as a structure. Mark what carries the weight." No next-action or if-then wording prints anywhere on the sheet. Capturing undone tasks without then planning them leaves the intrusions intact, so the continuation after scan-back is load-bearing, not cleanup: next actions get drafted there, kept deliberately lightweight — elaborately batch-planning many goals in one sitting undermines the planning benefit — and the machine transcribes the human's plans, never audits them. The mind-sweep is never pre-filled: the machine's inventory is precisely not the human's mind, and printing guesses there is the design-fixation failure mode. Weekly, not daily. Never fused with introspection; the Distanced pair keeps its own sheet. Skip the format when the machine holds no genuine visible inventory — never invent loops. Sources: evidence brief, weekly review cluster. **Grinnell field kit.** Deep **series** — the library's first multi-day format, and the multi-day sibling of the Observation log. One kit per sitting, up to 4 pages per kit: the Deep page cap counts per kit, not per project, and more entities means more sessions, not more pages. For projects with recurring entities worth tracking across days — stakeholders, subsystems, failure modes. Page 1, the journal: a printed header of labeled blanks the human fills at the time of writing — date, location, conditions — then a ruled narrative zone in the fact-forward register: what happened, who, where, with human-written times as data. Times are quantification, not schedule; the Observation log's timestamp column is the precedent. A dot-grid band runs beside the narrative for sketch maps and diagrams — the lineage's own instruction, and it satisfies the sketch requirement natively. Account pages, at most two per kit: one per recurring entity, the entity's name printed as the page header in the machine's voice — gathered inventory at category level, categories-not-instances territory — and every field below empty, human-only. A compact catalog closes the working pages: a numbered one-line register of the concrete artifacts captured this session — documents, links, quotes, samples. The journal is not a diary; the lineage's own doctrine excludes the reflective voice, so this kit never merges with the Distanced pair. The same-day writing discipline is real doctrine, and it lives in the chat framing and the human's practice; nothing on a sheet ever schedules, deadlines, or hurries. Kits return days apart and account pages accumulate: scan-back appends to an entity's running account across sessions rather than treating each scan as standalone, and reorders by the human-written dates, which are data. No generative zone, so no defixation pattern is required. Not for Light, not for ideation; a project with no recurring entities is already served by the Observation log. Sources: evidence brief, Grinnell field-note cluster. **Serial disclosure kit.** Deep **series**, the second multi-day format after the Grinnell field kit: four pages across three consecutive-day sittings — pages 1-3 are the sittings, page 4 is the distillation that closes the series. Cadence carries in the chat framing and a printed sitting label ("First sitting", "Second", "Third"); no date, deadline, duration, or schedule prints anywhere. The three disclosure pages are identical: one Serif provocation in the classic register — first person, deepest thoughts *and* feelings about the friction together — over a genuinely full ruled page. The page is the dose: every trial doses in minutes across days and none has run untimed, so the full page is this skill's own conversion, labeled untested wherever it is claimed and never shrunk — an undersized dose left writers with more negative appraisal, not less. The machine contributes the provocation wording and nothing else: no examples, no sub-prompts, no I PROPOSE zone. No format withholds more, because the writing is the intervention. Distanced wording is banned on them, a scoped exception to `prompt-craft.md` §5: instructed distancing raised depressive symptoms in vulnerable writers. Each prints, in the infrastructure voice, `PRIVATE — THIS PAGE STAYS WITH YOU. IT IS NOT SCANNED.` (the marker is specified in design.md §9; print it exactly). Scan-back honors it: never requested, never asked after, their absence never read as information. Page 4 is the only page that returns, and where the Deep sketch requirement is met: dot-grid, "Draw the shape of this over time," with Second reading; then a TLDR distillation zone in the distanced voice ("What would you tell someone standing where you're standing about what this was?"); open territory closes. Offer only for work friction and professional setbacks, never as therapy or treatment. Never in acute distress, after recent or severe trauma, or to anyone seeking trauma processing — clinical territory this skill does not enter — and never to a self-described low expresser, where anxiety worsened. Say in the offer that the first sittings reliably feel worse, a dip unrelated to whether it works; promise no relief. The effect is small, delayed, and accrues across sittings, which is why this is a series and not a sheet. Sources: evidence brief, expressive writing cluster. **Brainwriting rounds.** Deep, group-only: three to six co-located people, one sheet each, passed by hand. Never solo — the solo interleaved variant was evaluated and cut, since re-anchoring a person on their own first concepts may reinstate the fixation the break relieved. The Deep page cap counts per participant, not per table: it protects one human's completion load, and each participant completes one page, so a six-person session is one page pattern printed six times, not a six-page kit. The sheet: the shared problem in one Serif line; a Spent Ground category header aimed at teammates' obvious ideas as much as the machine's, since copying concentrates on common upstream ideas whoever wrote them; then rows of three idea cells per round, each cell lightly dot-gridded so a sketch is as welcome as a sentence. One protocol line, event-based: "When your three are down, pass left." Each recipient reads the column above and adds three that extend, combine, or fork it. One `HAND:` box per sheet in the infrastructure voice — furniture, not a heading — routes a follow-up to whoever wrote a line. Filling it is optional and leaving it blank is a fully supported choice; nothing printed or said asks for a name, and the nominal-group evidence may favor anonymity outright. Ink keeps meaning intent, never identity, and ideas stay unattributed to their authors in synthesis. Nothing else prints: no timer (the five-minute round is the brand's convention, never isolated as active), no round count as a promise, no idea-count arithmetic, and no attend-and-memorize instruction — reading the column above does the attention work, and a printed memory instruction cut fluency in the one study that tested it. Selection is exiled to a separate Fast pick sheet after the rounds, with the named criterion, distance frame, and counterweight, never on a passing sheet. A slim open-territory band closes each sheet. No group, no format; a task that is not idea generation belongs to a different kit. Sources: evidence brief, brainwriting cluster. ## Notebook translation (when the session is dictated, not printed) When the human cannot print, or cannot read the printed sheet, the session is dictated as a setup card to copy into whatever notebook they own — or into whatever stands in for one, since a person who cannot read the page may work the card as a typed file, a slate, or a scribe's transcription (`prompt-craft.md`, "Dictating instead of printing"). Choose against this gate before designing anything: some patterns survive the copy intact, some survive only under a restriction, and some do not survive at all. The restriction is one line: **machine handles never appear on a page carrying a generative zone.** Copying by hand is deeper exposure than glancing at a printed page — the working hypothesis, flagged as unsourced in `evidence.md` — which makes handles safe where reacting is the task and corrosive where generating is. It bans handles, not every machine word: a Spent Ground category is a fence, not material to work from. **Translate near-verbatim.** Nothing load-bearing is lost in the copy. - One provocation per page — the best case in the library. A question copied at the top of an empty page is the entire pattern. - The long list — with one procedural rule standing in for the gravity of printed slots: number every line before writing in any of them. A page numbered as it fills stops where the obvious categories stop. - Sketch field, and Second reading with it — arguably improved. Dot grid is what most notebooks already are, and the second-reading prompt is one line to copy. That holds where the card is worked on a page; where it is worked without sight of one, the scoping above applies — the prompt the drawing was there to ask still carries, and nothing takes the drawing's place. - Constraint box — the human draws the box. A hand-drawn box fits what it fits, and the limit still does the editing. - Distanced pair — the isolation a dedicated sheet gives it becomes a page turn. Start it on a fresh page, keep both zones there in order, and let no ranking or planning zone share it. - First Three Are Free — both halves were always human-only, so there is nothing to dictate but the prompt. - Spent Ground — through its own convention: the categories are written small and struck through as they are written, which delivers the already-walked fence in the pen protocol's own vocabulary. Categories-not-instances is unchanged. Dictate territories; an instance dictated here is the failure the pattern exists to prevent. **Translate only through handles, and only onto a page carrying no generative zone.** - Two-column react, Rank-and-circle, and the kill sheet each need machine work-product to push against. It crosses as numbered handles with the full items held in chat; the human rules one line down the page, or leaves the back half of each line empty. Light sessions get the same work offered in the chat instead, so in practice this is a Deep route. **Do not translate.** - Reaction margins. The pattern is space around an AI draft, and no draft reaches the page unless the human transcribes it whole — the one copy this gate exists to refuse. - Cut-apart cards. A bound notebook cannot be cut. Index cards restore the scissors but not the pattern, since the cards would carry hand-copied machine summaries; cards work only where the items are the human's own to invent. When the pattern you wanted lands here, choose a different pattern that does the job. Never invent a notebook substitute: it inherits the name of a mechanism and none of the evidence under it.