--- name: create-report description: >- Draft a standalone report for a reader who is not the operator and was not in the working session: a decision-maker, a client, or a reviewer who will read the report alone. Fires whenever the deliverable is a report, memo, position paper, background paper, or staff study that must carry its own conclusion, evidence, and stated uncertainty to that reader. Not for notes you keep for yourself, and not for prose written back to the operator. --- ## Meta-rule Operator instructions for this report outrank every rule in this skill. Operator instruction differs from anything below -> instruction wins. Following instruction against a rule here is a deviation, and the report records it. Record shape: references/checklist.md. ## Scope: this skill drafts, it does not analyze Analysis is work already done. Steps below turn that work into a report. Step naming an analysis procedure (key assumptions check, criteria and decision matrix, evidence tests) names a standard the finished work must meet and record, never a fresh task to run. ## The report contract Every run meets all of these. - The written-out question governs the report. - The thesis and main conclusion appear in the first paragraph, with the primary supporting reasons grouped into categories, not a flat list. - The first paragraph announces the order; the body follows it exactly. - The body contains only what the first paragraph promises. - Every container announces its point in its own label. The title states the conclusion, each subhead states its sub-conclusion, each paragraph opens by saying its point, and each graph is titled with its thesis. (This four-level container law is a synthesis across four of Daniel's rules, marked as such. No single page states it.) - Say nothing that is not part of the answer to the question asked. - Every claim carries specific support, every number is compared to another number, and every source is listed. Four more clauses, in the contract's own voice: - Every major judgment states its likelihood on the ladder the report publishes, and states confidence in the judgment separately. Likelihood and confidence never share a sentence. A judgment that is not empirical carries no band; it states its premise as a premise and names where the premise comes from. - Assumptions are marked as assumptions and never presented as information. Each critical assumption states what follows if it is wrong. - Where a recommendation is asked for and the analysis compared options, at least two genuinely distinct options were compared against criteria set before the options were generated. Where the analysis instead states that no matrix applies and gives its reason, a defended single option, a screening that emptied the field, or a question of fact, this clause is met and the report owes no deviation entry. A single option handed over with no such reason is not a recommendation, because nothing was compared. - Every major judgment names what would change it. A judgment with no stated falsifier is not finished. A judgment that is not empirical names instead what would have to change for it to stop holding. Design mark on those four: ICD 203 writes several as "should" and "as appropriate". Contract states them as musts. Skill design decision, not source wording. Report prints the ladder it uses. Ladder table: references/style.md. Contract is also the revision rule. Any revision leaves the first paragraph's promise true: new material merges into the announced structure, or the announcement changes with it. ## Do not include Directly under the contract. Non-negotiable. Do not restate the problem; do not announce that the research was completed; do not summarize background; do not define terms; do not explain calculations; and do not footnote your own conclusions, presentation explanations, or summaries. One more, at the register layer: do not state an inference in the grammatical form of a fact. Underlying information, assumptions, and judgments stay visibly distinct. ## The eight steps 1. **Assignment.** Write out the specific question the report answers before anything else. Topic is the question; thesis is the answer. Yes-or-no asked means yes-or-no answered. "Look into X" is not an assignment. One controlling purpose governs, and every sentence serves it. Where the question came from someone else, get the problem statement approved before working on it, cast as who, what, when, why, and how. Run the key assumptions check here, against the answer the assignment already presumes. It runs twice: here and at step 8. 2. **Audience.** Set two dials together: omit as much as possible (what the reader already knows) and be as technical as possible (what the reader can absorb). Default reader: intelligent, suspicious, busy. Address four roles, not one reader: primary receivers, secondary receivers, key decision makers, and gatekeepers. 3. **Notes.** Collect the facts the report will use in their own file beside the report, named `.notes.md`. Transform each idea into your own words rather than copying text, and record each fact with its source. This is the counter to pasting session context into the report. Check 22 fails when the file is not there. 4. **Thesis.** Form it from the notes, after them, never as a starting assumption. State it with its primary supporting reasons grouped into major categories. 5. **Plan.** Build a skeleton plan (the order of arguments), then a final plan that correlates every note with the point it supports. No prose yet. A note supporting no point gets cut here, which enforces the exclusion rules before drafting. Where the report recommends, this step lays out the criteria and the decision matrix the analysis already used; both are standards that analysis must have met, not work to do here. Analysis that arrives stating the matrix does not apply, with its reason, a defended single option, a screening that emptied the field, or a question of fact, has met the standard: draft from what it handed over. Analysis that owed a matrix and arrives without one has a gap this skill cannot close: say so and send it back. Writing a criterion, setting a weight, or building a matrix at this step is doing the analysis, which this skill never does. Two moves are compliant here and there is no third: send the work back, or draft from what was handed over. 6. **Draft.** Largely mechanical if step 5 was honest. Apply the exclusion list, the evidence rules, the sentence layer, and the uncertainty layer. 7. **Revise.** Read the draft aloud, or have it read to you. Run the connective sweep: references/style.md. Check the words a spell-checker passes but the sentence misuses. Pause before you send. Three passes, in order: big picture, then paragraphs, then sentences and words. 8. **Check.** Run references/checklist.md against the finished draft. Score every check on `.checks.md` beside the report, then run `bash /scripts/check-report.sh `, passing the folder the operator named, and fix what it names. Run the key assumptions check a second time and run the sensitivity test; record both outputs. The heavy challenge techniques (competing hypotheses, Team A and Team B, devil's advocacy, red team) run only when the operator asks for them, never as your default. ## Collisions Five rulings an agent otherwise gets wrong: data plural, hedging versus calibration, conclusion-first versus hostile reader, analysis order versus presentation order, named format versus exclusion list. references/collisions.md. Load when two rules here appear to conflict, or when the report uses a named paper format, whoever chose it. ## Failure modes Four named ways an agent-drafted report fails. - Context bleed. Session context or source text pasted into report, not transformed notes. - Agent narration. Report announces own process: restated problem, research announcement, explained calculations, footnoted own conclusions. - Revision as append. New material bolted on the end, not merged into the announced structure -> first paragraph's promise broken. - False confidence. Inference stated in the register of a fact, or conclusion given with no stated likelihood or confidence. ## Pick the format from the function Pick one shape from the function it serves. Hold it fixed whole report. Detail and skeletons: references/structure.md. - Point paper: one-page bullet brief, quick reference or decision now. - Talking paper: bullet talking points to carry into discussion or briefing. - Bullet background paper: background in bullets, reader needs facts fast. - Background paper: same background in prose, reader needs the reasoning. - Position paper: argues one position, asks reader to adopt it. - Staff study: full problem-to-recommendation report, five-part body. ## Reference files Load on demand. Each is this skill's own file. - references/planning.md. Steps 1 to 5: assignment, audience, notes, thesis, plan. Load when working steps 1 to 5. - references/recommendation.md. Criteria, screening and evaluation, weights, decision matrix, committing to one course. Load when report recommends among options. - references/structure.md. Organization pattern, paragraph template, cover letter, paper formats above. Load when choosing a shape or building the skeleton. - references/skeletons.md. Staff-study five-part body and running-estimate format, verbatim. Load when the report is a staff study or a running estimate, whoever chose it. - references/style.md. Sentence layer, numbers rules, likelihood ladder, weasel list. Load when drafting or revising sentences, or stating uncertainty. - references/evidence.md. Comparison mandate, graphs, footnote policy, source credibility, argument tests. Load when placing numbers, graphs, or sources. - references/checklist.md. Step 8 checks, key assumptions check, three-pass edit, structured challenge. Load at step 8 before finalizing. - references/collisions.md. Five tie-breaks. Load when two rules appear to conflict. - references/challenge-heavy.md. Four heavy techniques. Load only when the operator names one. - references/sample-reports.md. One bad report, one good rewrite, annotated. Load when you want a worked example of the contract.