--- name: improve description: "Critique an outside idea (a pasted tip, post, transcript or video summary) against what this system already has. Use when the founder pastes something and asks whether to adopt it, or says 'improve', 'should we do this', 'compare this to what we have'. Grounded by scripts/improve_ground.py in the lessons corpora and named files; returns adopt, skip, or already-built, and NEVER decides what to build, sell or publish." --- # Improve: an outside idea, judged against the system that exists Plan item 2d of the morning-brief overhaul (Bloom's `improve` pass). The founder pastes an idea. This skill answers one question: does it change how the SYSTEM works, and is that change already here? ## The two halves 1. **Deterministic (this skill must run it, never reason from memory):** ```bash python3 plugins/kipi-core/skills/improve/scripts/improve_ground.py --target "" ``` It prints a JSON verdict with `cites` (lessons paths or named files) and a `corpora` report saying which lessons directories were read, missing or unreadable. Exit 0 for `already-built` / `adopt`, 2 for `skip`. Set `KIPI_LESSONS_CORPORA=/path/one:/path/two` to ground against more than this instance's corpus; the report names each one either way. 2. **Judgment (the skill):** read what the script cited. Then write the critique in this shape: - The idea, in one sentence, as a SYSTEM change (a stage, a trigger, a rule, a context entry). If it cannot be phrased that way it is roadmap, and the verdict is skip regardless of how good the idea is. - What already exists: the cited file or lesson, opened, with the line that covers it. Not "we have something like that". - Verdict: `already-built` (name the file), `adopt` (name the smallest change and the check that would prove it), or `skip` (name the reason the classifier gave). ## Hard constraint This skill fixes the system. It never decides the roadmap. An idea about what to build, sell, price, publish or advise a client returns `skip` from the script (`roadmap_scope.py`, fail-closed on unknown) and the skill does not override that with judgment. If the founder wants the product conversation, that is a conversation, not this skill. ## Do not - Cite from memory. If the script's `cites` is empty the verdict is not a verdict; say so. - Widen the trigger. This is on-demand only; it never rides the brief or the weekly pass. - Read a corpus the report says is missing. Say it is missing. ## Changelog - 2026-09-01: created (prd-morning-brief-learns, issue mbl-improve-skill). Verdicts cite a lessons path or a named file; corpora contract via KIPI_LESSONS_CORPORA; roadmap ideas skip by classifier, never by judgment; one line per entry because test_skill_changelog.py reads each line.