Narrative

Narrative Skill

Two jobs, in order of importance:

1. Revision protocol — how to absorb PI/user feedback without the report degrading back into a patchwork (most narrative damage happens during feedback rounds, not first drafts).
2. Article-type templates — positive examples of section logic per article type, consumed at outline time.

This skill is the delta on top of brain.md's <report_synthesis_protocol> (outline-first, claims-as-titles, lab-book test, anti-stacking pass). It does not repeat those rules — it adds what they lack: per-type section logic, the figure-narrative binding, and what to do when feedback arrives.

First draft (thin — most rules live in brain.md)

1. Pick the article type and record it as the first line of notes/report_outline.md: type: empirical | feasibility | comparison | policy-zh | survey. Surveys: stop here and follow skills/review/ + skills/survey-methodology/ instead — never both pipelines.
2. Read templates/<type>.md BEFORE writing the outline. The template gives the section logic for that type; your outline instantiates it with this project's claims. The lab-book test catches structure that mirrors the experiment DAG; the template shows what to write instead.
3. Read references/figure_narrative.md before commissioning figures. The outline must name its Figure 1 (schematic) and its hero figure (the one figure that settles the central claim) — both are outline-level decisions, not afterthoughts.

Revision protocol (core)

Every feedback batch that touches the report gets classified BEFORE any edit. Mixed batches: tag every item first, edit second (tag-all-before- edit-any) — append-only reviews/pi_feedback.md means earlier instructions never vanish, so there is no need to rush.

Add the class tag to each checklist line in notes/memory.md (the existing pi_correction_protocol checklist — same boxes, one tag + one-line rationale extra):

- [ ] <instruction verbatim>  [class: local-fix | section-rewrite | restructure — <one-line why>]

Class 1 — local-fix (a number, a wording, a citation)

The pi_correction_protocol order applies unchanged: ledger first, report second. Then:

• grep report.tex for OTHER occurrences of the corrected number (abstract, captions, tables — printed values drift in packs, not alone).
• re-read the first sentence of the touched paragraph against its section's outline thesis (10 seconds — does the paragraph still serve the claim?).
• the outline is NOT touched — unless the corrected number appears in, or defines, the central-claim sentence (e.g. a threshold comparison like "99.987% < 99.99%"). Then this is not a local-fix; reclassify as restructure and follow that flow.

Class 2 — section-rewrite (a section's argument is wrong/weak)

Precondition: if the feedback strikes an L2 claim, the experiment re-spawn (ledger fix) comes FIRST; this class begins only after notes/experiments.md § L2.X reflects the new physics.

1. Edit that section's block in notes/report_outline.md first (thesis / evidence / synthesis move).
2. Rewrite the section from outline + ledger — single-section scope (mv the old text into the editor's view, don't patch sentence by sentence around a broken spine).
3. Check the transitions: the last paragraph of the previous section and the first of the next still hand over correctly.