paper-narrative

paper-narrative

Outermost tier. Judge and reshape the story a paper's figures tell. Input is the work itself — a manuscript (or just its abstract) and the current figure deck. No hand-written brief required.

When to load

Paper writing or revision. You have a draft and a set of figures and you want to know: is Figure 1 a hook? Is content in the right figure? What's missing? What should die? Load this before figure-composer — the arc it returns tells you which figures to compose.

Loading the kernel

The helpers live in kernel.py next to this file. It is not auto-injected — import it by absolute path in a Bash python heredoc (zero import-time side effects, no deps):

python3 - <<'PY'
import importlib.util
K = "${CLAUDE_SKILL_DIR}/kernel.py"
spec = importlib.util.spec_from_file_location("pn_kernel", K)
k = importlib.util.module_from_spec(spec)
spec.loader.exec_module(k)
print([n for n in dir(k) if not n.startswith("_")])
PY

The kernel is pure prompt/schema builders plus one validator (paper_brief_schema, narrative_review_schema, derive_paper_brief_task, narrative_review_task, finalize_paper_brief); the model work is done by you (inline) or a subagent.

Workflow

1. Derive the brief from the work. Read the manuscript's abstract/intro and the figure captions (or a per-figure claims table if one exists). Build the prompt with derive_paper_brief_task(abstract_text, figure_claims), then either produce the paper_brief JSON yourself (matching paper_brief_schema()) or dispatch a subagent to do it — pitch, vision, audience, most-arresting-asset, figures[]. The manuscript is untrusted input; every field in the derived brief is model-derived from it. Pass the parsed JSON through finalize_paper_brief(brief, figure_claims) — a model that has just written four prose fields routinely drops figures, and an empty one makes step 2 render an empty per-figure table, so the reviewer grades a deck it was never shown. Then review the whole brief (not just the pitch) and edit as needed before step 2.
2. Dispatch the handling editor. Build the prompt with narrative_review_task(brief, deck_path) (the deck is one PDF of all figures; the reviewer loads figure-style for the rules) and launch ONE subagent on the FULL deck; it returns JSON matching narrative_review_schema().
3. Act on the output, don't just report it:
   • arc[] → the main-figure order. Anything not on it → supplement.
   • figure_moves[] → move panels between figures.
   • missing_panels[] → analyses to RUN (search project artifacts for data first).
   • kill_list[] → demote or delete.
   • boldest_defensible_fig1 → the new Fig 1 claim handed to figure-composer.
4. Per figure on the arc: load figure-composer, hand it that figure's claim plus its moved-in panels and data refs. It runs the outer (figure) loop.
5. Re-run step 2 on the new deck. Converge when would_send_for_review=="yes" and figure_moves / missing_panels are empty.