Extract narrative methods

Extract Narrative Methods

A long narrative carries reusable methods — how a weak actor negotiates, how a manager allocates power — buried inside plot. This skill runs an auditable pipeline that turns those into a traceable methodology, and treats the source book as three things at once: a deliverable, a training set, and a regression suite for the workflow itself.

Two roles own the work:

• Codex owns the workflow and is the final evidence judge. It routes work between a cheap map model and stronger reasoning, audits against source, signs cards, and decides the final methodology.
• A cheap map model (DeepSeek Pro, Codex mini, or similar) does bulk reading and first-pass cards. It never declares its own output accepted.

This skill was validated end-to-end on a 556-chapter novel (《孺子帝》): 19 chunks, ~1.9M characters, reduced through R1–R5 to 8 stable methods across two lines. The constraints below come from that run; batch capacity remains a project-specific calibration result, not a universal constant. See references/model-routing.md for measured findings.

Two Tracks, Run In Parallel

Production track — finish the book. Research track — turn every audit finding into a reusable rule, sample, or cost number, so the next book is cheaper and the workflow stops repeating mistakes.

The single most important rule of the research track: when an audit finds a problem, do not paste a new sentence into the long prompt. Route it (see "Improve Constraints"), and keep one positive and one negative sample so a later prompt edit can't fix one error while reintroducing another. The first run's most expensive lesson — batching silently degrades reading mode — now lives as a stop threshold and a regression sample, not as prose in the producer prompt.

Start A Project

1. Run scripts/init_project.py <analysis-dir>.
2. Identify the source text, chapter markers, target lines (e.g. weak-side negotiation A, management B), and expected outputs.
3. Run scripts/split_source.py <source> <chapter-dir> or provide a project-specific splitter. Verify chapter count and continuity before production.
4. Only load the current chunk's source slice into a producer.
5. Freeze a project contract (source path, chunk boundaries, line definitions, schemas, model routing, current rule version). Store source-specific choices in the project, never in this generic skill.
6. Keep tests and legacy outputs outside analysis/cards.
7. Read references/workflow.md before designing the first producer prompt.

If adapting an existing project, inspect its current cards, audits, failed tests, cost record, and rule history before changing anything. Trust the artifacts' real state over a stale progress table.

Run One Chunk — Four Layers

Each chunk passes four layers in order. A layer never starts before the previous one closes. Mechanical validity is not semantic acceptance.

Self-use disposition rule (overrides any "auto-delete / auto-stop / re-run" wording anywhere below). No layer and no script ever deletes, downgrades, rejects, or re-runs a card on its own. Every gate is a marker, not a judge — it attaches a disposition or pushes an item onto the human review list, and the human (who has read the book several times) decides what to actually do. The only hard stop is a true parse failure (missing field, unreadable file) that would crash a downstream script — that is mechanical, not judgment. Re-running a chunk is always a human decision: by default every chunk runs exactly once and the markers are read afterward.

1. Evidence layer — chapter-by-chapter events, source evidence, result state, boundary record. No method claims yet. Mechanical checks are scripted.
2. Candidate layer — generate A/B card candidates from the evidence. Every card must answer actor · dilemma · action · counterpart response · cost · result. No closed loop → downgrade to an event, never force a card. ("本块无合格A线卡" is a correct, expected output.)
3. Review layer — Codex medium reasoning audits every candidate against source and marks (never deletes): causal inflation, duplicate splitting, A/B misrouting, result overstatement. Every finding becomes a disposition tag carried into the human review list, not an automatic removal. Low-risk formatting is left to scripts.
4. Synthesis layer — only after the whole book is done: cross-chunk dedup, stable methods, applicability and failure conditions, counterexamples, character-capability arc. This is the only layer that uses high reasoning.

Treat all producer output as raw. Require an event audit, candidate cards, and a self-check. Title scans, framework drafts, and placeholder prose are marked incomplete (a real parse-level placeholder is the one mechanical hard stop). Run scripts/validate_cards.py with --production, the assigned chapter range, the source, events, and self-check; do not use the cards-only shortcut to sign a chunk. The validator now reports three tiers: ERRORS (true parse failures only — missing field, unreadable file, unparseable card — these still block because a downstream script would crash); WARNINGS; and REVIEW (the human review list). The validator's job is anti-fabrication and field parsing, nothing more:

• It does not require quotes to be verbatim. The 关键原句 field may be paraphrased, truncated, or differ in punctuation — it is for meaning, not proof.
• Fabrication is checked by a separate short 定位锚 (6–15 chars) that must be grep-able in the assigned slice. If the anchor is not found, the card goes to REVIEW — it is not rejected and the chunk is not re-run. Prefer false positives (you dismiss them in three seconds) over false negatives.
• Chapter coverage, self-check length, event-density, and stop thresholds are now REVIEW markers, never automatic stops. Use --evidence-source when the book does not use 第X章 headings.

Then audit semantically. Dispositions are KEEP, REWRITE, MERGE, DOWNGRADE, ESCALATE, DELETE — but in self-use these are recommendations the human acts on, not actions the reviewer takes unilaterally. Default disposition for any semantic/depth concern is ESCALATE (mark, send to human), not DELETE. When the human decides a card needs source re-extraction, rebuild it from events and evidence — do not preserve the old title, route, or count; one old card may yield zero, one, or several new cards. Only Codex writes or approves files in analysis/cards; never silently overwrite v1.

Read references/quality-gates.md and references/extraction-rules.md for entry thresholds, references/review-rubric.md for the audit checklist, and references/templates.md when building prompts. Card and self-check shapes live in assets/.

Mark Depth Decay — Don't Auto-Stop

The validated run showed silent degradation under batch load: the first chunk was complete, the second compressed its self-check, and the third dropped to a title scan. Start conservatively at one calibrated chunk per run. Increase batch size only after a project-local test.

These depth signals are now REVIEW markers, not automatic stops or re-runs. The validator flags them; the human reads the flags and decides whether to补挖 or move on. By default the chunk still ran once and is kept:

• self-check under ~10 lines, or missing its chapter-coverage table;
• any phrase like 高速概览 / 标题扫描 / 框架初稿 / 批量初稿;
• a chunk's event density visibly below the previous one.

A real parse-level placeholder file (empty chapter, "本块……待补" stub) is the one exception that still blocks, because it is malformed, not merely shallow. Do not generate placeholder files for chunks you could not finish — stop after the last complete chunk and tell the human. Spot-audit every 2–3 chunks rather than waiting for the whole book; this spot-audit is the depth net, and it is the human's call.

Run the bundled regression suite after changing templates, rules, or the validator:

python3 -m unittest discover -s tests -v

Improve Constraints — Route, Don't Append

After every audit, for each distinct failure:

1. Decide if it is reusable. One-off text issue → fix that card only. Recurs ≥2 times → make a rule. Can corrupt a quote, fact, event identity, or final count → hard rule immediately.
2. Choose where it lives. Machine-checkable → validator script. Needs semantic judgment → review-rubric. Hard to describe but easy to imitate → a positive/negative sample pair. Do not keep pushing rules into the main prompt.
3. Keep a regression sample. One correct case + one failure case per important rule in analysis/meta/regression. Re-run them after any rule edit.
4. Record cost. Per chunk: model, input length, card count, error count, repair count.

Both 3 and 4 are detailed in references/regression-and-cost.md. Log each failure family in references/error-ledger.md. Version the rule set; review and merge rules every three accepted chunks; retire rules subsumed by a broader one. Constrain unsupported inference, never prescribe conclusions or a fixed card count.

Model Routing

Measured roles from the validated run (detail in references/model-routing.md):

Work	Executor
Chapter split, coverage scan, boundaries, field/quote checks	scripts
Event recall, evidence tables, result marking, format skeleton	map model / mini
First-pass A/B candidates	map model (Pro for locally hard cards only)
Card keep/cut, causal audit, dedup, rewriting	Codex medium
Cross-chunk synthesis, methods, counterexamples, boundaries	Codex medium→high
Whole-book design, hard adjudication, skill definition	Codex high

Never use a light model for final A/B routing, causal structure, motive, method stability, or acceptance. 0 error / 0 warning from the validator does not predict semantic acceptance — it only means fields parse.

Reduce Accepted Cards

Read only signed cards. Assign one stable event ID per independent situation; merge continuations and duplicate perspectives before counting method frequency — multiple cards for one event are one evidence node, not several. Cluster by underlying decision mechanism, not plot topic. Require independent contexts, counterexamples, boundaries, and failure modes. For limited-goal actions, require an unreachable original goal, a replacement target, observable success criteria, and a stopping rule. Keep every final principle traceable to accepted cards and source locations. Escalate to high reasoning when a conclusion spans ≥3 chunks, compares many cards, changes event counts, or enters the final methodology. See references/workflow.md §Reduce.

Tag Every Rule: [method] or [adapt]

Each rule in this skill — in the quality gates, the review rubric, the validator config — carries one of two tags, so the layer that is universal can be told apart from the layer that merely accommodates one model's quirks.

• [method] — true of any model extracting any novel: a closed exchange needs offer + concession + causal link + observable response; cost must be borne by the focal actor; explicit narration outranks inference. These are the real judgment crystallized from reading.
• [adapt] — there only to work around a specific model or Chinese-text quirk: the 6–15-char anchor length, the chapter-number regex, the placeholder phrase list, "Pro for locally hard cards only". These name a model by behavior.

Discriminator: "Would this rule still hold if I swapped the model or the book?" Yes → [method]. It mentions DeepSeek's / mini's behavior, or a punctuation/格式 fix → [adapt].

Why now, in self-use: both tiers are active and you obey both. The tag is the escape hatch for later — when this becomes a product, you re-examine only the [adapt] layer (it may be wrong for the next model) and keep [method] intact. 登楼撤梯：[adapt] is the ladder, tag it now so it can be pulled cleanly later.

Reuse On Another Book

Keep the workflow, four layers, quality gates, validator, rule lifecycle, regression mechanism, cost record, and generic schemas — i.e. keep the [method] layer. Replace the [adapt] layer: chapter parser and chunk size; project lines and entry thresholds; role-specific terminology; candidate method labels; source-specific examples and accepted conclusions.

Start the new book with a calibration chunk. Do not copy accepted conclusions or character assumptions from an earlier book. Blind test: after the methodology is built, extract two chunks of a book that never shaped any rule and check whether the gates and routing hold. Schedule this separately — it is research spend, not deliverable spend.