Scene–Concept–Entity Extraction

Input is a batch of linear source texts. Output sce-scenes.yaml, sce-concepts.yaml, and sce-entities.yaml in the user-selected directory, with progress in _sce-round1-progress.md. The value is reuse: one scene can directly invoke a concept established elsewhere.

Principles

1. Source first — define and IPO come from the source and stay within 50 words. Leave unsupported content empty rather than inventing it.
2. Concepts are the hub — Establish concepts first, attach entities to concepts, and let scenes invoke only established concepts and entities.
3. Prefer omission — A candidate without its own IPO or decomposition is not a concept. Merge abstractions into their parent and demote fragments to steps.
4. Bound references — Keep two to six relations per concept and only the strongest relation for a knowledge-element pair.
5. Stop for confirmation between rounds. Never start round two before approval.

Round One: Skeleton

Process five to ten texts per batch and update _sce-round1-progress.md.

1. Read the complete batch and record progress.
2. Extract candidates with only id + define + sources: scenes answer “how do I…”, concepts can express an IPO, and entities answer “did I use X?”
3. Deduplicate globally: merge at 80% or greater semantic overlap, normalize synonymous IDs, and return overly fine steps to their parent. Record every X → Y merge.
4. Write the three YAML files and mark remaining fields TODO.
5. Stop and present deduplication and concept coverage for confirmation.

Round Two: Content

Use sources to return to the original text. Sequence words indicate steps; requirement and input words indicate IPO inputs.

1. Fill from the hub outward: concepts first with complete IPO or decomposition, entities in parallel, then scenes with trigger, goal, and three to six progressive assembly stages.
2. Read references/relations.md; add entity hierarchy and concept–entity invocation relations.
3. Run completeness checks and write updated_at.

Completion requires a skeleton-only round one with a merge log; every concept has IPO or decomposition; every entity passes “I used X” and has hierarchy depth at most three; every scene invokes established concepts/entities in three to six stages; all references resolve; every definition has a source and stays within 50 words; there are no isolated concepts; and every concept has two to six relations.

During round two, read references/relations.md for the direction and selection order of the eight relation types.