ts-idea2story

ts-idea2story — idea → grounded research story (the upstream of ts-paper)

Distilled from the product's idea→story link. The idea is the PROTAGONIST; a recalled/searched pattern is the TOOL it wields. Claude is the packager, the recaller, the searcher, the storyteller, and the critic — all in one context (so coherence and anti-stacking come for free). Two scripts are irreducible: kg_recall.py (vector/graph retrieval) and story_lint.py (the gate); embeddings (optional) reuse ../ts-kg-build/scripts/embed.py's TS_EMBED_* config.

Inputs / Outputs

In: a raw idea (text); optional kg/ dir (from ts-kg-build, or the bundled kg_ts); optional TS_EMBED_* endpoint; a retrieval_focus dial in {trust_kg, balanced, go_search_web}. Out (in the workdir):

• story.json — the 8 fields: title, abstract, problem_framing, gap_pattern, solution, method_skeleton, innovation_claims[], experiments_plan.
• story_proposal.md — the Markdown projection of the 8 fields (this is what ts-paper-plan reads — a story IS a structured proposal; story2paper == proposal2paper).
• retrieved_papers.json — the real papers found, the citation seed for ts-paper-cite (§ Reuse).
• logs/*.io.md, optional novelty_report.json.

Procedure

1. Package the idea (Claude)

Write idea_brief.json: motivation, problem, assumptions_explicit[], assumptions_inferred[], constraints, retrieval_query. Faithfulness guard: keep assumptions you inferred separate from what the user stated — never fabricate user intent. The retrieval_query is an English, search-friendly reformulation. Never raise on a thin idea; normalize and proceed.

2. Recall candidate patterns (script + Claude)

If a kg/ exists: python3 scripts/kg_recall.py --idea "<retrieval_query>" --kg <kg_dir> --out recall.json --topk 8. It returns Top-K candidate patterns (semantic+graph if embeddings configured, else lexical-only, labeled — treat lexical hits as weaker). Then YOU reason over them comparatively in one pass, scoring each on three independent axes (no per-pattern round-trips needed):

• stability — reliable skeleton, well-trodden; novelty — differentiating, rare;
• domain_distance — how far the pattern's home domain is from the idea (sorted ASCENDING; a near-domain pattern transfers more safely, a far one is a riskier "storyteller" move). Size ≠ stability (a big cluster can be incoherent — check the pattern's coherence). If no KG: skip to search (or, with neither, generate the story from the idea alone — weakest, allowed).