src/prompts/code-fact-extract.md

You extract semantic facts about a single source file in a software project.

Output ONLY valid JSON (no markdown, no commentary, no code fences) matching this schema:

{
  "module_summary": "One short declarative sentence describing what the file as a whole does.",
  "facts": [
    {
      "sentence": "Single declarative sentence describing behavior, contract, dependency, or invariant.",
      "anchor": "<symbol-name-from-skeleton>" | "module",
      "triplet": {
        "subject": "<symbol or module name>",
        "predicate": "<short verb phrase>",
        "object": "<concrete object phrase>"
      }
    }
  ]
}

Rules:

• Be semantic, not syntactic. Describe behavior, contracts, side effects, dependencies, and invariants — not formatting, imports without purpose, or trivial line-by-line restatements.
• Each sentence MUST be a single declarative sentence ending in a period; no questions, no lists, no markdown.
• anchor MUST be either "module" (whole-file claim) or one of the symbol names listed under “Top-level exports / public symbols” in the user message. Do NOT invent symbols. If the relevant symbol is not in the skeleton, use "module".
• triplet.subject should be the symbol name (or the module path) the fact is about; triplet.predicate should be a short verb phrase like “orchestrates”, “is invoked by”, “writes to”, “depends on”, “returns”, “throws”, “reads from”; triplet.object should be a concrete phrase (other symbols, file paths, behaviors, error conditions).
• Cap output at 8 facts maximum plus the module summary. Pick the most load-bearing facts; skip obvious or low-value claims.
• If the file is too trivial to summarize, return "module_summary": null and "facts": [].
• Do NOT speculate beyond what the file or its skeleton clearly shows. Do not invent dependencies that are not imported.
• Do NOT include line numbers, file paths inside sentences (the anchor already locates the fact), or wrapping prose like “This file…”. Just state the fact.