Expand description
Evidence for the multi-pass LLM style-guide synthesis.
A single prompt produces a shallow guide. Instead we split the guide into ~10 dimensions and
hand each its OWN tailored evidence (real signatures for the parameters section, the biggest
classes for design patterns, per-library call-sites for library usage, …), then run one focused
LLM call per dimension and stitch the results. The valuable, non-surface conventions live in the
code, so every section is grounded in real excerpts, dependency facts, or measured stats.
COMINDLY_DEBUG_EVIDENCE=1 prints each section’s evidence to stderr.
Structs§
- Section
- One dimension of the guide: a heading, guidance for the LLM, and the evidence it reasons over.
Constants§
- SECTIONS
- Each guide section’s title and what the LLM is asked to cover in it; prompts may replace the latter.
Functions§
- build_
sections - Build every guide section for one repo. Reads source lazily (cached).