enhance

You are running auk's deep pass. auk has already extracted the structural facts (tree-sitter parsing, call graph, clusters). Your job is the semantic layer: turn its generic, machine-mined rules into sharp, codebase-specific guidance — the part only an agent that can read the code can do.

Do this end to end, without asking the user to copy-paste anything:

1. Emit the work. Run npx -y auk-develop generate --no-compile --emit-prompts in the project root. This writes .auk/prompts/enhance-rules-NN.md batches.

2. Read every batch file under .auk/prompts/. Each contains rules, code excerpts, the project overview, and a response JSON schema.

3. Do the analysis yourself. For each rule, open the evidence files (and any neighbours you need) and rewrite the description so it is specific to THIS repo: name the actual directories, types, frameworks, and functions. Add a rationale explaining why the convention exists here and what breaks if it's violated. Prefer concrete instructions an agent can act on when writing new code. Process batches in parallel where practical.

   • Good: "HTTP handlers in backend/internal/handlers/ take *gin.Context, validate with ShouldBindJSON, and delegate to a service in internal/services/ — never put DB calls in a handler."
   • Bad: "Use consistent error handling."

4. Write one enhance-rules-NN.response.json per batch, matching the schema exactly. Only reference files that exist. Keep descriptions 10–600 chars.

5. Apply each: npx -y auk-develop enhance --apply .auk/prompts/enhance-rules-NN.response.json. If auk reports skipped rules, fix that response and re-apply.

6. Compile: npx -y auk-develop compile to refresh CLAUDE.md, AGENTS.md, .cursor/rules, copilot-instructions, etc. with the enhanced text.

7. Summarize what you sharpened, grouped by category.

Note: auk's deterministic verification logic is never touched by this — you are only rewriting human-facing descriptions and rationales. The rules stay machine-checkable by auk verify.