instrlint

instrlint

Lint and optimize agent instruction files. Produces a scored health report across three dimensions: token budget, dead rules, and structure.

Command Resolution Protocol

When the user runs /instrlint [args]:

1. Detect language — identify conversation language → output: locale

   • 繁體中文 → zh-TW
   • English or any other → en

2. Check --verify — if args contain --verify → jump to LLM Verification Protocol below. Otherwise continue.

3. Map command — parse user args against this table → output: base command

   User args	Base command
   (none)	npx instrlint@latest
   budget	npx instrlint@latest budget
   deadrules	npx instrlint@latest deadrules
   structure	npx instrlint@latest structure
   --fix	npx instrlint@latest --fix
   ci --fail-on warning	npx instrlint@latest ci --fail-on warning

4. Build final command — append flags from steps 1 and 3: {base command} --format markdown --lang {locale} Exception: --fix omits --format markdown (output is a fix summary, not a report).

5. Execute and present — run the command → present the markdown output directly to the user. Never summarize or paraphrase.

LLM Verification Protocol (--verify)

contradiction / duplicate / structure detectors are text heuristics and may produce false positives. You (the host agent) can judge these semantically.

Protocol: instrlint never calls an LLM API. It writes suspicious findings as candidates.json, you judge them and write verdicts.json, then instrlint merges the results back into the report.

Quick steps

1. npx instrlint@latest --emit-candidates instrlint-candidates.json --skip-report --lang <detected>
2. Read instrlint-candidates.json and judge each candidate using criteria from references/judgment-framework.md (confirmed / rejected / uncertain)
3. Write instrlint-verdicts.json with your verdicts (id must match 12-char hex from candidates)
4. npx instrlint@latest --apply-verdicts instrlint-verdicts.json --format markdown --lang <detected>

Splitting Guidance

When instrlint reports either of these conditions, do not paste the suggestion text directly — run the splitting decision walkthrough instead:

• Budget warning: root instruction file exceeds recommended line count (> 200 lines)
• Structure findings contain path-scope suggestions (messageKey: structure.scopePathScoped)

Walkthrough flow

1. Read the file, list major sections (heading + line count)
2. Classify each section using the 4-bucket framework (see references/judgment-framework.md)
3. Present a decision table to the user (section / lines / recommended action / reason)
4. Ask the user: "Minimum split (bucket 1 only) / Full split (buckets 1+2+3) / Custom"
5. Execute: bucket 1 → extract to rules directory with paths: / globs: frontmatter; bucket 2 → extract without path-scope; bucket 3 → replace with one-line pointer