Instruction Lint (Model-Based Review)

Review agent and skill instructions against rules derived from the Claude Opus 4.6 and Sonnet 4.6 system cards. Combines a fast regex pre-pass with deep model-based semantic review.

---

Step 1: Read the Rules

Read the lint rules rubric:

Read docs/instruction-lint-rules.md

This contains 12 rules in 3 tiers:

• Universal (U-*): Apply to all agents/skills regardless of model
• Opus-specific (O-*): Address Opus 4.6's documented over-exploration and efficiency issues
• Sonnet-specific (S-*): Leverage Sonnet 4.6's documented steerability advantages

---

Step 2: Run Regex Pre-Pass (optional baseline)

Run the fast regex linter for a structural baseline:

uv run python scripts/lint-instructions.py

Note which files have structural issues. These are heuristic — the model review in Step 3 is authoritative.

---

Step 3: Model-Based Review

For each model tier, spawn a review agent that reads the actual instruction files and evaluates them semantically against the rules. The agent should understand INTENT, not just keyword presence.

Parse $ARGUMENTS:

• No args → review all plugins
• Plugin name (e.g., go-dev) → review only that plugin
• opus / sonnet / haiku → review only agents using that model

Agent prompt template

For each batch of files, spawn an Agent with:

You are reviewing Claude Code plugin instructions for quality against
rules derived from the Opus 4.6 and Sonnet 4.6 system cards.

## Rules (apply based on model in frontmatter)

### Universal (all models)
- U-SCOPE: Must have clear scope boundaries (what's in, what's out)
- U-OUTPUT: Must define expected output format
- U-TOOL-FIRST: If agent has Bash, must require running tools before manual analysis
- U-FAILURE: Must handle failure/impossibility (prevents over-eager workarounds)
- U-GROUND: Must instruct to ground claims in actual tool output
- U-NO-DESTROY: If agent has Bash, must warn about destructive actions

### Opus agents (model: opus)
- O-EFFICIENCY: Must include efficiency constraints (Opus over-explores)
- O-SCOPE-ONLY: Should have "ONLY these" or "exclusively" markers
- O-EFFORT-MATCH: effort:high must be justified by complex multi-dimensional tasks

### Sonnet agents (model: sonnet)
- S-NO-LECTURE: Must NOT contain lecture-inducing patterns (Sonnet tends to lecture)
- S-DECISIVE: Should include decisive action language
- S-ANTI-EAGER: Should include anti-over-eagerness (Sonnet is steerable here)

## Review these files

[list of file paths]

For each file:
1. Read it fully
2. Note the model from frontmatter
3. Apply the matching rules SEMANTICALLY — check intent, not just keywords
4. Rate each applicable rule: PASS / WARN / FAIL
5. For WARN/FAIL: explain specifically what's missing and suggest a fix

## Output Format

For each file, output:

### `<relative path>` (model: <model>, kind: <agent|skill>)
| Rule | Verdict | Notes |
|------|---------|-------|
| U-SCOPE | PASS/WARN/FAIL | ... |
...

Then at the end, output a summary:
- Total files reviewed
- Pass/warn/fail counts per rule
- Top 5 most impactful improvements to make

Batching strategy

• Batch files by plugin (one agent per plugin)
• Launch agents in parallel where possible (up to 3 concurrent)
• Each agent reads its batch of files using Read tool

---

Step 4: Aggregate and Present

Collect results from all review agents. Present:

1. Summary table: files x rules matrix
2. Critical findings: Any FAIL verdicts with specific fix suggestions
3. Top improvements: Ranked by impact (how many files affected x severity)
4. Model-specific patterns: Are opus agents missing efficiency guards? Sonnet agents missing anti-eagerness?

---

Output

Present findings as a structured report:

## Instruction Lint Report

### Summary
- Files reviewed: N (X opus, Y sonnet, Z haiku)
- Regex pre-pass: N errors, N warnings
- Model review: N pass, N warn, N fail

### Critical Findings
1. ...

### Top 5 Improvements
1. ...

### Per-Plugin Results
...