compare

Cross-Model Comparison Skill

Compare trigger phrase generation across model sizes. Same file, same prompt, three models.

Tools

• Agent (to spawn compare-researcher)
• Read, Edit (to store results)
• Glob (to resolve directories)

Process

1. Resolve Input

• File path: Use directly
• Directory: Glob for relevant files (*.{ts,tsx,js,jsx,py,go,rs,md,json,yaml,sh,sql,css,scss,html,svelte,vue}), exclude node_modules/, dist/, build/, .git/, *.lock, *.min.*
• No argument: Ask the user what to compare
• Cap at 10 files per run (15 agent calls per file)

2. Invoke Compare Researcher

Spawn the compare-researcher agent:

Use Agent tool with:
  subagent_type: total-recall:compare-researcher
  prompt: |
    Compare trigger phrase generation across models for these files:
    <list each absolute file path, one per line>

3. Store Results

Read .claude/compare-results.json if it exists, otherwise create it. Append a new run entry:

{
  "version": 1,
  "runs": [
    {
      "runAt": "<ISO timestamp>",
      "files": {
        "relative/path.md": {
          "models": {
            "haiku":  { "phrase": "...", "samples": [...], "convergence": [...], "confidence": 0.0 },
            "sonnet": { "phrase": "...", "samples": [...], "convergence": [...], "confidence": 0.0 },
            "opus":   { "phrase": "...", "samples": [...], "convergence": [...], "confidence": 0.0 }
          },
          "crossModelConvergence": [...],
          "modelSpecificTerms": { "haiku": [...], "sonnet": [...], "opus": [...] }
        }
      }
    }
  ]
}

Use relative paths as keys (relative to project root).

4. Display Results

Show a comparison table:

## Cross-Model Comparison Results

| File | Haiku | Sonnet | Opus | Shared Terms |
|------|-------|--------|------|--------------|
| testing.md | test behavior not implementation (0.99) | ... (0.xx) | ... (0.xx) | behavior, implementation |

### Analysis
- Files where all models agree: [list] — these are stable attractors
- Files where models diverge: [list] — larger models may surface deeper associations
- Average confidence: haiku X.XX | sonnet X.XX | opus X.XX

5. Do NOT rebuild the recall index

This is experimental data. Production triggers in triggers.json are not affected.