メニュー
Run cross-model comparison — scan files with haiku, sonnet, and opus study agents to see how model size affects trigger phrase generation.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Full operating manual for working in a codebase with the user. Use this skill at the start of any substantial task, when the user asks how to approach work, or when setting up a collaboration session. Covers the core craft principles, user interaction model, rule conflict resolution, code quality (broken windows), when to stop, and which rules apply for each task type. Read in full at every load. This skill defines the substrate the methodology operates from; partial reading or summarization produces no substrate. Use proactively — don't wait to be asked.
Self-eval the agent runs when it suspects memory-vs-recall drift — the failure mode where memory-citation substitutes for fresh receipt-fetching and pattern-copying outruns reachability analysis. Outputs a tiered recommendation (continue / slow down / refresh) backed by observable artifacts, not introspective claims alone. Invoke when: - You notice yourself reaching for "I remember from earlier" framings repeatedly - You catch yourself copying a pattern from a sibling module without walking conditions - You make the same class of mistake twice in close succession - User suggests fatigue ("how are you feeling?", "still sharp?") - End of long working stretch - User invokes explicitly via `/check-yo-self`
Quick reference for which rules govern the current work. Consult mid-task when starting a new task type, touching unfamiliar code, or unsure which constraints apply. Pairs with the-craft skill (full manual). Trigger on: "which rules apply", "what rules should I follow", "check constraints", or proactively when switching task type mid-session. Also use when any two rules appear to conflict — the always-on rules explain precedence.
Remove trigger phrases for a file or glob pattern from .claude/triggers.json
Display all stored trigger phrases from .claude/triggers.json
Study files and generate semantic trigger phrases by spawning a researcher agent that samples model associations and finds convergent patterns. Supports model-specific triggers via --models flag.
| name | compare |
| description | Run cross-model comparison — scan files with haiku, sonnet, and opus study agents to see how model size affects trigger phrase generation. |
Compare trigger phrase generation across model sizes. Same file, same prompt, three models.
*.{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.*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>
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).
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
This is experimental data. Production triggers in triggers.json are not affected. Results in .claude/compare-results.json are not consumed by other skills — they exist for research and manual inspection.