CC Opus First Pass
Use this skill to classify a task into the 3-tier system and produce an analysis brief.
Model order (always Opus-first for analysis):
claude-opus-4-6claude-sonnet-4-6opussonnet
3-Tier Classification
Classify into exactly one tier: TINY, MEDIUM, or HIGH-RISK.
The classifier is risk-first: the highest-risk signal present determines the tier.
Risk Signals
| Signal | Low | Moderate | High |
|---|---|---|---|
| Files / modules / layers | 1 file, local path | 2-4 files, 1 module boundary | Crosses modules or layers |
| Auth / security | None | Minor permission check | Trust boundary, secrets, auth flow |
| State / concurrency | None | Simple state update | Shared state, locks, async coordination |
| Schema / migration | None | Additive field | Breaking change or migration |
| Build / config / deps | None | Config value tweak | Build system, CI, or new dependency |
| Public API surface | None | Backward-compatible addition | Breaking change or redesign |
| Blast radius clarity | Obvious | Mostly clear | Hard to reason about |
| Design tradeoffs | None | Minor choice | Significant alternatives |
Decision
- TINY: All signals low. <= 15 lines, 1 file, obvious blast radius. -> Skip this skill; Codex handles directly.
- MEDIUM: Any moderate signal and no high signals. ->
cc-main-implement -Tier medium(Sonnet-first). - HIGH-RISK: Any high signal. -> Full orchestration: plan review ->
cc-main-implement -Tier high-risk(Opus-first) -> checkpoint reviews.
Highest-risk signal wins.
Workflow
- Classify the task.
- If
TINY, skip this skill and proceed in Codex. - If
MEDIUMorHIGH-RISK, invoke the bundled script:
powershell -NoProfile -ExecutionPolicy Bypass -File "scripts/invoke-cc-opus-first-pass.ps1" `
-UserRequest "$ARGUMENTS" `
-WorkingDirectory "$PWD"
- If specific files are already known, include them as context paths:
powershell -NoProfile -ExecutionPolicy Bypass -File "scripts/invoke-cc-opus-first-pass.ps1" `
-UserRequest "$ARGUMENTS" `
-WorkingDirectory "$PWD" `
-ContextPaths "src/app.ts","src/auth.ts"
- Read Claude Code output and use it as a first-pass brief, not as unquestioned truth.
- Codex still owns:
- final implementation decisions
- actual edits
- verification
- final user-facing explanation
Failure Handling
If Claude Code is unavailable, not authenticated, or the command fails:
- note the failure briefly
- continue locally in Codex
- do not block the task solely because the external first pass is unavailable
Expected Claude Output
The script asks Claude Code for:
- Tier Verdict: TINY, MEDIUM, or HIGH-RISK
- Risk Signal Assessment (which signals triggered and at what level)
- Why (concise justification)
- Likely Affected Files
- Implementation Plan
- Risks and Edge Cases
- Suggested Tests
- Patch Sketch
Use that output to guide the Codex implementation and review pass.