All Hands

Dispatches all 9 audit roles in parallel. Each role does a fast scan and returns their top 3-5 recommended action items for the week -- not a full audit, just the most important things they see right now.

Roles

The 9 roles are:

1. audit-engineering -- CTO: architecture, code quality, test coverage, CI, tech debt
2. audit-security -- CISO: auth, secrets, injection, dependencies, credential exposure
3. audit-criticism -- Devil's advocate: strategy, unit economics, self-deception, moat
4. audit-design -- CDO: visual identity, typography, color, component coherence
5. audit-ux -- CXO: user flows, accessibility, onboarding, friction, error states
6. audit-customer -- Customer advocate: PMF, adoption friction, churn signals, feedback
7. audit-financial -- CFO: spend caps, margin, SaaS creep, runway
8. audit-legal -- Head of Legal: missing documents, compliance gaps, marketing claims
9. audit-marketing -- CMO: positioning, messaging, conversion, brand coherence

Dispatch Process

Step 1 -- canary. Dispatch audit-engineering alone first. Give it this prompt:

You are running a lightweight weekly priorities scan, not a full audit. Read the codebase quickly. Return ONLY your top 3-5 action items for this week, each as a single bullet: [P0/P1/P2] <role>: <action item in one sentence>. The R-804 evidence requirement (pasted code with file:line) is waived for this format; each bullet still names the file, directory, or surface it is about, and resolves precedence before flagging (a documented override is not a finding). Do not write a full audit report. Do not write to any file. Return the bullets only.

Wait for clean return. If the canary errors or produces garbage output, stop and report the error to the user before fanning out.

Step 2 -- fan out. Dispatch the remaining 8 roles in a single parallel batch, each with the same lightweight prompt above (substituting their role name). Pass model: sonnet on every dispatch: the role file's own model line (Opus for engineering, security, criticism, and customer) is the default for a full audit, not for this scan, and the Agent tool applies the role file's model unless the call overrides it.

Step 3 -- consolidate. Collect all bullet lists. Deduplicate overlapping items (same issue flagged by multiple roles -- keep the highest-priority instance and note how many roles flagged it). Sort by priority (P0 first, then P1, then P2). Write the consolidated output.

Output Format

Write to docs/audits/YYYY-MM-DD-all-hands.md:

# All Hands -- YYYY-MM-DD

## Top Priorities This Week

| Priority | Role | Action Item |
|----------|------|-------------|
| P0 | Security | Rotate the leaked API key in git history |
| P0 | Financial | Set a hard spending cap on Anthropic API |
| P1 | Engineering | Fix the 3 unpaired fix: commits with no test change |
| ... | | |

## Per-Role Recommendations

### Engineering (CTO)
- [P1] ...
- [P2] ...

### Security (CISO)
- [P0] ...

### Criticism (Devil's Advocate)
- [P1] ...

### Design (CDO)
- [P2] ...

### UX (CXO)
- [P1] ...

### Customer Advocate
- [P2] ...

### Financial (CFO)
- [P0] ...

### Legal (Head of Legal)
- [P1] ...

### Marketing (CMO)
- [P2] ...

## Notes

_Items flagged by multiple roles are consolidated in the top table. The per-role sections show each role's raw output before deduplication._

Commit the file to the current branch. Do not create a separate branch. Route the items per R-802: a P0 goes to the current effort now; a P1 or P2 waits for that role's full audit, or goes to ISSUES.md with its priority tag when no audit is scheduled.

What This Is NOT

This is a 15-minute pulse check, not a deep dive. Each role is scanning quickly and returning their most visible concerns. If a role flags something that needs a full investigation, the follow-up is to dispatch that role's dedicated audit agent (e.g. the audit-security agent for a full security report). These roles are agents, not skills; there is no /audit-security slash command.

Model

Sonnet for every role, passed as model: sonnet on each dispatch (the role files set Opus for the four standing and customer roles, and that default stands for a full audit). This is pattern-matching against a clear rubric, not deep architectural reasoning. If the user explicitly requests Opus for a specific role, override that role only; the task-start Model Routing table defers to this skill for the scan.