the-rabak/compound-engineering-plugin/adversarial-plan-audit

Adversarial Plan Audit

Run a rigorous pre-implementation audit for architecture and plan quality, not a superficial review. Produce a scored, evidence-cited report that pressure-tests viability, operational readiness, and differentiation.

Announce at start: "I'm using the adversarial-plan-audit skill to run a scored adversarial architecture and plan review."

When to Use

• Evaluate a major architecture or implementation plan before building
• Stress-test a memory/plugin/agent system intended for daily use
• Decide whether a design is ready for implementation or needs redesign
• Generate investor-grade or leadership-grade recommendation artifacts

When Not to Use

• Fix isolated code bugs with clear root causes
• Perform style-only PR review
• Produce implementation plans from scratch without existing artifacts

Required Inputs

• Primary architecture/plan/contracts documents
• Any benchmark/quality gates and constraints
• Target output location for final assessment artifact

If scope is ambiguous, ask for the target decision this audit should support (go/no-go, re-scope, or rank options).

Core Workflow

Step 1: Build the Audit Matrix

Load references/audit-rubric.md and convert scope into explicit tracks:

1. Action parity
2. Tools as primitives
3. Context injection
4. Shared workspace
5. CRUD completeness
6. UX/feedback integration (including CLI and artifact visibility)
7. Capability discovery and onboarding
8. Prompt-native extensibility
9. Adversarial risk and contradiction analysis
10. Market positioning and differentiation

Apply weighting from the rubric and predefine scoring scale before collecting evidence.

Step 2: Run Parallel Adversarial Tracks

Launch track investigations in parallel with subagents.

Dispatch rules:

1. Run up to platform concurrency limits (typically 4 concurrent). Use waves if needed.
2. Use specialized reviewers where available for architecture, security, performance, and data integrity.
3. Resolve local named agents before specialist dispatch:

• Verify the agent metadata/source, then dispatch the resolved subagent itself.
• Do not include the agent file body in the dispatched prompt; pass only the audit track payload.

1. Require each track to return:

• numeric score in X/Y and percentage
• concrete evidence with file/section citations
• top gaps
• high-impact recommendations first

Step 3: Add Dedicated Red-Team and Market Passes

After core principle tracks complete:

1. Run red-team pass:

• identify likely adoption failures
• identify hidden coupling and contradiction risks
• rank by impact x likelihood

1. Run market pass:

• benchmark against credible competitors
• classify differentiators vs commodity areas
• identify strategic opportunities and traps

Step 4: Synthesize, Normalize, and Prioritize

1. Normalize scores into a single weighted composite.
2. Resolve contradictions across tracks; do not average conflicting claims blindly.
3. Produce a prioritized recommendation list using:

• impact on trust/adoption
• effort
• risk reduction
• dependency order

1. Classify strengths separately from gaps.

Step 5: Produce a Reusable Assessment Artifact

Load references/report-template.md. Write the full report to a persistent docs path:

• Prefer docs/assessements/ if that path exists in the repository.
• Otherwise use docs/assessments/.

Name format:

YYYY-MM-DD-<project-or-topic>-adversarial-assessment.md

The report must include:

1. Executive verdict and readiness rating
2. Per-principle score table with status badges
3. Adversarial risk matrix
4. Market positioning analysis
5. Top prioritized recommendations
6. Clear next-step decision framing (go / re-scope / redesign)

Step 6: Return Tight Summary to User

Return concise summary in chat:

• overall score and confidence
• biggest 3 strengths
• biggest 3 blockers
• immediate next action

Do not paste full report in chat when an artifact file already exists.

Output Quality Bar

• Cite all critical claims to source docs or credible external references.
• Avoid generic advice; bind recommendations to observed evidence.
• Distinguish design defects from planned non-goals.
• Flag uncertainty explicitly when evidence is weak.
• Prefer smallest high-leverage interventions before large redesigns.

Failure Handling

• If subagent limits block full parallelism, run in waves and continue.
• If external research hits rate limits, proceed with available internal evidence and clearly mark confidence.
• If required docs are missing, stop and request exact missing artifacts.

Success Criteria

• [ ] Core and adversarial tracks completed
• [ ] Numeric scored output generated per track
• [ ] Final artifact written to repository docs path
• [ ] Recommendations prioritized by impact
• [ ] Summary delivered with explicit decision framing