Spec Auditor
teterouge/spec-auditor
Auditoria adversarial de specs em 5 passes antes da engenharia
Ver no GitHubSkill que realiza auditoria adversarial de especificações, PRDs e feature briefs em cinco passes estruturados antes de chegar à engenharia. Identifica bloqueadores, critérios de aceitação não testáveis, ambiguidades, suposições ocultas e riscos de scope creep — funcionando como um revisor sênior automatizado em sprint planning.
Instalação
npx convez-ai teterouge/spec-auditor1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Revisar PRDs e feature briefs antes de sprint planning
- Validar testabilidade de critérios de aceitação
- Identificar ambiguidades e suposições ocultas em especificações
- Mitigar scope creep e dependências não confirmadas
- Garantir completeness de especificações técnicas
Veredito

Ferramenta bem estruturada e pragmática para times de produto que enfrentam specs incompletas. O design em cinco passes (completeness, testability, ambiguity, assumptions, scope) reflete problemas reais de sprint planning. Recomendado para equipes que priorizam documentação de qualidade e querem economizar ciclos de rework.
README
A tradução para o português deste README ainda está na fila.
spec-auditor
An adversarial spec review plugin for Claude Code. Audits product specs, PRDs, and feature briefs across five passes before they reach engineering — surfacing blockers, untestable acceptance criteria, ambiguities, hidden assumptions, and scope creep risks.
The goal is to catch everything a senior engineer or QA lead would raise in sprint planning, before sprint planning happens.
Why This Exists
Most spec reviews are superficial. Someone reads it, says "looks good," and engineering picks it up. The problems surface mid-sprint: a missing error state, an untestable AC, a dependency nobody confirmed, a scope phrase that expands to three weeks of work.
This skill runs a structured adversarial audit — not a proofreading pass. It asks the questions nobody wants to hear until it's too late to ask them cheaply.
What It Does
Five sequential audit passes, each targeting a different failure mode:
| Pass | What It Finds |
|---|---|
| 1. Completeness | Missing required fields — problem statement, success metrics, out-of-scope, error states, dependencies, rollback |
| 2. Testability | Acceptance criteria a QA engineer can't write a deterministic test for — rewrites each one |
| 3. Ambiguity | Language two engineers would interpret differently — pronoun drift, undefined terms, implicit conditionals, conflicting requirements |
| 4. Assumptions | Things the spec silently assumes — system behavior, data availability, permissions, platform, state/sequence |
| 5. Scope Risk | Unbounded requirements that will expand in implementation — flags each with min/max interpretation range and a bounded rewrite |
Every finding gets a severity rating:
- 🔴 Blocker — Sprint should not start without resolving this
- 🟡 Warning — Sprint can start but mid-sprint confusion is likely
- 🔵 Note — Worth addressing; won't block delivery
The output ends with a Sprint Readiness Verdict: Ready / Ready with fixes / Not ready.
Installation
Marketplace (coming soon)
spec-auditor will be available through the pm-discipline Claude Code marketplace, which will aggregate several product-management-focused skills.
Once available:
/plugin marketplace add teterouge/pm-discipline
/plugin install spec-auditor
Or manually:
git clone https://github.com/teterouge/spec-auditor
claude --plugin-dir ./spec-auditor
Usage
The skill triggers automatically. Just share a spec and ask if it's ready:
"Can you review this PRD before I send it to engineering?"
"Is this spec ready to go into the sprint?"
"Take a look at these user stories and tell me what's missing."
"Quick gut check on this feature brief."
You can also upload a spec file directly. If a .md, .txt, .html, or .docx file is attached, the spec_parser.py script will extract structured elements before the audit begins.
Structure
spec-auditor/
├── .claude-plugin/
│ ├── plugin.json ← Plugin manifest
│ └── marketplace.json ← Marketplace catalog
├── skills/
│ └── spec-auditor/
│ ├── SKILL.md ← Core audit instructions
│ ├── references/
│ │ ├── audit-criteria.md
│ │ ├── acceptance-criteria-patterns.md
│ │ └── severity-guide.md
│ ├── scripts/
│ │ └── spec_parser.py
│ └── evals/
│ └── evals.json
└── README.md
Design Notes
Adversarial, not collaborative. Calibrated to find problems, not validate the work. A spec that passes all five passes clean has earned that verdict.
Severity avoids over-flagging. Not everything is a blocker. High-risk features (payments, auth, irreversible actions) have a lower escalation threshold.
Rewrites, not just flags. Pass 2 and Pass 5 model what fixed looks like — original and rewrite side by side.
Part of the pm-discipline Suite
- spec-auditor ← you are here
- feature-roi-forecaster — Financial business case before the spec is written
- post-launch-retro — Closes the loop between hypotheses and reality
- pmf-signal-analyzer — Multi-signal product-market fit assessment
- dependency-risk-mapper — Cross-team dependency graph and delivery risk
- competitive-intel-synthesizer — Competitive signals to roadmap implications
- pricing-packaging-intel — Pricing and packaging recommendation engine
- stakeholder-comm-translator — Audience-specific message translation# spec-auditor
An adversarial spec review plugin for Claude Code.
spec-auditor audits product specs, PRDs, and feature briefs across five structured passes before they reach engineering — surfacing blockers, untestable acceptance criteria, ambiguities, hidden assumptions, and scope creep risks.
Most product failures don't originate in code. They originate in specifications that hide assumptions, edge cases, and operational constraints. This skill is designed to surface those issues while they are still cheap to fix.
The goal is to catch everything a senior engineer, QA lead, or systems thinker would raise in sprint planning — before sprint planning happens.
Why This Exists
Most spec reviews are superficial.
Someone reads the document, says “looks good,” and engineering picks it up. The real problems surface mid-sprint:
- a missing error state
- an untestable acceptance criterion
- a dependency nobody confirmed
- a scope phrase that quietly expands to three weeks of work
These issues are rarely caught because the review process is collaborative and polite.
spec-auditor runs a structured adversarial audit instead.
It deliberately interrogates the document from the perspective of engineering, QA, operations, and failure modes — asking the questions nobody wants to hear until it's too late to ask them cheaply.
What It Does
The skill runs five sequential audit passes, each targeting a different class of specification failure.
| Pass | What It Finds |
|---|---|
| 1. Completeness | Missing required fields — problem statement, success metrics, out-of-scope definition, error states, dependencies, rollback strategy |
| 2. Testability | Acceptance criteria a QA engineer cannot write a deterministic test for — includes rewritten AC examples |
| 3. Ambiguity | Language two engineers would interpret differently — pronoun drift, undefined terms, implicit conditionals, conflicting requirements |
| 4. Assumptions | Silent assumptions about system behavior, data availability, permissions, platform constraints, state/sequence logic |
| 5. Scope Risk | Unbounded requirements likely to expand during implementation — flagged with interpretation range and bounded rewrite |
Every finding receives a severity rating:
- 🔴 Blocker — Sprint should not start without resolving this
- 🟡 Warning — Sprint can start but confusion or rework is likely
- 🔵 Note — Worth addressing; unlikely to block delivery
The report concludes with a Sprint Readiness Verdict:
- Ready
- Ready with fixes
- Not ready
Installation
Marketplace install
# Add this repository as a marketplace
/plugin marketplace add <your-github-username>/spec-auditor
# Install the plugin
/plugin install spec-auditor@pm-discipline
Manual install
git clone <repo-url>
claude --plugin-dir ./spec-auditor
Usage
The skill triggers automatically during spec discussions.
Example prompts:
"Can you review this PRD before I send it to engineering?"
"Is this spec ready to go into the sprint?"
"Take a look at these user stories and tell me what's missing."
"Quick gut check on this feature brief."
You can also upload a spec file directly.
If a .md, .txt, .html, or .docx file is attached, the spec_parser.py script extracts structured sections before the audit begins.
Repository Structure
spec-auditor/
├── .claude-plugin/
│ ├── plugin.json
│ └── marketplace.json
├── skills/
│ └── spec-auditor/
│ ├── SKILL.md
│ ├── references/
│ │ ├── audit-criteria.md
│ │ ├── acceptance-criteria-patterns.md
│ │ └── severity-guide.md
│ ├── scripts/
│ │ └── spec_parser.py
│ └── evals/
│ └── evals.json
└── README.md
Design Notes
Adversarial by design.
The tool is calibrated to find problems, not validate the work. A spec that passes all five passes clean has earned that verdict.
Severity prevents noise.
Not every issue is a blocker. High-risk systems (payments, auth, irreversible actions) escalate faster than routine features.
Rewrites instead of vague criticism.
Where possible, the audit produces concrete rewrites — particularly for acceptance criteria and scope boundaries.
Focus on pre-engineering clarity.
The goal is not stylistic improvement. The goal is eliminating ambiguity and hidden scope before implementation begins.
Part of the pm-discipline Suite
spec-auditor is one component of a broader set of tools for operationalizing product discipline.
- spec-auditor ← adversarial spec review
- feature-roi-forecaster — financial business case before the spec exists
- post-launch-retro — closes the loop between hypotheses and outcomes
- pmf-signal-analyzer — multi-signal product-market fit assessment
- dependency-risk-mapper — cross-team dependency graph and delivery risk
- competitive-intel-synthesizer — competitive signals to roadmap implications
- pricing-packaging-intel — pricing and packaging recommendation engine
- stakeholder-comm-translator — audience-specific message translation