SuperAPI AI Directory

spec-review

General & other

View repository

Updated 10-16 06:16

About this skill

Spec Review Skill

Purpose

Run 3 focused review agents against a feature's spec files, then consolidate into an actionable PASS / NEEDS REVISION report.

Agents run sequentially by default. Run in parallel only if the user explicitly requests it.

Agent count by mode:

  • Quick mode (plan.md only): 2 agents (skip Agent 3 — Buildability)
  • Full mode (all 3 files present): 3 agents
  • Partial Full mode (some files missing): 2 agents (skip Agent 3)
Agent Name Covers
1 Spec Quality Business validity, EARS correctness, 10-pattern ambiguity smell scan
2 Completeness Missing scenarios, safety invariants, liveness properties, testability scoring
3 Buildability Compatibility, implementation blockers, REQ→Design→Tasks traceability

Activation Triggers

  • "review the spec", "review feature spec", "check the spec"
  • "is this spec good?", "validate the spec", "audit the spec"
  • Before moving from planning to implementation
  • /dev-workflow:spec review

Step 1 — Locate the Spec

  1. List docx/features/ to find all feature directories
  2. If a feature name was given as argument, match it
  3. If ambiguous, show the list and ask which feature to review
  4. If docx/features/ does not exist, stop: "No feature specs found. Run /dev-workflow:spec create first."
  5. Identify which files exist and record their absolute paths — do NOT read contents into context. Agents read files themselves.
    • Quick mode: plan.md
    • Full mode: requirements.md, design.md, tasks.md (note which are present)
    • Set traceability_eligible = true only if all three Full mode files exist

Dispatch Prompt Template

Each agent prompt must be self-contained. Include:

  1. Role: "You are reviewing a feature spec for [agent name]."
  2. File paths: Absolute paths — tell the agent to read them with the Read tool. Do NOT paste content inline.
  3. Checklist: The check items for that agent (from Step 2 below)
  4. Output format: PASS or NEEDS REVISION, then labeled findings with severity icons

Step 2 — Run Review Agents

Default: sequential (1 → 2 → 3). Summarize each agent's findings in one line before the next.

Parallel: Only if user explicitly says "run in parallel" — emit all Agent tool calls in a single message.


Agent 1: Spec Quality

Covers: Business validity + EARS correctness + ambiguity smell scan

Prompt includes: absolute paths to all spec files

Business:

  • Problem being solved clearly stated?
  • Value / benefit to user or system explicit?
  • Goals realistic and well-scoped?
  • Gold-plated or out-of-scope requirements?
  • Acceptance criteria meaningful?
  • Unstated business assumptions?

EARS Correctness (Full mode):

  • Every requirement uses proper EARS syntax (WHEN/THEN, WHILE, IF/THEN, WHERE, ubiquitous SHALL)
  • Each requirement is one atomic statement (no compound AND)
  • Active voice, SHALL for mandatory, SHOULD for desirable
  • No contradicting requirements

Plan Correctness (Quick mode):

  • Tasks unambiguous and actionable
  • Clear done criteria per task

Ambiguity Smell Scan — flag every occurrence:

Smell Pattern
Vague intensifier "quickly", "appropriate", "reasonable", "good", "user-friendly"
Implicit actor "the system" with no named role
Combinatorial explosion A and/or B
Ambiguous pronoun "it", "they", "this" with unclear referent
Missing subject Passive voice with no actor
Unbounded quantifier "all", "every", "any" without scope
Escape clause "where possible", "if applicable", "as needed"
Unverifiable adjective "secure", "reliable", "scalable" without metric
Temporal vagueness "soon", "eventually", "periodically"
Implicit assumption Behavior implied but not stated

For each smell: cite requirement ID, name the smell, suggest a concrete rewrite.

Return: PASS or NEEDS REVISION + severity-labeled findings


Agent 2: Completeness

Covers: Missing scenarios + safety/liveness invariants + testability scoring

Prompt includes: absolute paths to all spec files

Completeness:

  • Happy path covered?
  • Error / failure scenarios?
  • Edge cases: empty inputs, zero values, max bounds, concurrent access?
  • Auth / permission requirements?
  • NFRs: performance, security, scalability?
  • Data validation rules?
  • Rollback / undo behavior (if relevant)?
  • All actors / user roles accounted for?
  • Implied behaviors not written down?

Safety Invariants — what must NEVER be true?

  • State: "shall never allow X while Y"
  • Security: "unauthorized users shall never access Z"
  • Data integrity: "balance shall never go negative"
  • Concurrency: "two requests shall never simultaneously modify the same record"

Flag specs with async/stateful/multi-actor behavior that have no safety boundaries stated.

Liveness — what must EVENTUALLY happen?

  • No infinite-wait guarantees
  • Every pending request eventually gets a response or timeout
  • Every queue eventually drains

Testability (0–3 per requirement):

Score Meaning
3 Automatable — clear oracle, deterministic
2 Manual-testable — defined procedure
1 Partially testable — some aspects vague
0 Not testable — vague or no pass/fail criteria

For each 0 or 1: explain the blocker and suggest a rewrite that raises the score.

Return: PASS or NEEDS REVISION + missing scenarios + invariants + testability scorecard + avg score


Agent 3: Buildability

Covers: Compatibility + implementation blockers + traceability

Only dispatch when traceability_eligible = true (all 3 Full mode files present). Skip for Quick mode and partial Full mode.

Prompt includes: absolute paths to all spec files

Compatibility:

  • Design aligns with existing tech stack?
  • Breaking changes to existing interfaces?
  • Dependencies available and versioned?
  • External API contracts defined?
  • Integration points with third-party services specified?
  • Backward compatibility addressed?
  • Rate limits, quotas, SLA constraints mentioned?
  • Regulatory / compliance implications (GDPR, PCI, HIPAA)?

Implementation Blockers — flag anything that will immediately block a developer:

  • Requirements that need an unresolved decision ("store user data" — where?)
  • References to undefined systems, APIs, or services
  • Contradicting requirements that make a single implementation impossible
  • Missing technical definitions developers will ask about immediately

Traceability chain:

  • REQ→Design: every requirement addressed by at least one design decision?
  • Design→Tasks: every design element has at least one implementing task?
  • Tasks→REQ: every task traces back to a requirement or design element?
  • Acceptance criteria in requirements consistent with done-criteria in tasks?

If requirements have no REQ-XXX IDs, trace by text — don't fail silently.

Return: PASS or NEEDS REVISION + compatibility concerns + blockers (🔴) + traceability matrix (REQ | Design | Tasks | Status)


Step 3 — Consolidate

Wait for all agents to return before consolidating.

# Spec Review: [Feature Name]
**Mode:** Quick | Full | Partial Full  **Date:** [today]

## Verdict: PASS ✅ | NEEDS REVISION ❌

## 1. Spec Quality [PASS|NEEDS REVISION]
[business + correctness + smell findings]

## 2. Completeness [PASS|NEEDS REVISION]
[missing scenarios + invariants + testability scorecard + avg score]

## 3. Buildability [PASS|NEEDS REVISION|N/A]
[compatibility + blockers + traceability matrix]

---

## Action Items
| # | Sev | Agent | Issue | Fix |
|---|---|---|---|---|
...

## Summary
Critical: N  Warnings: N  Suggestions: N  Avg testability: X.X/3.0
→ [Proceed to implementation | Revise spec first]

Verdict Rules

  • PASS — zero 🔴 Critical issues across all agents
  • NEEDS REVISION — any 🔴 Critical issue blocks implementation
  • Agent 3 is N/A for Quick mode or partial Full mode

If NEEDS REVISION: offer to open the spec file or activate spec-driven-planning to revise.


Severity

Icon Meaning
🔴 Blocks implementation
🟡 Risk of rework — fix before implementing
🟢 Optional improvement