AgentSkillsCN

spec-verification

验证生成的规格文档是否保留了上下文并避免质量问题。在文档生成后自动触发。检查上游需求是否被保留,并识别常见的规格问题。

SKILL.md
--- frontmatter
name: spec-verification
description: Verify generated specification documents for context preservation and quality issues. Automatically triggered after document generation. Checks that upstream requirements are preserved and identifies common specification problems.

Spec Verification Skill

Verify specification documents for context preservation (nothing important lost from upstream docs) and quality issues (vague language, missing rationale, untestable criteria, etc.).

When This Skill Runs

This skill runs automatically after generating:

  • TECHNICAL_SPEC.md (verifies against PRODUCT_SPEC.md)
  • EXECUTION_PLAN.md (verifies against TECHNICAL_SPEC.md)
  • FEATURE_TECHNICAL_SPEC.md (verifies against FEATURE_SPEC.md)
  • Feature EXECUTION_PLAN.md (verifies against FEATURE_TECHNICAL_SPEC.md)

Can also be invoked manually via /verify-spec <document-type>.

Workflow

Copy this checklist and track progress:

code
Spec Verification Progress:
- [ ] Step 1: Detect document type
- [ ] Step 2: Context preservation check
- [ ] Step 3: Quality check
- [ ] Step 4: Present issues
- [ ] Step 5: Collect user choices
- [ ] Step 6: Apply fixes
- [ ] Step 7: Re-verify

Workflow Overview

code
1. Identify document type and locate upstream document(s)
2. Context Preservation Check: Extract key items from upstream, verify in current
3. Quality Check: Scan for common specification anti-patterns
4. Present issues inline with resolution options
5. Collect user choices for each CRITICAL issue
6. Apply fixes to document(s)
7. Re-verify (max 2 iterations, then escalate)

Step 1: Document Type Detection

Determine what's being verified and its upstream dependencies:

Document Being VerifiedUpstream Document(s)Generation Prompt
TECHNICAL_SPEC.mdPRODUCT_SPEC.md.claude/skills/technical-spec/PROMPT.md
EXECUTION_PLAN.md (greenfield)TECHNICAL_SPEC.md, PRODUCT_SPEC.md.claude/skills/generate-plan/PROMPT.md
FEATURE_TECHNICAL_SPEC.mdFEATURE_SPEC.md.claude/skills/feature-technical-spec/PROMPT.md
EXECUTION_PLAN.md (feature)FEATURE_TECHNICAL_SPEC.md, FEATURE_SPEC.md.claude/skills/feature-plan/PROMPT.md
PRODUCT_SPEC.mdNone (quality check only).claude/skills/product-spec/PROMPT.md
FEATURE_SPEC.mdNone (quality check only).claude/skills/feature-spec/PROMPT.md

Step 2: Context Preservation Check

2a: Extract Key Items from Upstream Document

Read the upstream document and extract:

From PRODUCT_SPEC.md / FEATURE_SPEC.md:

  • Core features (each listed feature)
  • User flows (each step-by-step flow)
  • Constraints (security, performance, accessibility, etc.)
  • Edge cases explicitly mentioned
  • Non-functional requirements
  • Integration points (feature only)
  • Backwards compatibility concerns (feature only)

From TECHNICAL_SPEC.md / FEATURE_TECHNICAL_SPEC.md:

  • Architecture decisions
  • Data model entities and relationships
  • API endpoints/contracts
  • Technical constraints
  • Implementation sequence items
  • Edge cases and boundary conditions

Create a checklist of extracted items with unique IDs (e.g., CTX-001, CTX-002).

Verify the upstream document was read successfully by confirming at least one key item was extracted. If zero items are extracted, re-read the file and check for formatting issues before proceeding.

2b: Verify Each Item in Downstream Document

For each extracted item, check if it appears in the downstream document:

StatusMeaning
PRESENTItem found with clear reference
CONSOLIDATEDItem merged into broader item (acceptable)
DEFERREDExplicitly marked as future/out-of-scope (acceptable)
MISSINGNot found, not deferred - FLAG AS CRITICAL

2c: Context Preservation Output

code
CONTEXT PRESERVATION CHECK
==========================
Upstream: PRODUCT_SPEC.md
Downstream: TECHNICAL_SPEC.md

Extracted Items: 12
Present: 9
Consolidated: 2
Deferred: 0
Missing: 1 <- CRITICAL

MISSING ITEMS
-------------
[CTX-007] "Rate limiting for API endpoints"
  - Source: PRODUCT_SPEC.md, Section "Security Requirements"
  - Not found in TECHNICAL_SPEC.md
  - Not marked as deferred

Step 3: Quality Check

Scan the document for common anti-patterns. Be conservative - only flag obvious, clear problems.

Universal Quality Checks (All Documents)

IDCheckPattern to DetectSeverity
Q-001Vague/Unmeasurable Language"fast", "user-friendly", "intuitive", "simple", "easy", "quickly", "responsive" without metricsCRITICAL
Q-002Subjective Terms"better", "improved", "enhanced" without baselineCRITICAL
Q-003Missing RationaleArchitecture/tech decisions without "because" or "to enable"MAJOR
Q-004Scope CreepItems in downstream not traceable to upstreamCRITICAL
Q-005Implicit AssumptionsDecisions that assume unstated contextMAJOR
Q-006Conflicting RequirementsTwo statements that cannot both be trueCRITICAL

See references/quality-check-patterns.md for the complete quality check pattern tables for PRODUCT_SPEC, FEATURE_SPEC, TECHNICAL_SPEC, and EXECUTION_PLAN documents.

Quality Check Output

code
QUALITY CHECK
=============
Document: TECHNICAL_SPEC.md

Issues Found: 3

[Q-001] CRITICAL: Vague Language
  Location: Section "Performance Requirements"
  Text: "API responses should be fast"
  Problem: "fast" is unmeasurable

[Q-TS-003] CRITICAL: Undefined API Contract
  Location: Section "API Endpoints", /api/users
  Text: "POST /api/users - Creates a new user"
  Problem: No request body or response shape defined

[Q-003] MAJOR: Missing Rationale
  Location: Section "Architecture"
  Text: "Use Redis for caching"
  Problem: No explanation of why Redis vs alternatives

Step 4: Issue Presentation

Present all CRITICAL issues inline. Do NOT present MAJOR issues unless there are no CRITICAL issues.

Issue Presentation Format

For each CRITICAL issue, present:

code
────────────────────────────────────────────────────────────
ISSUE 1 of N: [Issue ID] - [Issue Type]
────────────────────────────────────────────────────────────
Location: [Document], [Section/Line]
Problem: [Clear description]
Text: "[The problematic text]"

How would you like to resolve this?

Then use AskUserQuestion with appropriate options based on issue type.

Resolution Options by Issue Type

Vague/Unmeasurable Language (Q-001, Q-EP-007):

  • Option 1: Use suggested specific wording (provide suggestion)
  • Option 2: Specify your own measurable target
  • Option 3: Mark as intentionally vague (adds TODO for later)
  • Option 4: Remove this requirement

Missing from Upstream / Scope Creep (CTX-*, Q-004):

  • Option 1: Add to upstream document (expands scope)
  • Option 2: Remove from current document
  • Option 3: Mark as explicitly deferred

Missing Rationale (Q-003):

  • Option 1: Add suggested rationale (provide suggestion)
  • Option 2: Provide your own rationale
  • Option 3: Remove this decision

Undefined Contract/Model (Q-TS-002, Q-TS-003):

  • Option 1: Define now (will prompt for details)
  • Option 2: Mark as TBD with TODO

Untestable Criteria (Q-EP-001):

  • Option 1: Use suggested testable version (provide suggestion)
  • Option 2: Rewrite criterion
  • Option 3: Remove criterion

Missing Verification Metadata (Q-EP-008):

  • Option 1: Add suggested verification type + method (recommended)
  • Option 2: Mark as MANUAL with a reason
  • Option 3: Remove criterion

Conflicting Requirements (Q-006):

  • Option 1: Keep first, remove second
  • Option 2: Keep second, remove first
  • Option 3: Clarify intent (will prompt for resolution)

Step 5: Collect User Choices

Use AskUserQuestion to collect resolution for each CRITICAL issue.

Example:

code
questions:
  - question: "API responses should be fast' is unmeasurable. How should we fix this?"
    header: "Vague term"
    options:
      - label: "API responses return within 200ms p95 (Recommended)"
        description: "Sets a specific, measurable latency target"
      - label: "Specify custom target"
        description: "You'll provide your own performance number"
      - label: "Remove requirement"
        description: "Delete this requirement entirely"

If user selects "Other" or a custom option, prompt for their input.

Step 6: Apply Fixes

For each collected resolution:

  1. Read the target document
  2. Locate the problematic text (use exact match from issue)
  3. Apply the fix using Edit tool
  4. Verify the edit — Re-read the edited section to confirm the change was applied correctly. If the Edit tool reported no match or the content is unchanged, retry with corrected old_string.
  5. For upstream document changes: Require explicit confirmation before editing

Fix Application Rules

  • Make minimal edits - only change what's necessary
  • Preserve document formatting and structure
  • For additions, place in appropriate section
  • For removals, ensure no orphaned references remain

Upstream Document Edits

When a fix requires editing an upstream document (e.g., adding a requirement to PRODUCT_SPEC.md):

code
This fix requires editing the upstream document.

Document: PRODUCT_SPEC.md
Change: Add "Rate limiting: 100 requests/minute per user" to Security Requirements section

Confirm this upstream change? (This expands project scope)

Only proceed with explicit confirmation. After editing an upstream document, re-read the modified section to confirm the change was applied correctly and did not corrupt surrounding content.

Step 7: Re-Verification

After applying fixes:

  1. Run context preservation check again on modified documents
  2. Run quality check again on modified documents
  3. If new CRITICAL issues: Present them (max 2 total iterations)
  4. If still issues after 2 iterations: Escalate to user

Escalation Format

code
VERIFICATION INCOMPLETE
=======================
After 2 fix iterations, these issues remain:

[Q-001] Vague Language in Section X
  - Attempted fixes: [list]
  - Recommendation: Manual review needed

Proceeding with current state. Review flagged sections before execution.

Step 8: Final Report

code
SPECIFICATION VERIFICATION COMPLETE
===================================
Document: TECHNICAL_SPEC.md
Status: PASSED | PASSED WITH NOTES | NEEDS REVIEW

Context Preservation:
  - Items checked: 12
  - All present or explicitly deferred: Yes/No

Quality Check:
  - Critical issues found: 2
  - Critical issues resolved: 2
  - Major issues noted: 1 (non-blocking)

Fixes Applied:
  - TECHNICAL_SPEC.md: 2 edits
  - PRODUCT_SPEC.md: 0 edits

Notes:
  - [Q-003] Missing rationale for Redis - marked for future documentation

Ready to proceed: Yes/No

Conservative Flagging Principles

To avoid noisy reports:

  1. Only flag obvious problems - If it requires interpretation, skip it
  2. Prefer false negatives - Missing a minor issue is better than crying wolf
  3. Context matters - "Fast" in a performance section is vague; in casual description, maybe not
  4. Trust consolidation - If upstream item seems covered by broader downstream item, mark as CONSOLIDATED
  5. Limit output - Maximum 5 CRITICAL issues per verification. If more exist, show top 5 by severity

Severity Definitions

SeverityMeaningAction
CRITICALWill cause implementation problems or lost requirementsBLOCKS - must resolve
MAJORShould be addressed but won't break implementationNote for user, continue
MINORStyle/preference issueIgnore unless no other issues

Manual Invocation

The skill can be invoked manually:

code
/verify-spec technical-spec     # Verify TECHNICAL_SPEC.md
/verify-spec execution-plan     # Verify EXECUTION_PLAN.md
/verify-spec feature-technical  # Verify FEATURE_TECHNICAL_SPEC.md
/verify-spec feature-plan       # Verify feature EXECUTION_PLAN.md
/verify-spec product-spec       # Quality check only (no upstream)
/verify-spec feature-spec       # Quality check only (no upstream)

Example Workflow

Input: Just generated TECHNICAL_SPEC.md from PRODUCT_SPEC.md

Step 1: Detect document type = TECHNICAL_SPEC, upstream = PRODUCT_SPEC

Step 2: Extract 8 key items from PRODUCT_SPEC:

  • CTX-001: User authentication
  • CTX-002: Data export feature
  • CTX-003: Rate limiting
  • ... (5 more)

Verify in TECHNICAL_SPEC:

  • CTX-001: PRESENT (Section "Authentication")
  • CTX-002: PRESENT (Section "Export API")
  • CTX-003: MISSING <- Flag

Step 3: Quality check finds:

  • Q-001: "API should respond quickly" <- CRITICAL
  • Q-TS-003: POST /users has no request shape <- CRITICAL

Step 4-5: Present 3 issues, collect resolutions:

  • CTX-003: User chooses "Add to TECHNICAL_SPEC.md"
  • Q-001: User chooses "200ms p95"
  • Q-TS-003: User provides request/response shape

Step 6: Apply 3 edits to TECHNICAL_SPEC.md

Step 7: Re-verify - all clear

Step 8: Report PASSED, ready to proceed

When Verification Cannot Complete

If issues persist after 2 fix iterations:

  • Stop attempting automatic fixes
  • Report: "Verification loop limit reached (2 iterations)"
  • List all remaining issues with their current state
  • Recommend: "Manual review required for these {N} issues"
  • Ask user: "Continue with known issues or pause for manual resolution?"

If upstream document is missing or unreadable:

  • Skip context preservation check
  • Proceed with quality checks only
  • Report: "Context preservation skipped (upstream document unavailable)"
  • Note which checks were limited

If user declines to resolve CRITICAL issues:

  • Do NOT silently proceed
  • Report: "Proceeding with {N} unresolved CRITICAL issues"
  • List the specific items that remain unresolved
  • Warn about potential downstream impacts

Error Handling

SituationAction
Upstream document not found at expected pathSkip context preservation check; proceed with quality checks only and note the limitation in the report
Document is not valid Markdown (cannot parse headings/sections)Report "Unable to parse document structure" and provide manual verification steps as fallback
Edit tool fails to match old_string in the documentRe-read the document to get current content, then retry with corrected context
User declines to resolve all CRITICAL issuesDo NOT silently proceed; report unresolved count and warn about downstream impacts
Re-verification loop exceeds 2 iterationsStop automatic fixes, list remaining issues, and escalate to user for manual resolution

If document format prevents parsing:

  • Report: "Unable to parse document structure"
  • List what could not be extracted
  • Suggest: "Check document follows expected markdown format"
  • Provide manual verification steps as fallback