2.5 KiB

Raw Blame History

Verification Pattern

Mandatory verification before claiming any work complete.

Purpose

Prevent false completion claims. Every tool must specify how to verify it works.

Core Rule

Never claim work is complete without running verification.

This applies to:

Building new tools
Modifying existing tools
Fixing issues found in audits
Any task that changes behavior

Implementation Checklist

Tools implementing this pattern must:

Declare verification method in tool definition
Verification must be executable (not just "review")
Run verification before completion claim
Report verification output, not just "it passed"
If verification fails, do not claim completion

Verification Types

Script-based

verification:
  type: script
  command: python scripts/validate_skill.py <skill-path>
  success_criteria: "Exit code 0, no errors"

Example-based

verification:
  type: golden_examples
  examples_path: examples/
  check: "Output matches expected for all examples"

Behavioral

verification:
  type: behavioral
  test: "Invoke with test input, check output matches invariants"
  invariants_path: invariants.md

Manual (last resort)

verification:
  type: manual
  steps:
    - "Run skill with test case X"
    - "Verify output contains Y"
    - "Check file Z was created"

Workflow Integration

Work complete?
    │
    ↓
Run verification ──→ Failed ──→ Fix issues ──→ (loop)
    │
    ↓
   Passed
    │
    ↓
Check golden examples ──→ Changed ──→ Flag for review
    │
    ↓
Report completion with verification output

Anti-patterns

Verification-by-assertion: "I verified it works" without running anything. Always show evidence.

Partial verification: Checking one thing when multiple things changed. Verify all affected behaviors.

Verification deferral: "Will verify later." Verify now or don't claim completion.

Silent failure: Verification failed but reporting success. If it fails, say so.

Evidence Format

When reporting completion, include:

## Verification Results

**Method:** [script/examples/behavioral/manual]
**Command:** [what was run]
**Output:**
[actual output or summary]

**Status:** PASSED / FAILED

Cross-references

Audit Pattern - Audits require verification of fixes
Reversion Protection - Golden examples are a verification method

2.5 KiB Raw Blame History