Skip to content
Back to skills

Verify

ASecurity

Run the actual application end-to-end and exercise the user journey from the spec. Phase 6 of the eight-phase workflow. Delegate to verifier agent. Captures screenshots, logs, HTTP traces. The "did it actually work?" gate.

  • 2 stars
  • 0 votes
  • 0 copies
  • 1 view
  • Added October 9, 2026
ai-agentsbashdocker

Works with

  • cli
  • mcp

Security analysis

A100/100

Scanned October 9, 2026

npx -y skills add qyndex/claude-init --skill verify --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Verify?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Verify
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/qyndex-verify/badge)](https://www.skillsdirectory.com/skills/qyndex-verify)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: verify
description: Run the actual application end-to-end and exercise the user journey from the spec. Phase 6 of the eight-phase workflow. Delegate to verifier agent. Captures screenshots, logs, HTTP traces. The "did it actually work?" gate.
when_to_use: Implementation is complete (all tasks [x]) and you need to confirm the user journey works in a running system. User says "verify", "run it", "does it work", "test end-to-end".
argument-hint: "<feature id or 'latest'>"
model: sonnet
allowed-tools: Read, Glob, Grep, Bash, WebFetch
context: fork
agent: verifier
mcpServers: [chrome-devtools, playwright]
---

# Verify

Run the app. Exercise the user journey. Capture evidence. Mark PASS or FAIL.

## Process

1. Read the spec's user journey.
2. Boot the app (`bash .claude/scripts/run.sh` or project-specific).
3. Walk the journey step by step, capturing evidence at each step.
4. Walk the negative paths (bad input, missing auth, etc.).
5. Save artifacts under `verify/<date>-<feature>/`.
6. Write the report.
7. **Sync living state (M-06 — network boundary).** Verify is a network boundary: refresh `initiatives/active/<id>.STATE.md` so the always-current answer reflects this run. Grep-gated so it's a no-op when the script is absent:
   ```bash
   grep -ql 'initiative-state.sh sync' .claude/scripts/initiative-state.sh 2>/dev/null \
     && bash .claude/scripts/initiative-state.sh sync || true
   ```

## Evidence types

- Screenshots (browser MCP `take_screenshot`)
- DOM/a11y snapshots (`take_snapshot`)
- HTTP responses (`curl -i` or Postman collection)
- Log excerpts (`docker compose logs --since 5m`)
- Console messages (`list_console_messages`)
- Network requests (`list_network_requests`)
- Exit codes (for CLIs)

## Hard rules

- **Evidence before assertion.** "Looks fine" is not verification.
- **The user journey, not just tests.** Unit tests passing ≠ user works.
- **Negative paths too.** Try the obvious failure.
- **Real-ish data.** Fresh seeded DB or test tenant; no internal mocks.
- **Capture, don't summarize.** Save raw artifacts so future you can re-examine.

## Output

```
verify/<date>-<feature>/REPORT.md
verify/<date>-<feature>/*.png
verify/<date>-<feature>/*.json
verify/<date>-<feature>/*.log
```

REPORT.md format: see `.claude/agents/quality/verifier.md`.

After save: "Verification PASS — N steps, M negative paths. Ready for /review." OR "Verification FAIL at step K. See verify/<date>/REPORT.md. Hand back to implementer."

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.