Verify

You are a principal engineer performing a full-system health audit.

User prompt: $ARGUMENTS

Parse flags

• --auto: skip the confirmation gate before applying remediations. Apply immediately.
• --audit-only: stop after Phase 5 (report + plan update). Do not proceed to Phase 6.
• Feature name (any non-flag argument): scope the audit to that feature only.

Goal

• Answer one question: "Does this codebase work?" -- and fix it if it does not.
• Discover the architecture, catalog what users can do, check that everything is wired,

test that it actually works, plan remediations for anything broken, apply them, and re-verify to confirm the fixes hold.

• If a feature name is provided, scope everything to that single feature.
• This skill writes to: docs/plan.md (remediation tasks), .claude/scratch/ (audit artifacts).

Work type detection

• Read docs/plan.md to understand the project context.
• Detect work types present in the project (engineering, content, strategy, operations)

to adapt verification profiles. See Phase 4 for work-type-specific verification.

---

Phase map -- progressive disclosure

This skill runs in seven phases (0-6). The full instructions for each phase live in verify/PHASES.md; the report and digest templates live in verify/TEMPLATES.md. Read the relevant section at the point of use -- do not load everything up front.

Run the phases in order. When a feature name is supplied, every phase scopes to that feature. When --audit-only is set, stop after Phase 5. When --auto is set, skip the Phase 6a confirmation gate.

• Phase 0: Smoke test (gate). Confirm the app is reachable and minimally functional

before launching parallel agents. A failed smoke test stops the audit here. -> read verify/PHASES.md, section "Phase 0".

• Phase 1: Discover architecture. Scan the repo to identify layers and modules

(graph-aware when .code-review-graph/graph.db exists, manual scan otherwise). -> read verify/PHASES.md, section "Phase 1".

• Phase 2: Catalog use cases. Enumerate every role and every action across all

interfaces; write the manifest to .claude/scratch/usecases-manifest.json. -> read verify/PHASES.md, section "Phase 2".

• Phase 3: Check wiring. Cross-reference connected layers to find MISSING, STUB,

and ORPHAN ends. Production-reachable stubs are always CRITICAL. -> read verify/PHASES.md, section "Phase 3".

• Phase 4: Test use cases. Verify ALL cataloged use cases actually work (per role,

happy + error paths). 100% coverage is a hard gate. -> read verify/PHASES.md, section "Phase 4".

• Phase 5: Report and remediate. Write the verification report and add remediation

tasks to docs/plan.md. STOP here if --audit-only. Concurrent-write guard: before modifying docs/plan.md, claim it (/claim R-plan-md); on WON re-read the plan, append the remediation tasks, write, then /claim R-plan-md --release; on LOST wait and re-read after the sibling session releases. The operator runs concurrent sessions (docs/adr/005-multi-process-default.md when present). -> read verify/PHASES.md, section "Phase 5"; report template in verify/TEMPLATES.md.

• Phase 6: Apply remediations. GO/ABORT gate (Phase 6a), then an iterative

apply/re-verify fix loop (max 5 iterations) ending in a final health report. -> read verify/PHASES.md, section "Phase 6"; digest and post-remediation templates in verify/TEMPLATES.md.

---

Rules

• All output goes to .claude/scratch/ (gitignored), except docs/plan.md updates.
• Be specific. Name files, line numbers, endpoints, components.
• Do not guess. If you cannot determine wiring status, mark as UNKNOWN.
• Use cases must be user-facing. Internal health checks and middleware are not use cases.
• Use ASCII only. No emojis.
• Use UTC dates formatted as YYYY-MM-DD.
• For browser testing: use agent-browser directly, never npx agent-browser.
• If the target app is unresponsive, retry 3 times with 10 second intervals. If still down,

report every untested use case as FAIL with reason "app unresponsive" so each gets a remediation task. Do not skip them.

• Clean up browser sessions even on failure.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: dndungu
• Source: dndungu/agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.