/verification-before-completion
Assess nontrivial completion evidence.
$ npx -y skills add TheGoat395/Codex-Skills --skill verification-before-completion --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
- Fires itselfAuto-invocation. Claude auto-loads it when your prompt matches the work.Auto-invocation is when the right skill fires by itself at the right moment, driven by a FLOW.md router and a hook, instead of you invoking it by name. It is the difference between a skill being installed and a skill actually getting used.Read the full definition →
- You can call itInvoke it directly when you want it.
- Slash command
/verification-before-completion
Context preview
The summary Claude sees to decide when to auto-load this skill.
Assess nontrivial completion evidence.
SKILL.md
verification-before-completion.SKILL.mdname: verification-before-completion
description: "Assess nontrivial completion evidence."
Verification Before Completion
Core principle
Evidence before claims. Verification should be fresh, decisive, and proportionate to the result being asserted. For an obvious low-impact local edit, apply the direct check without loading this skill or creating a formal verification artifact. Use this fuller workflow when the evidence choice, failure interpretation or completion boundary is nontrivial, or when explicitly invoked.
Lightweight gate
Before claiming that work is complete, fixed, passing, deployed, sent, published, or otherwise in a verified state:
1. Name the exact claim and its consequence. 2. Choose the smallest fresh check that can actually prove it. 3. Run the check and read the exit status, failures, and material output. 4. Report the result with exact coverage and anything not verified.
Do not replace a required runtime check with source inspection, a build with lint, or end-to-end evidence with an agent's success message. Do not run an entire test universe when a targeted regression plus the repository's required gate proves the claim.
Match evidence to the claim
| Claim | Decisive evidence | |---|---| | Targeted bug fixed | Reproduce the original symptom or run its regression test | | Tests pass | The named test command exits successfully with zero relevant failures | | Build succeeds | The actual build command exits successfully | | Visual issue fixed | Inspect the rendered affected state at relevant viewports | | Requirements met | Check each material acceptance criterion against code, output, or runtime evidence | | External state changed | Read back the exact account, message, deployment, or service state | | Research complete | State sources searched, semantic coverage, exclusions, and unresolved limits |
Coverage language
Use precise labels such as:
- source-inspected;
- structurally verified;
- targeted test passed;
- runtime-tested;
- visually inspected;
- external state confirmed;
- untested or blocked.
Never imply whole-system certification from a sample. Previous evidence may provide context, but a current completion claim needs current verification when the state could have changed.
Delegated work
A subagent report is evidence about what the subagent observed, not automatic proof of completion. Inspect the relevant diff or artifact and run the smallest decisive verification before adopting its claim.
Stop condition
Stop when the acceptance criteria are supported at the required confidence. Record residual limitations instead of adding ceremonial checks that cannot change the decision.
Read more
name: verification-before-completion description: "Assess nontrivial completion evidence."
Verification Before Completion
Core principle
Evidence before claims. Verification should be fresh, decisive, and proportionate to the result being asserted. For an obvious low-impact local edit, apply the direct check without loading this skill or creating a formal verification artifact. Use this fuller workflow when the evidence choice, failure interpretation or completion boundary is nontrivial, or when explicitly invoked.
Lightweight gate
Before claiming that work is complete, fixed, passing, deployed, sent, published, or otherwise in a verified state:
1. Name the exact claim and its consequence. 2. Choose the smallest fresh check that can actually prove it. 3. Run the check and read the exit status, failures, and material output. 4. Report the result with exact coverage and anything not verified.
Do not replace a required runtime check with source inspection, a build with lint, or end-to-end evidence with an agent's success message. Do not run an entire test universe when a targeted regression plus the repository's required gate proves the claim.
Match evidence to the claim
| Claim | Decisive evidence | |---|---| | Targeted bug fixed | Reproduce the original symptom or run its regression test | | Tests pass | The named test command exits successfully with zero relevant failures | | Build succeeds | The actual build command exits successfully | | Visual issue fixed | Inspect the rendered affected state at relevant viewports | | Requirements met | Check each material acceptance criterion against code, output, or runtime evidence | | External state changed | Read back the exact account, message, deployment, or service state | | Research complete | State sources searched, semantic coverage, exclusions, and unresolved limits |
Coverage language
Use precise labels such as:
- source-inspected;
- structurally verified;
- targeted test passed;
- runtime-tested;
- visually inspected;
- external state confirmed;
- untested or blocked.
Never imply whole-system certification from a sample. Previous evidence may provide context, but a current completion claim needs current verification when the state could have changed.
Delegated work
A subagent report is evidence about what the subagent observed, not automatic proof of completion. Inspect the relevant diff or artifact and run the smallest decisive verification before adopting its claim.
Stop condition
Stop when the acceptance criteria are supported at the required confidence. Record residual limitations instead of adding ceremonial checks that cannot change the decision.
An unofficial, open-source, Codex-first Agent Skills library for premium frontend, website, design-engineering, and reliable agent workflows.

