done done

Developers often waste hours fixing "completed" work from agents that never actually ran the code. This skill installs a hard definition-of-done protocol that forbids the agent from declaring a task finished until it provides verifiable proof of execution.

What it does

• Automatic detection identifies the project's native verification commands by inspecting lockfiles, CI configs, Makefiles, and manifests.
• Real-world execution runs the detected test, lint, or build commands in the current environment instead of predicting outcomes.
• Evidence capturing pastes exact terminal commands and raw exit outcomes directly into the chat for human verification.
• Self-correction loops attempts to fix failures up to three times before stopping to prevent token thrashing and code degradation.
• Honest reporting generates a standardized BLOCKED report when verification fails, pinpointing the exact error and recommended human intervention.

How it works

1. The agent identifies the project's source of truth for "passing" by scanning scripts, CI workflows, and task runners.
2. It executes the full relevant check or specific subset covering the changes.
3. It captures the raw output and exit codes.
4. If green, it issues a structured DONE report; if red, it attempts a fix or issues a BLOCKED report after three failures.

Frameworks & tools

Works with any stack defining verification in package.json, Cargo.toml, .github/workflows/, Makefile, justfile, or Taskfile. It maps to tools like npm, pytest, cargo, go test, and standard linters.

Why this beats prompting it yourself

Standard prompts often result in "should work" hallucinations or silent abandonment when builds fail. This skill enforces a rigid state machine that makes it impossible for the agent to say "done" without evidence, replacing vague prose with terminal output.

Use cases

• Closing a complex refactor where you need to ensure no regressions occurred.
• Fixing bugs in unfamiliar repos where you want the agent to prove the fix passes the CI suite.
• Onboarding agents to new projects to identify missing environment dependencies immediately.

Known limitations

Requires a runnable environment with necessary tools installed. It cannot verify work in projects that have no declared test, lint, or build scripts.