Master Checklist: Definition of Done

One page for any skill, agent, or MCP server. For the detailed gates of each phase, use the section checklists: Skills · Agents · MCP.

Before building

• Problem statement and 3–5 real example requests written
• Baseline measured without the extension
• Lightest mechanism chosen (decision tree), with the reason logged
• Checked for an existing skill, agent, or server that already does this

Design

• Name follows conventions (skill/agent: lowercase-hyphens; MCP tool: verb_noun)
• Description or tool description states what and when, in the third person
• No overlap with sibling skills, agents, or tools
• Location and audience chosen (platform matrix)
• Least privilege: read-only unless writing is the job

Build

• Started from a template in this guide
• Only includes what the model wouldn’t already know
• Output format defined exactly
• Errors are actionable
• No secrets in files; configuration comes from the environment

Test

• npm run qa passes
• Should-trigger / should-call tests pass (fresh chat each time)
• Should-not-trigger tests pass
• Behaviour and outcome tests pass against a rubric or golden tasks
• Guardrail tests pass (forbidden actions blocked)
• Tested on every host claimed (Cursor / Claude Code / Claude Desktop / Claude.ai / SDK)
• Eval sheet saved next to the artifact

Ship

• README with install steps and 2–3 example prompts
• Version and changelog entry
• Owner assigned
• Colleague installed and used it from the docs alone
• Decision logged in your project’s decision log (for example DECISIONS.md)

Operate

• Re-run evals on model or host updates
• Review usage quarterly: improve, merge, or retire