Agentskills Project Onboarding

AgentSkills Project Onboarding

Core Workflow

1. Inspect the repository before recommending anything: language, framework, package files, CI, test commands, docs, deployment files, existing agent instructions, and likely work modes.
2. Identify the smallest useful AgentSkills profile: one skillset, or 3-8 atomic skills when a full skillset would be too broad.
3. Choose install scope: prefer project scope for repo-specific focus; use user scope only for durable personal workflows across many repos.
4. Plan project instructions: preserve existing AGENTS.md, CLAUDE.md, GEMINI.md, .github/copilot-instructions.md, and Cursor rules; propose a patch instead of replacing them.
5. Recommend MCP presets only when the repo workflow needs them and explain the tool/security implication.
6. Verify setup with a dry run or a small trigger prompt that should activate one selected skill.
7. Report the focused setup plan, files touched, commands to run, verification evidence, rollback notes, and remaining questions.

Freshness Rule

Verify current official host documentation before giving exact host-specific paths, commands, supported agents, scope behavior, or MCP config formats when those details matter. Keep AgentSkills portfolio decisions separate from current host installer facts.

Decision Rules

• Do not recommend all unless the user explicitly wants every skill or is maintaining AgentSkills itself.
• Prefer the narrowest skillset that matches the repo's main job.
• Prefer atomic skills when the repo needs only a few workflows from different bundles.
• Keep AGENTS.md short: repo facts, selected AgentSkills routing, validation commands, and safety constraints. Put long procedures in docs or skills.
• If repo intent is unclear, ask only the missing questions that would change the skill selection or install scope.

Safety Rules

• Do not run install, network, or MCP-enabling commands without explicit user approval.
• Do not overwrite existing project instruction files; show a patch or create a clearly scoped new section.
• Do not invent build, test, lint, deploy, or environment commands.
• Do not enable broad MCP access, browser automation, or network-backed tools without explaining why they are needed.
• Do not add secrets, credentials, private URLs, or sensitive customer/project data to instruction files.

Deliverable Shape

• Repo readout
• Recommended AgentSkills profile
• Skillsets or atomic skills to install
• Host and install-scope recommendation
• Project instruction plan
• MCP preset recommendation
• Verification steps and expected trigger prompt
• Approval-required actions
• Rollback notes
• Open questions

References

• Read references/onboarding-checklist.md when creating a focused AgentSkills setup plan or editing project agent instructions.