Claude Code Workflow

The problem

Coding agents without guardrails wander the repo, ignore conventions, and produce changes no one reviews. A team needs repo-level instructions, scoped skills, and an automated review loop to make agent-assisted work safe.

System architecture

A repo carries a CLAUDE.md operator runbook, a set of scoped skills, and MCP servers exposing internal tools; a GitHub Action runs an agent review on each pull request so changes are checked against the conventions before merge.

Workflow

• Repo instructions and conventions are codified in a CLAUDE.md runbook.

• Scoped skills package repeatable tasks the agent can invoke.

• MCP servers expose internal tools to the agent under auth.

• A developer makes agent-assisted changes on a branch.

• A GitHub Action runs an agent review on the pull request.

• Findings post back as comments before the change is merged.

Stack

• Claude Code as the coding agent

• A CLAUDE.md runbook for repo conventions

• Scoped skills for repeatable tasks

• MCP servers for internal tool access

• GitHub Actions for the review loop

What gets logged

• Which skill or MCP tool an agent invoked

• The branch and pull request the change landed on

• The review action's findings per pull request

• Conventions the change passed or violated

• Approvals and merge decisions

Where evals run

The review runs in CI on every pull request, checking changes against the repo conventions and surfacing findings before merge; a baseline set of review cases guards the reviewer itself from regressing.

Failure modes

• The agent ignores a convention — the CI review flags it and blocks merge.

• An MCP tool is called without authorization — access is gated so the call is refused.

• A skill goes stale against the codebase — skills are versioned in-repo and updated with the conventions they encode.

• The reviewer over-flags — findings are advisory comments, and a baseline set keeps its signal honest.

What this demo proves

That coding agents can be made repeatable and reviewable across a team — codified conventions, scoped skills, and an automated review gate instead of ad-hoc prompting.