Keep one engineering contract across Claude Code
Claude Code carries guidance in CLAUDE.md, .claude/rules, imported @AGENTS.md, and Skills, plus auto memory it writes itself. Anthropic's own docs call these context, not enforced configuration, with no guarantee Claude follows them. Baz discovers that same guidance, scopes it to the changed paths, and verifies every Change against it, so the standard holds no matter the session, the model, or what stayed in personal memory.
Claude Code tells the agent how to work. Baz keeps the standard.
Native customization tells an agent how to work. Baz turns the durable guidance already stored with your code into scoped requirements it can verify on every Change.
Native customization
CLAUDE.md, .claude/rules, Skills, hooks, subagents, plugins, MCP, and automatic memory
Where instructions drift
CLAUDE.md loads as a user message, not the system prompt, so adherence is best-effort. Long context dilutes it, Skills fire only when selected, subagents start fresh, and auto memory never reaches the repository.
What Baz adds
Baz discovers CLAUDE.md, its @-imported files, .claude/rules, and Skills, then loads only the rules relevant to the changed paths and verifies the Change against them.
A plan Claude Code can defend
Claude Code has plan mode. Baz gives that plan cross-repository evidence, so the approach is grounded before implementation begins.
Install the Baz plugin and run /baz:plan-with-baz in your session.
Claude enters plan mode and searches the organization through Baz, not just the local checkout.
The plan is routed through risk checks: known bugs, vulnerabilities, CI failures, and team requirements.
The completion hook captures the final plan with the model, usage, and relevant repositories.
Result A plan grounded in cross-repository evidence and stored as a reviewable artifact before any code exists.
One independent merge standard
Claude Code can produce work many ways: Terminal sessions, plugins, Skills, hooks, plan mode, and specialized subagents. Baz reviews the result after the code exists and decides one thing: Whether the Change follows repo guidance, satisfies the intended specification, avoids known risks, and is ready to merge.
| Baz capability | Evidence used | Decision supported |
|---|---|---|
| AI Coding Guidelines | AGENTS.md, CLAUDE.md, Cursor rules, and Skills matched to changed paths | Did the agent follow the team's durable instructions? |
| Spec Reviewer | The approved requirements and the resulting implementation | Did the agent build what the team asked for? |
| Checks | CI failures mapped to the relevant files and lines | Does the Change pass the tests and checks that matter? |
| Usage and impact | Services, APIs, consumers, entry points, and runtime context | What could the Change affect beyond the edited files? |
| Advanced Security | Code, dependencies, vulnerability patterns, and adversarial review | Does the Change introduce a security risk? |
| Merger | Findings, discussions, approvals, checks, and team merge requirements | Is the Change ready to merge, blocked, or waiting for a human decision? |
Keep every agent’s strengths. Add one control layer.
Each agent keeps its own way to shape behavior. Baz adds the same evidence before code and the same independent verdict after, so the toolchain can change without resetting the standard.
| Capability | Claude Code | OpenAI Codex | Cursor | Devin |
|---|---|---|---|---|
| Durable repo guidance | CLAUDE.md, imported AGENTS.md, .claude/rules | AGENTS.md | Project Rules, AGENTS.md | AGENTS.md |
| Reusable workflows | Skills and plugins | Skills and plugins | Skills and plugins | Skills and Playbooks |
| Specialized workers | Subagents | Subagents | Subagents | Subagents |
| Tool connections | MCP and plugin-bundled | MCP and plugin-bundled | MCP and plugin-bundled | MCP |
| Baz before code | Packaged Planner plugin and plan capture | Packaged Planner plugin and plan capture | Baz Rule, MCP, and plan capture | Baz MCP with repo-held guidance |
| Baz after code | Independent review and merge verdict | Independent review and merge verdict | Independent review and merge verdict | Independent review and merge verdict |
One control layer, whichever agent your team picks
OpenAI Codex
Baz discovers AGENTS.md and Skills, then applies the matching requirements during an independent review in GitHub, outside the session that produced the code.
Baz + OpenAI CodexCursor
Baz discovers repo-held .cursor/rules, AGENTS.md, and Skills, scopes them to the changed paths, and verifies the result outside the editor.
Baz + CursorDevin
Baz enforces repo-held AGENTS.md and Skills. Teams can move durable requirements out of Knowledge or Playbooks into versioned repo guidance so every agent and reviewer works from the same contract.
Baz + DevinQuestions teams ask first
Does Baz replace CLAUDE.md?
No. Baz uses the CLAUDE.md and Skills your team already maintains. It normalizes the instructions, applies the relevant parts during review, and verifies the final Change.
Can Baz enforce Claude's automatic memory or personal settings?
Baz enforces guidance it can read from connected repositories. Personal memory and uncommitted settings stay private. Requirements that must apply across the organization belong in versioned repo guidance.
Does Baz run inside the terminal?
Planner runs inside your Claude Code session through the plugin. Review runs independently in GitHub after the Change exists, so creation and approval stay separate.
Let developers keep Claude Code. Keep one standard for the code.
Connect Baz to your repositories and turn the guidance you already maintain into requirements verified on every Change, before code and again before merge.