Skip to content

Navigation Menu

Sign in
Appearance settings

Search code, repositories, users, issues, pull requests...

Provide feedback

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly

Sign up
Appearance settings

integration guides claude md driven

jpicklyk edited this page Apr 30, 2026 · 1 revision

Tier 2: CLAUDE.md-Driven Workflow

What You Get

  • Consistent agent behavior without installing a plugin
  • Workflow conventions embedded in project instructions
  • Team-shareable via checked-in CLAUDE.md
  • Personal overrides via gitignored CLAUDE.local.md

Prerequisites

How CLAUDE.md Works

Claude Code loads CLAUDE.md from your project root at the start of every session. Instructions in this file act as persistent context — the agent reads them before responding to any prompt.

  • CLAUDE.md — checked into source control. Shared with all team members who clone the repo.
  • CLAUDE.local.md — gitignored. Personal conventions layered on top of the shared file.

This means you can embed MCP workflow conventions once and have them apply to every session, for every team member.

Example CLAUDE.md Snippets

Workflow Protocol

Add this block to your project's CLAUDE.md:

## MCP Task Orchestrator Conventions
- Use `create_work_tree` to create features with child items — never create items one at a time
- Use `advance_item(trigger=...)` for all role transitions — never update role directly via manage_items
- Call `get_context(itemId=...)` before starting work on any item to check gate status and expected notes
- Call `get_context()` (no params) at session start to see active and stalled work
- Use `get_next_item()` when uncertain what to work on next
- Fill required notes with `manage_notes(operation="upsert")` before attempting to advance phases
- Use `get_blocked_items()` to check for dependency blockers before starting an item

Session Resume

## Session Start
At the beginning of each session, call `get_context()` to see what's active, blocked, or stalled.
If resuming from a previous session, call `get_context(since="<last-session-timestamp>")` to see
recent transitions.

Dependency Awareness

## Dependencies
Before starting any item, check `get_blocked_items()` to confirm no upstream blockers.
When completing an item, check the `unblockedItems` field in the advance_item response
to see what's now ready to start.

Role Semantics

## Work Item Roles
- **queue** — planned but not started. Fill required queue-phase notes here.
- **work** — actively being implemented. Fill implementation notes.
- **review** — implementation complete, under review. Fill review notes.
- **terminal** — done. Items here are complete, cancelled, or archived.
- **blocked** — paused due to external dependency or manual hold.

CLAUDE.local.md for Personal Conventions

Personal preferences that shouldn't be shared with the team:

## My Preferences
- Default priority for new items: high
- When I say "TO", I mean Task Orchestrator
- Preferred complexity range: 3-7 (skip trivial and overly complex items)
- When creating features, always add a "test-strategy" child item

Team Scenario

A typical team setup:

  1. CLAUDE.md (checked in): MCP workflow conventions, project-specific tool patterns, shared terminology
  2. CLAUDE.local.md (gitignored): Personal shortcuts, preferred priorities, individual work style
  3. A new team member clones the repo and gets all MCP conventions automatically

The agent reads both files, with CLAUDE.local.md layered on top. No plugin installation required — the conventions travel with the repo.

Combining with Note Schemas

CLAUDE.md can instruct agents to fill specific note keys even without schema gating:

## Note Conventions
When creating a feature item, always fill these notes:
- `requirements` (role: queue) — what the feature must do
- `implementation-plan` (role: work) — how you'll build it
- `test-results` (role: review) — test output and coverage

This is convention-based — the agent follows the instructions but nothing enforces them. For enforcement, add Note Schemas (Tier 3).

When to Level Up

Signal: You want automated quality gates — required documentation before items can advance, not just conventions the agent might forget.

Next: Note Schema Gating — add .taskorchestrator/config.yaml to enforce required notes at each phase.

Alternative signal: You want skill-based automation (plan-mode pipeline, subagent protocols) without intermediate schema setup.

Next: Plugin: Skills and Hooks — install the Claude Code plugin for automated workflows.

Clone this wiki locally

AltStyle によって変換されたページ (->オリジナル) /