Gen AI for code generation. CDD for control.
CDD Skills wrap agentic coding in the guardrails busy developers need: scoped plans, auditable TODO steps, isolated worktrees, validation evidence, and clear human approval points.
Use the core skills for precise human-guided work. Use Master Chef when the plan is ready for autonomous execution.
Powered by cdd-boilerplate.
- Human-guided feature work:
[CDD-0] Boot→[CDD-2] Plan→[CDD-3] Implement. - Post-implementation review:
[CDD-4] Audit→[CDD-2] Planfor approved follow-up. - Repo upkeep:
[CDD-5] Maintainfor doc drift, source cleanup, index refresh, and refactor review. - Autonomous delivery:
[CDD-6] Master Cheffor controlled multi-step execution after kickoff approval.
Here's a refined version for your cdd-skills repo README section:
Multiple ways to run autonomous development steps using CDD boilerplate:
- Reliability:
⚠️ Can be unreliable ongpt-5.4 xhigh.
- Reliability: ✅ Works well and fast on
Opus 4.8 xhigh. - Token use: Can be token-hungry, watch your budget.
- Process: Executes each step methodically in a highly controlled manner.
- Reliability: ✅ Highly dependable; ideal for unattended runs (e.g., overnight).
- Budget: Predictable token usage due to careful pacing; avoids excessive costs.
- Performance: Might be slower compared to Claude.
$cdd-plan docs/specs/blueprint.md into TODO.md
# Optional audit step:
$cdd-audit TODO.md and confirm cdd-master-chef readiness
$cdd-master-chef TODO.md all steps, session gpt-5.5 xhigh builder gpt-5.4 xhigh, worktree and branch todo-123Note
The session model is fixed to ensure smooth operation and consistency. In the authors’ opinion, gpt-5.5 xhigh is better for longer-running tasks, while gpt-5.4 xhigh is better for detailed work, albeit slower.
Tip
Autonomous runs work best with well-defined, properly chunked steps. Run $cdd-audit TODO.md and confirm it's cdd-master-chef ready first to review and adjust step size.
Master Chef works through runnable TODO steps in a branch-backed worktree, with persistent Builder context, recovery rules, commits, pushes, and a final mission report.
- One kickoff approval controls how far the run can go.
- Branch-backed worktrees keep autonomous changes isolated and reviewable.
- Persistent Builder sessions carry context across steps instead of starting cold every time.
- Oversized work is reviewed first and split only when the split cost is justified.
- Mission reports make the result auditable: completed steps, checks, pushes, decisions, and next actions.
Start cdd-master-chef via $cdd-master-chef for Codex or /cdd-master-chef for Claude Code or OpenClaw runtime.
The core loop is intentionally simple: boot context, plan work, implement one step, audit the result.
-
[CDD-0] Boot cdd-boot Ingest AGENTS.md, project docs, and current work context; decide whether to stay in the current checkout, continue in a linked worktree, or create a branch-backed worktree under .cdd-runtime/worktrees/.
-
[CDD-1] Init Project cdd-init-project Initialize or adopt CDD in the current repo. gh is a great tool to have when the repo is GitHub-backed.
-
[CDD-2] Plan cdd-plan Convert change requests or audit findings into implementation-ready TODO steps.
-
[CDD-3] Implement cdd-implement Implement exactly one bounded task; prefer TODO-backed steps and use direct mode only when the task is already decision-complete.
-
[CDD-4] Audit cdd-audit Audit spec, code, tests, complexity, and docs, then route approved follow-up into planning.
-
[CDD-5] Maintain cdd-maintain Review doc drift and repo upkeep, run source cleanup sweeps, refresh indexes, and audit refactor candidates.
-
[CDD-6] Master Chef cdd-master-chef Run the autonomous multi-step workflow through the canonical Master Chef skill and runtime adapters.
Tip
Fastest path — no-clone install or upgrade for all detected default runtime homes:
bash <(curl -fsSL https://raw.githubusercontent.com/ruphware/cdd-skills/main/install-remote.sh) --allNo-clone update:
bash <(curl -fsSL https://raw.githubusercontent.com/ruphware/cdd-skills/main/install-remote.sh) --all --update --yesUninstall:
bash <(curl -fsSL https://raw.githubusercontent.com/ruphware/cdd-skills/main/uninstall-remote.sh) --allUse the local installer when you want a specific runtime, target directory, or direct control.
Current concrete Master Chef adapters:
- Codex — subagent-backed adapter docs in
skills/cdd-master-chef/CODEX-ADAPTER.mdandskills/cdd-master-chef/CODEX-RUNBOOK.md - Claude Code — subagent-backed adapter docs in
skills/cdd-master-chef/CLAUDE-ADAPTER.mdandskills/cdd-master-chef/CLAUDE-RUNBOOK.md - OpenClaw — packaged adapter installed with
./scripts/install.sh --runtime openclaw
Runtime details live in:
skills/cdd-master-chef/RUNBOOK.mdskills/cdd-master-chef/CODEX-ADAPTER.mdskills/cdd-master-chef/CLAUDE-ADAPTER.mdskills/cdd-master-chef/RUNTIME-CAPABILITIES.mdskills/cdd-master-chef/openclaw/README.md
git clone git@github.com:ruphware/cdd-skills.git
cd cdd-skills
./scripts/install.shCommon variants:
./scripts/install.sh --all
./scripts/install.sh --runtime claude
./scripts/install.sh --runtime openclaw
./scripts/install.sh --helpQuick first install for Codex, Claude Code, and Gemini CLI only:
npx skills add https://github.com/ruphware/cdd-skills/ --skill cdd-boot --skill cdd-init-project --skill cdd-plan --skill cdd-implement --skill cdd-audit --skill cdd-maintain --skill cdd-master-chef -a codex -a claude-code -a gemini-cli -gThis path does not provide the repo-managed update or uninstall flow and does not preserve managed prune semantics.
Free-to-use-adjust-just-don't-blame-me-for-anything-license. ✌️
This repository follows the structured CDD Project and CDD Skills workflow using the local AGENTS.md contract.