A blueprint repository for setting up and running projects with Claude Code, with a Codex compatibility layer via AGENTS.md. Contains the RPI (Research-Plan-Implement) methodology, a catalog of known agent errors, and operational rules learned from hundreds of real sessions.
- Claude Code CLI installed and configured
- Git
Clone the repository:
git clone https://github.com/juan294/cc-rpi.gitThen tell Claude Code in your target project:
Go read the cc-rpi repository and set up this project following all the best practices. Read the quick reference and methodology first, use the error-patterns skill or full error catalog only if needed, then configure CLAUDE.md, AGENTS.md, and slash commands for this project.
Bootstrapped and adopted projects now also get an AGENTS.md
compatibility layer so the same methodology can be operated from Codex /
GPT-5.x without changing the workflow.
If you also use Codex, the blueprint ships a Codex-only
codex-simplify skill at cc-rpi/.codex/skills/codex-simplify/.
Copy it into ~/.codex/skills/codex-simplify/ if you want a reusable
equivalent of Claude Code's native /simplify without creating a
project skill named simplify.
Principle: one harness per blueprint. cc-rpi is authored for Claude
Code and maintained for a single author syncing 15+ downstream
projects. Every additional harness is ongoing sync surface --
parallel command trees, parallel wrappers, extra work in /update on
every blueprint bump -- so by default new harness support lives
outside this repo.
Codex is the one exception. AGENTS.md is a widely adopted
cross-agent markdown convention that Codex reads natively, so the
compatibility layer is a single file and a handful of translations,
not a parallel command tree. The ongoing cost is low enough to carry
upstream without dragging on /update across every project.
Other harnesses (OpenCode, GitHub Copilot, etc.) are better served by
sibling projects or community-maintained overlays. copilot-rpi is
maintained as a separate repository rather than layered on top of
cc-rpi for exactly this reason. For an ad-hoc harness such as
OpenCode, the recommended path is a small overlay repo with a
generator script that derives the harness-specific command directory
from .claude/commands/, a long-lived downstream fork, or a minimal
AGENTS.md + harness config file with no command wrappers at all.
New here? Read GUIDE.md — a human-readable walkthrough of the philosophy, the workflow, and every command. It covers everything you need to know without diving into every file. Also works great as source material for NotebookLM podcasts or articles.
The full Research-Plan-Implement pattern adapted for Claude Code, based on HumanLayer's opencode-rpi and ACE-FCA framework. Organized by topic (10 files, in reading order):
- Philosophy — Core tenets, error amplification principle, mental alignment
- Context Engineering — The foundational discipline: compaction, context quality, settings & permissions
- Four Phases — Research, Plan, Implement, Validate with detailed processes
- Agent Design — Documentarian rule, tool restrictions, subagent catalog, Anthropic-native commands (
/simplify,/batch), agent teams, autonomy principles - Pseudocode Notation — Compact notation for writing implementation plans
- Testing — Automated-first verification hierarchy, TDD protocol
- Push Accountability — Post-push CI ownership, background polling, fix-and-repush cycle
- CI & Guardrails — Pre-commit hooks, CI workflows, development guardrails, enforcement stack
- Scheduled Agents — Recurring quality agents on cron/launchd, shared context system
- Error & Success Logging — Framework for systematic improvement
A catalog of recurring Claude Code agent errors documented from real sessions. Each entry includes the symptom, root cause, correct approach, and what to avoid:
- Shell behavior (parallel calls, cwd resets, tilde paths)
- Git operations (worktrees, pre-commit hooks, push rejections)
- GitHub CLI (
ghfield names, CI status checking) - Node.js/TypeScript (ESM shebangs, Buffer vs string)
- CI & workflow (push-and-forget, skipping TDD, suggesting manual steps)
Sample documents illustrating the methodology in practice — a research document, implementation plan with phase files, error/success log entries, and additional pseudocode notation examples. Use these as reference when producing your own RPI artifacts.
Ready-to-use starting points for new projects:
- CLAUDE.md template — Slim project configuration (~70 lines) with universal instructions
- AGENTS.md template — Codex compatibility layer that teaches Codex how to interpret the cc-rpi
.claude/layout - Codex-only skills —
.codex/skills/holds personal Codex helpers that intentionally stay outside.claude/skills/; currently includescodex-simplify - Rule templates —
.claude/rules/files with conditional loading (deployment, Supabase, testing) and universal rules (RPI details, push accountability) - settings.json template —
.claude/settings.jsonwith Agent Teams, hooks, and permissions - Setup checklist — Step-by-step guide including rules, skills, hooks, CI, and scheduled agents
- Slash commands —
/bootstrap,/adopt,/update,/research,/plan,/implement,/validate,/describe-pr,/pre-launch,/remediate,/triage,/status,/fix-ci— plus Anthropic-native/simplifyand/batch - Scheduled agent scripts — Nightly blueprint sync and multi-project morning triage with launchd/cron templates
When you discover a new recurring error or best practice:
- Add it to
patterns/agent-errors.md(detailed entry with symptom/root cause/solution) - Add a one-liner to
patterns/quick-reference.md - Keep entries generic — no project-specific references
See CONTRIBUTING.md for full guidelines.
- GitHub Discussions — Ask questions, share ideas, discuss the methodology
- Contributing Guide — How to report patterns, propose improvements, submit PRs
- Code of Conduct — Expected behavior for all participants
- HumanLayer — ACE-FCA framework and opencode-rpi implementation
- Adapted for Claude Code's native capabilities (CLAUDE.md, Task tool, slash commands)
See CHANGELOG.md for release history.