A structured development workflow for AI coding agents that brings memory, consistency, and reduced hallucination (only humans should) to AI-assisted development. TRIP helps you enter flow state and eat features like buttered noodles.
It is also the acronym (reversed) of the historical 4-phases development cycle: Plan, Implement, Review, Test.
Note: Since v2.0.0 the flow is even simpler Plan → Implement → Release — review and test moved inside Implement as a testing gate and an automatic Codex review loop, every feature passes through all 4 phases with fewer commands.
TRIP was initially designed for Claude Code using the Agent Skills open standard (SKILL.md). Also compatible with OpenCode, Codex CLI, Mistral Vibe and more.
There are tons of AI coding workflows out there like Superpowers, BMAD, Gastown and countless others. They might be powerful, but overwhelming for many of us dumb asses.
Even the "simple" ones come with:
- 47 different commands & skills to memorize
- Sub-agents swarm for God-knows-what
- Multi-chapters courses (sometimes paid lol)
TRIP is different. It's deliberately minimal:
| That's it | Just these |
|---|---|
/TRIP-1-plan |
Think before you code |
/TRIP-2-implement |
Implement, test & review |
/TRIP-3-release |
Version, changelogs, docs, commit, tag, merge, push |
3 numbered skills. 1 architecture file. 0 PhD required.
The onboarding is: copy the folder, run init, start coding. If you can count to 3, you can TRIP.
It was kept stupid simple because the goal is to ship features, not to master a workflow. The workflow should disappear into the background, not become a project of its own.
- Copy the
skills/folder contents to your repo's.claude/skills/or whatever - Run
/TRIP-init [YourProjectName] - Follow the interactive prompts
- Review and approve the generated ARCHI.md
Also copy AskUserQuestion/ to your agent /skills/, it provides the AskUserQuestion hook that TRIP workflow rely on.
Et voila ! Start using the skills like /TRIP-1-plan auth for this webapp, /TRIP-2-implement @auth-plan.md, etc.
demo.mp4
The ARCHI.md file is the central nervous system of this workflow. It serves as the AI agent's long-term memory of your codebase.
1. Persistent Context Across Sessions
AI agents have no memory between sessions. Every new conversation starts from zero. ARCHI.md solves this by providing a comprehensive, always-up-to-date snapshot of your architecture that the agent reads at the start of each task. Unlike tool-specific files like CLAUDE.md or AGENTS.md, ARCHI.md is purely about architecture. It's tool-agnostic, so it works with any agent. You can still reference it from your CLAUDE.md to include it in all conversations.
2. Token Savings & Reduced Hallucination
Without ARCHI.md, your agent must glob, grep, and read multiple files to piece together the architecture from scratch for every single session. This wastes tokens and leads to guessing: "There's probably a utils folder...", "This project likely uses Redux...". ARCHI.md eliminates both problems. The agent gets the full picture in one read for minimal exploration & hallucination.
3. Balanced Detail vs Token Usage
ARCHI.md is designed to be:
- Detailed enough to provide meaningful context, concise enough to not waste tokens
- Structured for quick navigation
- Updated after every architectural change
It's not a dump of your entire codebase, rather a curated architectural guide.
The TRIP-init skill is a script written in human language that programmatically bootstraps the TRIP workflow in any repository.
- Creates the docs structure - Folders for plans, changelogs, reviews, tests, memos
- Explores your codebase - Identifies languages, frameworks, patterns, conventions
- Classifies your project - Web frontend? CLI tool? Embedded firmware? Library?
- Generates ARCHI.md - Tailored to your specific project type
- Customizes the skills - Replaces placeholders with your project's specifics
The generic TRIP skills contain placeholders like:
[PROJECT_NAME]- Your project's name[VERSION_FILE]- Where your version is stored (package.json, Cargo.toml, etc.)[ADAPT_TO_PROJECT: ...]- Sections to customize
Init walks you through questions and replaces these placeholders based on your answers, creating a workflow tailored to your project.
Implementation delegated to Codex CLI in a workspace-write sandbox.
Iterative review loops powered by Codex CLI.
A grounded second opinion on anything. TRIP-research uses it to brainstorm findings before presenting them.
The former steps 3 and 4, reborn as on-demand support skills.
Upgrades an existing project's TRIP skills to a newer version without losing project customizations. Copy the new skills to new-TRIP/, run the skill, done.
Streamlined workflow for production emergencies. Bypasses full TRIP for genuine crises (or lazy debugging).
Exploratory investigation with defined compute level. For feasibility studies and technology evaluation. Produces documented findings, not production code.
Run this skill to compact ARCHI.md size while preserving relevance, accuracy, and coverage through summarization and restructuring. Token calculator script included.
Just like you wouldn't smell your own fart, an LLM is unlikely to catch bugs in its own implementation. Some people conduct adversarial review with a different session but still the same model, which is...meh. The best approach is to introduce a different model in the same reasoning ballpark as the first one, that will most likely catch what the other missed.
As of v2.0.0, this multi-agent approach is the default workflow.
Considering Claude as your main and Codex as the copilot:
Fable writes the plan, 5.6 Sol reviews it, Luna implements, back to Fable who reviews and fixes the diff, runs the testing gate, then a new Sol thread reviews again the code. All in one claude code session. Writer and reviewer are never the same thread.
flowchart TD
A["<b>/TRIP-1-plan</b><br/>Discovery and plan draft"] --> B{"ChatGPT Sol<br/>plan review"}
B -->|"REQUEST_CHANGES"| Bf["Fable fixes the plan"]
Bf -->|"re-review"| B
B -->|"APPROVED"| D["<b>/TRIP-2-implement</b><br/>Branch + split<br/>to-dos into batches"]
Bf ~~~ D
D --> E["ChatGPT Luna<br/>implements a batch"]
E --> F["Fable reviews the delta,<br/>fixes directly"]
F -->|"next batch"| E
F -->|"all batches done"| G["Fable final pass<br/>+ testing gate"]
G --> H{"ChatGPT Sol<br/>full code review"}
H -->|"REQUEST_CHANGES"| Hf["Fable fixes + re-tests"]
Hf -->|"re-review"| H
H -->|"APPROVED"| K["<b>/TRIP-3-release</b><br/>Version bump · changelog<br/>docs/ARCHI update<br/>commit · tag · ff-merge · push"]
Hf ~~~ K
As of mid july 2026, this Fable + GPT5.6 harness combo is absolute peak.
Last piece of advise before your new coding quest: Every MCP server you add is extra context, extra latency, and extra confusion. Keep it minimal. The one use case where MCP genuinely shines is up-to-date documentation, so your agent stops hallucinating deprecated APIs/whatever. Two servers cover it: Context7 for current library & framework docs, and Exa for web search when the answer isn't in any doc. No bloat beyond that.
PRs & forks are welcome
Happy tripping ! 🍄