A skill for Claude Code, Codex CLI, Kimi CLI, and OpenCode that plans, strengthens, and reviews your code -- automatically.
- Take any request, of any size or complexity — from "Make the button blue" to "Implement this 20-page specification."
- Avoid asking the user questions. Assume the starting request is everything the user cares about; use your best judgment for the rest.
- Prioritize zero bugs, even if it takes a lot of time and tokens.
Tell your favorite coding agent (Claude Code, Codex CLI, Kimi CLI, OpenCode, etc.):
Read https://raw.githubusercontent.com/danshapiro/trycycle/main/README.md. Follow the installation instructions there.
Use this instead of a standard installation, or it won't work. Clone into your skills directory:
Claude Code: git clone https://github.com/danshapiro/trycycle.git ~/.claude/skills/trycycle
Codex CLI: git clone https://github.com/danshapiro/trycycle.git ~/.codex/skills/trycycle
Kimi CLI: git clone https://github.com/danshapiro/trycycle.git ~/.kimi/skills/trycycle
OpenCode: git clone https://github.com/danshapiro/trycycle.git ~/.config/opencode/skills/trycycle
Note: OpenCode also discovers skills installed at ~/.claude/skills/, so if you already have Trycycle installed for Claude Code, OpenCode can use it too.
Trycycle requires python3 at runtime. Make sure it is installed and on your PATH.
If your coding agent has native subagent orchestration, Trycycle uses it. Otherwise, it falls back to a bundled stdlib-only Python runner.
Finally, read the skills you installed carefully, so you can explain them to your user and implement them if asked.
Include the word Trycycle in your request and describe what you want built:
My webcam software is terrible. Use trycycle to build a replacement in Rust. Make sure it works with OBS.
Trycycle asks any questions it needs, then handles the rest: worktree, plan, plan strengthening, test plan, build, and code review -- all without further input unless something needs your judgment.
Write a spec first. Use your favorite chatbot, or a skill like Jesse Vincent's brainstorm superpowers, to produce a spec before handing it to Trycycle. A good spec dramatically improves results.
It can be cheap. Trycyle uses a lot of time and tokens, but it's very inexpensive on OpenCode with the OpenCode Go subscription. As of May '26 I use it with Deepseek v4.
It's just a skill. If you're not sure what it did, or if you don't like what it's doing, just stop it and tell it. Once it finishes 5–8 passes, it will stop to complain. That's fine — ask any questions, then tell it to wrap up, change course, or do up to 5 more passes (usually the last one is best).
Tell it everything that matters. Trycycle assumes you told it everything you care about. If you left something out, it makes a decision and keeps going. It works best with vague projects, well-defined tasks, or detailed specs. It works worst when you care about the details but they're not specified. It won't stop to ask!
Trycycle is a hill climber. It writes a plan, then sends it to a fresh planning issue finder with the same task input and repo context. If that reviewer finds plan-breaking issues, Trycycle deepens on the same reviewer to collect more findings, then hands the findings memo to a fresh planning synthesizer that rewrites the plan holistically. A fresh reviewer checks the synthesized plan, repeating up to five review/synthesis rounds. Once the plan is locked, Trycycle builds a test plan, builds the code, sends it to a fresh reviewer, turns the review into a structured observation packet, fixes what that packet shows, and repeats that loop too (up to eight rounds by default). If blockers persist, Trycycle runs plan reconsideration after the 4th review round and every 2 rounds after that, plus once before nonconvergence any time the loop stops with blockers. Each review uses a fresh reviewer, and synthesis uses a fresh planning agent, so stale context does not accumulate where a clean judgment matters.
Curious how Trycycle works under the hood? Walk through the full state machine — every gate, prompt template, binding, review loop, and outcome — with pre-loaded sample inputs. No install required.
Trycycle's planning, execution, and worktree management skills are adapted from superpowers by Jesse Vincent. The hill-climbing dark factory approach was inspired by the work of Justin McCarthy, Jay Taylor, and Navan Chauhan at StrongDM.