Shipping an app without coding: Claude designs and reviews, a cheaper model writes the code in OpenCode, and scripts gate the tests
Chinese builder @niukutzu open-sourced the workflow behind an Android app (~70 task packages, shipped): Claude only designs, splits tasks and reviews; ChatGPT/DeepSeek inside OpenCode write the code; Bash scripts gate on real exit codes and bounce failures back for fixes.
Visit site Claude Code + OpenCode (claude-dev-workflow) ↗- Pricing
- Paid
- Region
- Global
- Where it's from
- Chinese community
- Added
- Oct 17, 2026
- Tested on
- Oct 17, 2026
- Version we used
- niukutzu/claude-dev-workflow @ e276897 (first release, 2026-10-07)
- Categories
- CodingAutomation & Agents
- Original author
- niukutzu @niukutzu ↗
- Our test
- Partly tested
- Cost / time
- Template is free (MIT). You pay for Claude Code plus whatever API/subscription your OpenCode coding model uses — the whole point is moving coding tokens off the expensive model.
The task
A non-coding product owner wants to keep the expensive model for design and review, let a cheaper model do the coding, and have scripts — not the model's own report — decide whether tests passed.
Steps
- Copy the files from github.com/niukutzu/claude-dev-workflow into your project root (MIT).
- Install Claude Code and OpenCode and log OpenCode into your coding model. The template defaults to openai/gpt-5.6-sol and falls back to deepseek/deepseek-flash#high when quota runs out (edit tools/workflow.conf).
- Edit GATE_STEPS in tools/workflow.conf — one line per step, "name :: error-summary regex :: command" — with your real build/test commands; replace every <placeholder> in AGENTS.md (modules, versions, hard bans).
- In Claude Code, say "read HANDOFF.md first" and describe the feature. Flow: scenario list → glossary → clickable HTML mock → you say "final" → change request + task packages (goal + tests Claude writes first).
- Claude dispatches each package to OpenCode with tools/dispatch-coder.sh; afterwards tools/review-gate.sh runs automatically, failures go back for up to 2 fix rounds, then Claude reviews independently (PASS / changes needed).
- Track progress in docs/status.html (python3 tools/build-status.py renders it from status-data.json in plain language). Switch chats via HANDOFF.md; run tools/flow-stats.py every ~3 tasks as a health check.
Prompt
Read HANDOFF.md and AGENTS.md first. I want to build an expense-tracking app: give me a scenario list and a clickable HTML mock first; only after I say "final", write the change request and task packages with tests written up front, then dispatch them with dispatch-coder.sh.
What the author says
@niukutzu (2026-10-07) says they built an app with Claude Code + OpenCode (ChatGPT/DeepSeek via API), about 70 tasks from requirements to release, and open-sourced the workflow; the README says it came from a shipped Android app. In a reply they add that a handoff doc tracks current progress.
Our test
We cloned the repo and ran the model-free parts on Linux: the gate script, the token-saving wrapper and the status board generator. We did not install OpenCode or have any model write code, so the "70 task packages, shipped" claim is not reproduced.
What we found when we tried it:With GATE_STEPS set to one always-pass step (true) and one always-fail step (a Python assert), review-gate.sh printed "[通过]" / "[失败] exit=1", surfaced only the AssertionError line, and exited 1 overall. run-quiet.sh returned 0/1 for success/failure, printing a summary while full logs went to .coordination/logs/. build-status.py generated docs/status.html. dispatch-coder.sh passed a bash syntax check.
How to get in
Repo is on GitHub with Chinese rules and an English summary. The scripts are plain Bash/Python, so non-Chinese readers can use the structure after translating the rule files; OpenCode works with any provider you have.
Who it suits
Non-coding indie builders and product owners; Claude Code users who want lower cost but fear cheaper models drifting off-spec.
Pitfalls
- Rule files are written in Chinese (the original owner reads Chinese); translate them yourself — the structure is language-agnostic.
- Replace every placeholder in AGENTS.md and workflow.conf, or the gate runs the sample command (npm test by default).
- dispatch-coder.sh needs the opencode CLI on PATH (or the CLI bundled with the macOS desktop app).
- Cheap models drift: task packages must say what not to touch, Claude writes the tests and the coder may not edit them; the README caps tasks at 30–40 turns with auto-pause handoffs.
- Brand-new repo (first commit 2026-10-07), no community track record yet — start with the read-only TRIAL_DISPATCH task.
More in this category
- A demo video per docs page that flags itself when the UI changesVideoScreenplayClaude CodeTool free; uses your agent's quota◐ Partly testedOverseas
- AI backend code that guards against races, leaks and bad migrationsCodingbackendskillsClaude CodePostgreSQLFree; about 2× tokens per task◐ Partly testedOverseas
- An afternoon of AI work, a 3D action game you play in the browserCodingClaude CodeTripoJimengSkill free; Claude quota plus asset tools◐ Partly testedChina
- Keep a long-running AI-built project from turning into a messCodingClaudeCodexoil-delegateNeeds Claude and Codex plans◐ Partly testedChina
- Use a web app once and get an AI write-up of how it worksResearchREAClaude CodeChromeFree tools; needs Claude or Codex◐ Partly testedOverseas