A coding workflow for Claude Code in which nobody grades their own work. You and Claude Code plan a feature together; a different model attacks the plan and the acceptance tests; cheap models implement it goal by goal in a sandbox; and a separate Claude session reviews every goal's diff before it lands.

Role	Model	What it does
Architect	Claude Code (your session)	Interviews you, writes and locks PLAN.md, cuts it into a goal DAG (GOALS.md) with acceptance tests, dispatches, reads every stop
Attacker	Grok (grok-4.7), headless, read-only	Attacks the locked plan, then the acceptance tests ("oracles"), round by round
Implementer	DeepSeek Flash through DeepSeek's Anthropic-compatible endpoint, in headless sandboxed Claude Code; Claude Haiku 5.5 when DeepSeek refuses, declines or runs out of balance	Implements one goal at a time, test-first
Goal reviewer	Claude Opus 5.5, headless, separate session	Judges each goal's diff against the goal; DONE or FIX

The implement/review loop runs in host scripts, not in your session: a goal lands only when its proof (the acceptance test) is green and the reviewer says DONE. Anything else stops with a stop file that says why.

Optionally the workflow runs unattended ("always-on"): you sign a walk into a queue and a macOS LaunchAgent walks it, waits through usage limits, retries operational stops within fixed limits and messages you on iMessage for everything else. See docs/guides/always-on.md.

The flow

/crucible    recon + interview → PLAN.md → Grok attacks it → you sign off
/to-goals    PLAN.md → GOALS.md (a DAG of goals) + red acceptance tests → Grok attacks the tests
/run-goal    one ready goal: implementer → proof → Opus review → (FIX rounds) → done
/run-goals   walk the whole DAG, one goal at a time or in parallel waves (--jobs 2-4)

Each step is a Claude Code skill this repo installs. The full walkthrough is docs/guides/using-with-claude-code.md.

Requirements

• macOS (the implementer and proof sandboxes use macOS's sandbox; the always-on mode uses launchd and Messages).
• Claude Code 2.1 or later, signed in (claude login) or with an ANTHROPIC_API_KEY.
• Grok CLI (xAI's command-line client), signed in (grok login): the attacker.
• A DeepSeek API key (DEEPSEEK_API_KEY) for the default implementer. With IMPLEMENTER_PRIMARY=haiku, Haiku implements first and DeepSeek is only the fallback.
• git, bash, python3 (3.9 or later; the system one is fine), curl.
• A clone of mattpocock/skills at ~/mattpocock-skills for the borrowed helper skills (install.sh tells you the pinned commit).
• Optional: the Codex CLI, only for the legacy IMPLEMENTER=codex launcher.

Install

git clone -b release <this repo's URL> ~/claude-coding-workflow
git clone https://github.com/mattpocock/skills ~/mattpocock-skills
cp ~/claude-coding-workflow/.env.example ~/claude-coding-workflow/.env   # then set DEEPSEEK_API_KEY
~/claude-coding-workflow/scripts/install.sh

install.sh links the four skills (/crucible, /to-goals, /run-goal, /run-goals) and the borrowed ones into ~/.claude/skills, installs the two Claude Code mods (walk, lessons), runs the test suite and checks your setup. Details and the expected output: docs/guides/getting-started.md.

Clone it to exactly ~/claude-coding-workflow, though the repository is named Radiant: the skills call the scripts by that path.

Use it

In the repository you want to build in (the "target repo"), open Claude Code and copy this repo's CLAUDE.md into it (or merge it into the target's own), then:

/crucible add rate limiting to the public API

and follow the interview. After you sign off the plan, /to-goals, then /run-goals. While a walk runs, the walk mod keeps one task per goal in Claude Code's task list (ctrl+t) and /walk-status shows the DAG as a live graph.

Guides

Guide	For
Getting started	Installing, keys, install.sh --check, smoke tests
Using it with Claude Code	The phases, each skill, a walk end to end, reading a stop
Configuration	Every environment variable, where files live
Always-on	Unattended walks on a Mac: setup, signing, day to day
Troubleshooting	Stops, wakes, logs, common failures

What's in this branch

This is the release branch: the code needed to run the workflow and its instructions. The design history (every plan, the adversarial review logs, the build journals, evidence and research) lives on the main branch.

CLAUDE.md          the Architect's instructions: copy into a target repo
assets/            the icon
skills/            crucible, to-goals, run-goal, run-goals
scripts/           the host scripts (run-goal.sh, run-goals.sh, launchers, goals.py, ...) and tests/
mods/              Claude Code mods: walk (task list, /walk-status, always-on push), lessons
GOALS.example.md   a tiny goal DAG the install check and tests use
docs/guides/       the guides above

The name

In Isaac Asimov's Foundation, the Prime Radiant is the device that shows the Seldon Plan as glowing lines, where every step of the plan and every deviation from it can be seen. Radiant does that for a feature: a signed plan, cut into a graph of goals that light up as each one is built, proven and reviewed.

Credits

Built on ideas from chaseai-yt/crucible (MIT) and borrows helper skills from mattpocock/skills.

License

MIT — see LICENSE.