Talos, the bronze automaton, walked the same circuit of Crete three times a day without supervision. This runs a plan the same way.
A sequential loop for working a plan to completion overnight. One item at a time, implemented by a subagent, attacked by an independent red cell, ruled on by an independent judge, then committed by the main session.
- Sequential. One item at a time, clean tree between items. No worktrees, no lanes, no merge gate, no requeue.
- Adversarially verified. The implementer never grades itself. The red cell sees the item text and the diff, never the implementer's account of it. The judge runs on every item, including the ones the red cell passed.
- The main session drives. All plan edits and all commits are made by the driving session, never by a subagent. Subagents return text; the driver acts.
- State is on disk, not in context.
plan.md+decisions.md+ git history are the entire state, so/clearand re-invoke resumes for free.
| Path | What it is |
|---|---|
skills/talos/SKILL.md |
The driver procedure, and the bulk of the harness. |
skills/talos/FORMAT.md |
plan.md / decisions.md format spec. |
skills/talos/prompts/implementer.md |
Implementer subagent prompt. |
skills/talos/prompts/red-cell.md |
Red-cell subagent prompt. |
skills/talos/prompts/judge.md |
Judge subagent prompt. |
skills/talos-mini/SKILL.md |
Ad-hoc variant: distil a conversation's task list into mini-plan.md, then run the same loop. |
mkdir -p <repo>/.claude/skills
cp -r skills/talos <repo>/.claude/skills/talos
cp -r skills/talos-mini <repo>/.claude/skills/talos-miniThen write a plan.md in that repo, in the format given in
skills/talos/FORMAT.md. Author it in an ordinary session with a strong model:
a badly written plan is the one failure the loop cannot recover from. Commit
it, then run /talos.
decisions.md is created by the loop on first need. For task lists agreed in
conversation, run /talos-mini and skip the plan-authoring session.
plan.md:[x]done,[ ]not reached,[!]blocked.decisions.md: every judge override, plan edit, and block, with reasoning.git log --oneline: onefeat:orfix:commit per item, each optionally followed by itsrefactor:simplification commit.