Menu
Split the Plan/Act/Verify loop across three model tiers — frontier planner, cheap executor, frontier judge — via env vars read by run.sh.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
| name | model-routing |
| description | Split the Plan/Act/Verify loop across three model tiers — frontier planner, cheap executor, frontier judge — via env vars read by run.sh. |
| when_to_use | when a long unattended loop is being tuned for cost, or when you want a stronger judge than executor. |
run.sh reads three optional env vars and threads them into each claude invocation as --model. All default to unset, in which case the CLI default model is used (behaviour unchanged from a bare run).
CLAUDE_PLANNER_MODEL — reserved for /spec workflows that draft PROMPT.md up front. Not read by the current run.sh loop, but claimed here so future planner passes bind to it.CLAUDE_EXECUTOR_MODEL — used on the "do the next step" call. This is the workhorse; it runs on every iteration. Pick something cheap and fast.CLAUDE_JUDGE_MODEL — used on the /verify call. Runs once per iteration to adversarially check the executor's diff. Pick a frontier model — a weak judge is worse than no judge.planner = frontier (Opus-class, runs once at /spec time)
executor = cheap-fast (Haiku-class, runs every turn)
judge = frontier (Opus-class, runs every turn but on a small diff)
export CLAUDE_PLANNER_MODEL="claude-opus-4-7"
export CLAUDE_EXECUTOR_MODEL="claude-haiku-4-7"
export CLAUDE_JUDGE_MODEL="claude-opus-4-7"
./run.sh
A cheap executor paired with a frontier judge outperforms a frontier executor with no judge on long loops. The judge catches the executor's premature-victory claims that a mono-model run rationalises away when it runs out of context. Cost stays low because the judge only sees the diff, not the working history.
Before compaction Loopkit extracts decisions into claude-decisions.json (machine-readable). Read it alongside claude-progress.txt at session start — prose is for humans, JSON is for the loop.
Run the fixed 6-step session-opening sequence — pwd, read progress, git log, count remaining features, init.sh, smoke-test last feature — before touching any new work. The orientation ritual that lets fresh-context sessions reconstruct project state in under a minute.
Read git log since initial scaffold + last 3 progress notes, draft feature additions to a SEPARATE suggestions file so feature_list.json stays immutable
Use when starting any conversation in a loopkit-enabled project - establishes how to find and use loopkit's 49 skills, requiring skill invocation before ANY response including clarifying questions.
One sentence that describes WHEN to invoke this skill. Write it as a trigger phrase, not a summary — include the words a user would say ("fix bug", "add feature", "review diff"). The agent routes on this field.
Escalate blocked runs to a human via configured channel or fallback to BLOCKED.md and exit the loop. Use when the loop hits an ambiguous spec, a missing credential, a destructive action needing approval, or 3+ consecutive verify failures on the same task.