| name | cheaploop |
|---|---|
| description | Compact a running context into `fable.md`, then send the rewritten prompt onward with `cheaploop` disabled by `cheaploop stop`. Use when a session is getting expensive, when you want a repeatable compress-then-send loop, or when browser/Chrome work should stay on the cheapest available model path. |
Use this skill when a prompt needs one compact pass before it is sent onward.
- Read the current conversation and any local project docs that matter.
- Compress the prompt into a shorter, cleaner form.
- Write the compressed handoff to
fable.md. - Send only the compacted prompt onward.
- Keep the rewritten prompt focused on the active task.
- Preserve facts, constraints, file paths, and verification criteria.
- Drop repetition, brainstorming debris, and stale context.
- Treat
cheaploop stopas a hard bypass: no compression, nofable.mdrewrite, no extra routing. - Keep browser/Chrome work on the cheapest available model or lightest viable path.
- Never route Chrome work through the expensive Fable-style context loop unless the task explicitly requires it.
Record the handoff in fable.md before the downstream prompt is sent.
Minimum fields:
- source prompt
- compacted prompt
- why content was dropped
- any browser/Chrome routing decision
- rough before/after size
When installing this skill in another workspace:
- add the skill folder under that workspace's Claude skills directory
- add a project note in that workspace's
CLAUDE.mdsaying browser/Chrome work should use the cheapest available model path - keep
cheaploop stopas the explicit bypass phrase
Use a small fixture set:
- short prompt
- context-heavy prompt
- stale-context prompt
cheaploop stopprompt
Pass criteria:
- the compacted prompt is shorter
- critical facts survive
- stale facts are removed
- stop bypasses compression
- Chrome routing stays separated from the expensive loop