|
// you commented fix
never hit your Claude limit again
The reel showed you the idea. Here’s the whole setup — the 3-step loop, the exact commands, the role prompts to steal, and how to run it on autopilot.
| > the_idea WHY IT WORKS | Your best model ‚Äî Fable 5 or Opus 4.8 ‚Äî is too powerful (and too expensive) to waste on your smallest tasks. So you stop using it as the doer. Instead it becomes the planner and the judge. A cheaper, faster model (Sonnet or Haiku) does the grunt work. Your top model reviews the result and loops until it’s right. Same usage limits, far more output. |
| > the_loop 3 STEPS | 1. PLAN — your top model breaks the job into a clear plan and acceptance criteria. 2. EXECUTE — a cheap/fast executor subagent does the actual work, cheaply and quickly. 3. JUDGE — your top model reviews against the criteria. Not good enough? It loops back to step 2. Done? It stops. |
| > build_it IN CLAUDE CODE | Create an executor subagent pinned to the cheaper model: | /agents ‚Üí new subagent "executor", model: haiku |
Then wrap the work in the goal command so it self-checks against your criteria and keeps going: | /goal "ship X until: tests pass, lint clean, matches spec" |
And let it run on its own with a loop or a schedule: /loop "work the goal, then re-check criteria"
# or schedule it to run unattended |
Top model plans + judges. Cheap model executes. The loop closes the gap while you do something else. |
| > the_prompts STEAL THESE ROLES | “You are the PLANNER. Break this into steps + explicit acceptance criteria. Do NOT write the code.”
forces the smart model to think, not grind “You are the EXECUTOR (Haiku). Do exactly step N. Nothing more.”
cheap model stays in its lane “You are the JUDGE. Score the result against the criteria. If it fails, say exactly what to fix.”
the review step that makes it actually work “Keep looping execute ‚Üí judge until every criterion passes, then stop.”
the whole system in one line |
| > pro_tips FROM RUNNING IT | 1. The magic is in the criteria. Vague criteria = endless loop. Be specific: tests pass, matches this spec, under X lines. 2. Pin the executor to the cheapest model that can do the task. Upgrade only if it keeps failing judgement. 3. Cap the loop (e.g. max 5 rounds) so a hard task can’t burn your whole budget. 4. Schedule it overnight. You wake up to work that already passed your own bar. |
Set it up on one task today, then reply and tell me what it shipped while you were away.
Talk soon,
Mike
|