Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 
 
 
 
 

Repository files navigation

fable-prep

Stage your hardest work now. One-shot it the moment your best model arrives.

Your most capable model is the chef you can only book for a few hours — and the booking may turn metered the instant it's back. You don't make that chef read the menu. You prep every station so it only cooks.

The edge of a frontier model isn't "only it can do this." Cheaper models are good. The edge is round-collapse: it one-shots what they reach only after rounds of iterate → review → fix. So you do the slow, cheap work — diagnosis, selection, sequencing, acceptance criteria, test scaffolds — now, on a cheaper model. You produce execution-ready packets. When the frontier model returns, you fire one loop and it straight-executes with zero planning tax.

One loop, two modes

detect frontier availability
 ├─ UNAVAILABLE → PLAN-MODE    diagnose repos → adversarially verify → rank → write a wave queue
 └─ LIVE        → EXECUTE-MODE  next wave → run packets → tiered gates → pause for your nod

It flips automatically on a sentinel file. No frontier access yet? It plans. Frontier live? It ships. Same command either way.

What makes the queue worth more than a TODO list

  • Ruthless selection, not a dump. A 500-item pile wastes a scarce model as badly as no prep. Each task is scored by rounds saved × ceiling lift; anything a cheaper model one-shots too is dropped from the queue.
  • Cross-domain. Code, design/UI, sales copy, marketing — anywhere a frontier model collapses rounds, not just hard engineering.
  • Execution-ready packets. Concrete files, exact operation, an acceptance criterion, and a machine_check shell oracle that proves done. No re-planning on expensive time.
  • Findings are adversarially verified before they enter the queue — no phantom work.
  • Safety rails built in. Never touches money/trading paths, archives instead of deleting, respects your branch policy, gates every commit, pauses at each wave boundary.

Install

git clone https://github.com/Dallionking/fable-prep ~/.claude/skills/fable-prep
cp ~/.claude/skills/fable-prep/config.example.json ~/.claude/skills/fable-prep/config.json
bash ~/.claude/skills/fable-prep/lib/detect.sh --init   # auto-discovers your repos + audit tools

Use

/fable-prep            # detect mode and go
/fable-prep plan       # force diagnosis → refresh the queue
/fable-prep execute    # force execution (needs the live sentinel)
/fable-prep status     # show the queue summary + next wave

Configure your models, repos, and rails in config.json. The core is repo-agnostic and detects what exists — bring your own stack.


Built the way the June 2026 harness-optimization was: one frontier pass, a real before/after report, and a skill anyone can install.

About

Stage your hardest work now. One-shot it the moment your best model arrives. A Claude Code skill for frontier-model prep.

Resources

Stars

8 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages