Load the skill so Claude Code tackles the riskiest part of a task first instead of last.
Force the model to back up any completion claim with pasted evidence rather than assertion.
Get a clear NOT DONE report when verification is incomplete instead of a false success claim.
Extend the skill yourself using its test-first process for adding new rules.
| wsbm393/fable-skill | 0xpira/sskills | 8thpark/geode | |
|---|---|---|---|
| Stars | 25 | 25 | 25 |
| Language | — | JavaScript | TypeScript |
| Setup difficulty | easy | easy | moderate |
| Complexity | 2/5 | 2/5 | 2/5 |
| Audience | developer | developer | vibe coder |
Figures from each repo's GitHub metadata at analysis time.
Requires Claude Code to load and trigger the skill automatically.
Fable Skill is a Claude Code Agent Skill that changes how the model handles hard, multi-step tasks: how it breaks a task down, how it checks its own work, and how it decides what to do next. Instead of writing generic advice about thinking carefully, the author first ran an unguided model through pressure scenarios such as deadline pressure, sunk cost, a tired user, and missing runtime environments, and watched exactly where it failed. The skill was then written to counter only those specific observed failures, tested again until each one stopped recurring, and reviewed by an adversarial critic looking for wording a pressured agent could talk its way around. The skill enforces rules in three phases. When decomposing a task, it requires stating what done actually looks like as something checkable, tackling the riskiest part of the work first instead of saving it for last, and fixing a pasteable checkpoint for each subtask before work begins. When verifying, it sorts claims into three buckets: verified with pasted evidence produced after the last edit, believed but not verified, and not checked at all, and it will not let the model declare a task done unless the done statement itself is in the verified bucket. When deciding what to do next, it tells the model to debug immediately on a failed checkpoint, to question its diagnosis rather than try a third patch after two failed fixes, and to stop and just list follow ups once a verified goal is met. The skill was built and tested using a test driven approach: failing scenarios first, then a minimal rule added only to fix an observed failure, followed by adversarial review to close loopholes. After release, nine pressure scenarios in total were tested against the shipped skill, and none of the candidate rules that failed to catch a new failure were added, keeping the rule set deliberately small. Installing it means copying a single SKILL.md file into the Claude Code skills folder, either with curl or by cloning the repository. It is released under the MIT license.
A Claude Code skill that changes how the model decomposes hard tasks, verifies its own work with pasted evidence, and decides what to do next.
Use freely for any purpose, including commercial use, as long as you keep the copyright notice.
Setup difficulty is rated easy, with roughly 5min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
double-check against the repo, no cap.