Add structured planning, delegation, and review discipline to Claude Opus or Sonnet coding sessions.
Run multiple AI agents on the same hard problem and compare their answers to catch real issues.
Cut token cost by routing well specified implementation work to a cheaper model while keeping review on a stronger one.
| sambwarren/parable | abuzar-ansarii/androidlinux-gpu | flashpoint493/unreal-angelscript-skills | |
|---|---|---|---|
| Stars | 20 | 20 | 20 |
| Language | Shell | Shell | Shell |
| Setup difficulty | moderate | moderate | easy |
| Complexity | 2/5 | 2/5 | 2/5 |
| Audience | developer | developer | developer |
Figures from each repo's GitHub metadata at analysis time.
Requires filling in several project-specific placeholders and merging hook config into settings.json.
parable is a set of Claude Code skill files and hooks that try to give Claude Opus and Sonnet the same tight working discipline the author noticed in Claude Fable 5. It is an unofficial, community made project, not built or endorsed by Anthropic, and the author is upfront that its claims come from one person's study on one project rather than a broad benchmark. The idea behind it is that the gap between a careless AI session and a tightly run one is mostly about procedure, not raw model intelligence. So the repo packages that procedure as files you drop into your Claude Code setup: a skill folder with instructions for planning, delegating, and reviewing work, plus hook scripts that keep reminding the model of its rules as a session runs long. The author found that these reminders matter because rules tend to fade from a model's attention during a lengthy task, so a script re-injects a condensed version of the rules every few minutes. The system defines roles. A director plans the work out loud and delegates it but does not write code itself. Workers carry out the delegated tasks quietly and mark their progress with simple done markers. A coordinator sits above both, running multi step programs and checking results by hand rather than trusting a model's own claim that something was verified. For harder problems, the project supports running several agents on the same prompt without letting them see each other's work, then having a separate agent compare their answers, on the idea that disagreements between them tend to point at real issues. Setup involves copying the skill folder into a personal Claude skills directory, filling in a handful of placeholder values like the project's root path and build check commands, and merging a provided hook configuration into a project's Claude Code settings file. Usage is meant to be simple: a user prompts normally and adds a short mode suffix, such as asking for a cheaper model on well specified work or asking for the multi agent review mode on harder problems. The repo also includes a written testing protocol so others can measure the approach on their own projects rather than trust the author's numbers directly.
A community-made Claude Code skill and hook set that tries to make Opus or Sonnet work with the disciplined planning and review process seen in Claude Fable 5.
Mainly Shell. The stack also includes Claude Code, Shell, Markdown.
The README does not state a license.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
Verify against the repo before relying on details.