fable-skills
39 engineering-discipline skills and 5 subagents for Claude Code: root-cause debugging, security, edge cases, honest reporting
- 99 ★ on the repository
- Last commit
Minimal access
Why this level
- The skills and subagents are instructions and run nothing themselves
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/adamentwistle-fable-skillsDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewMinimal accessThe tool runs nothing and makes no network calls.
- The skills and subagents are instructions and run nothing themselves
This is third-party code. Review the repository files before installing.
What it does
fable-skills packs the habits of a careful engineer into 39 short skills. Each fires situationally: `root-cause-debugging` for a bug hunt, `security-reflexes` for a diff that touches user input, `data-migration-safety` for a schema change. Others cover codebase orientation, migrations, concurrency, change review, honest reporting, untrusted-content defense, scoping and git hygiene. The `agents/` folder holds five subagent definitions for Claude Code's Agent tool: `blind-judge`, `done-gate`, `fresh-eyes-reviewer`, `plan-premortem` and `skeptic`. They run in a clean context, so they do not inherit the main session's wrong assumptions. The verification agents are pinned to the opus model, and the model field can be changed. The author ships a reproducible benchmark: 315 graded runs over nine task areas. By the author's numbers, Opus 5 with the skills scored 95% against 94% for Fable 5.1 bare, and Haiku 4.5 gained 9 points, though it lost slightly on debugging and edge cases.
Who it is for. Developers who run Claude Code on Sonnet or Opus and want review, debugging and reporting discipline closer to a senior model.
Good fit when
- A bug keeps coming back and the agent should prove the cause before fixing it
- A diff touches a shared interface and the review should list every consumer
- The agent reads emails and web pages and needs the `untrusted-content-guard` skill against injection
- Before handing in a task you want an independent check by a subagent in a clean context
Not a fit when
- You need framework-specific advice for React or Django: the skills describe work habits and no particular library
- The model is small and context is tight: a 10K-character bundle can distract Haiku
- The task is trivial and extra checklists would only slow the answer
Example request
Load the root-cause-debugging skill from fable-skills and find the cause of the failing test before changing codeLimitations
- The author's runs are single-turn with no tools and measure what a model says, not how it behaves across a long agentic session. The author calls n = 5 per cell directional.
- The skills were tested as system-prompt instructions; other placements are not measured in the README.
- On debugging Haiku and Sonnet scored slightly lower with the skills, and Haiku lost 0.8 on edge cases.
- Subagents such as `blind-judge` are pinned to opus, the priciest model of the line.
How to disable. Delete the skill folders from `~/.claude/skills/` or the project's `.claude/skills/`, and the agent files from `~/.claude/agents/`. Changes take effect on the next session start.
FAQ
How do I install fable-skills in Claude Code?
Copy the skill folders you need into `~/.claude/skills/` or the project's `.claude/skills/`, and the Skill tool picks them up by description. Subagents go into `~/.claude/agents/` and register at the next start.
What do the skills deliver in the author's benchmark?
The set lifts the mean for Haiku 4.5 from 44% to 53%, Sonnet 5 from 74% to 80% and Opus 5 from 90% to 95%. These are the author's own numbers over 315 runs, and `benchmark/harness.py` reproduces them.
How do the subagents differ from the skills?
A skill nudges the main model toward a habit, while a subagent runs in a separate clean context and does not share its assumptions. That is why self-review goes to `fresh-eyes-reviewer` or `skeptic`.
How many skills are there?
Thirty-nine, one per habit: `codebase-orientation`, `security-reflexes`, `data-migration-safety`, `edge-case-sweep`, `test-design`, `git-hygiene`, `stop-thrashing` and others. Each sits in its own folder with a `SKILL.md`.
Can I use the set in Cursor or Codex?
The repository contains `AGENTS.md`, but the README describes installation only for Claude Code. In another agent the skills have to be loaded by hand as plain instructions.
Discussion
No questions or reviews yet.
Related
Skills14 skills that find the gaps in a task before the code: interviews, plans, assumption tests and proof that a bug fix works
Claude Code kit with eight phase agents, 21 skills and hooks that keep the agent inside set limits.
Over twenty skills for coding agents: MCP and hooks, plan mode, a bug-fix protocol, prompting and a bash safety hook
A library of 30-plus instruction skills for coding agents: planning, building, testing, hallucination guards and shipping
A skills library for Claude Code, Cursor and Codex: team code review, TDD, stacked PRs and repository readiness scoring
Nine personal engineering skills: an idea-to-PR chain, standards search, a sandboxed review loop and wizard pages for changes
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| fable-skills (this card) | no | 99 | |
| Superpowers | no | 316,679 | a development process with clarification, a plan and TDD, while fable-skills focuses on discipline and independent checks |
| Karpathy-Inspired Skills | no | 124 | fourteen skills inspired by Karpathy's posts, with no subagents |
README in short
The README explains the idea: write a strong model's habits down as explicit checklists. Next come the Benchmark 3 results with a table over nine areas, five findings, caveats about limits and a catalog of skills with descriptions. A separate section covers the five subagents and installing them into `~/.claude/agents/`.