fable5-skills

Fable 5 Skills

AI reasoning skills, mathematically distilled from 4,665 real Claude Fable 5 chain-of-thought traces.

Tuned to Grade A · 100.00% emulation accuracy against MiniMax M3 across 9 iterative test rounds.

Installation · The 5 Skills · Benchmark · CLI Reference · How It Was Built

Why

Modern coding agents are powerful, but they often skip verification, guess instead of investigate, and leave loose ends. Fable 5 doesn't. Across 4,665 real turns captured in the source dataset, Fable 5:

• Verifies after 87.7% of actions
• Reads the file before editing in 93.5% of edits
• Self-corrects in 56.4% of reasoning turns — then continues forward 74.4% of the time
• Uses a tight per-turn loop: ACKNOWLEDGE → OBSERVE → EXECUTE → VERIFY
• Speaks in professional, action-heavy prose (4.39 tool-to-text ratio)

fable5-skills packages these measurable behaviors into 5 drop-in SKILL.md files so any agent that reads markdown — Claude Code, Cursor, Cline, Windsurf, Continue, MiniMax M3, any custom agent — can adopt Fable 5's discipline in one command.

npx fable5-skills init

📦 Installation

You don't need to install anything globally. Just run npx:

# Default — installs into .claude/skills/ (Claude Code layout)
npx fable5-skills init

Other agent layouts

npx fable5-skills init --agent=cursor       # .cursor/skills/
npx fable5-skills init --agent=cline        # .cline/skills/
npx fable5-skills init --agent=windsurf     # .windsurf/skills/
npx fable5-skills init --agent=continue     # .continue/skills/
npx fable5-skills init --agent=generic      # ./skills/
npx fable5-skills init ./my-project         # install into a specific project

Install via skills.sh (open agent skills marketplace)

This package is also indexed on skills.sh — the open agent skills directory. If you use the skills CLI, you can install any of the 5 Fable 5 skills into 70+ supported coding agents:

# List available skills
npx skills add ahmdd4vd/Fable5res --list

# Install a specific skill (e.g. fable-debug) into Claude Code + Cursor
npx skills add ahmdd4vd/Fable5res --skill fable-debug -a claude-code -a cursor

# Install all 5 skills into all detected agents
npx skills add ahmdd4vd/Fable5res --all

The skills CLI auto-detects which coding agents you have installed (Claude Code, Cursor, Cline, Codex, Windsurf, Continue, Gemini CLI, GitHub Copilot, OpenCode, and 60+ more) and installs the SKILL.md files into the correct paths for each.

Useful flags

npx fable5-skills init --only=fable-debug   # install a single skill
npx fable5-skills init --force              # overwrite existing
npx fable5-skills init --dry-run            # preview without writing
npx fable5-skills doctor                    # verify package integrity

🧠 The 5 Skills

Each SKILL.md is self-contained, composable, and verifiable against the source dataset.

Skill	Purpose	Core Loop
fable-think	Foundational per-turn reasoning	ACKNOWLEDGE → OBSERVE → EXECUTE → VERIFY concise per turn, repeats across turns
fable-code	Writing & editing code	Read → Understand → Plan → Write → Verify → Iterate
fable-debug	Hypothesis-driven root cause analysis	OBSERVE → INVESTIGATE → HYPOTHESIZE → ROOT CAUSE → FIX → VERIFY
fable-architect	System design via vertical slices	UNDERSTAND → DESIGN → VERTICAL SLICE → VERIFY → ITERATE
fable-verify	Quality assurance vocabulary	5-phrase verification: should be · to verify · to ensure · to confirm · to make sure

The skills compose cleanly. For example, fable-debug opens with fable-think's per-turn loop and closes with fable-verify's verification vocabulary.

📊 Benchmark Results

Static skill verification — claims vs. actual data

Every quantified claim in every skill file is verified against the source dataset:

✓ Passed    31 / 31 claims
✗ Failed    0 / 31 claims
■ Pass rate 100.00%

Claims passed     ████████████████████████████████  100.00%  31/31

MiniMax M3 emulation — Round 9 (final)

After 9 iterative test rounds, MiniMax M3 emulates Fable 5's reasoning patterns at 100.00% across 7 test scenarios:

Test	Scenario	Score	Checkpoints	Bar
T1	reasoning	100.00%	11/11	████████████████████ 100%
T2	debug	100.00%	9/9	████████████████████ 100%
T3	code	100.00%	8/8	████████████████████ 100%
T4	verify	100.00%	10/10	████████████████████ 100%
T5	architect	100.00%	8/8	████████████████████ 100%
T6	self_correct	100.00%	6/6	████████████████████ 100%
T7	loop	100.00%	7/7	████████████████████ 100%
	Average	100.00%	59/59	████████████████████ Grade A

Iterative improvement curve

Round	Score	Δ	Key Fix
1	80.82%	—	baseline harness
2	93.93%	+13.11	extract <think> block as scored text
3	93.99%	+0.06	strengthened prompts
4	96.79%	+2.80	bumped max_tokens for long scenarios
5	97.96%	+1.17	explicit "MUST use ALL 5 phrases" in T4
6	97.40%	−0.56	stochastic variance on T1
7	97.62%	+0.22	broadened T6 continues_forward check
8	98.70%	+1.08	added self-verify instruction to T1
9	100.00%	+1.30	explicit verbatim "Thus"/"Therefore" instruction

See docs/GRADE_A_FINAL_REPORT.md for the full report.

🎯 Behavioral Profile (from 4,665 traces)

These are the measurable signatures of Fable 5's reasoning that the skills encode. Every number is verified against the source dataset.

Chain-of-thought structure

Metric	Measured
Traces with CoT	100.0%
Average words per CoT	409
Average paragraphs per CoT	7.19
CoTs opening with "Alright,"	53.1%
CoTs opening with "Okay,"	10.8%

Linguistic signature

Metric	Measured
First-person pronouns (share of all pronouns)	75.6%
First-person pronouns per CoT	11.29
Contractions per CoT	1.53 (professional, not casual)
CoTs with self-correction	56.4%
CoTs containing "actually"	32.4%
CoTs containing "however"	23.0%
Reasoning connectors per turn (because/since/therefore/thus)	2.14

Tool usage & verification

Metric	Measured
Tool-use turns (vs. pure-text turns)	81.4%
Tool-to-text ratio (action-heavy)	4.39
Traces using inline code with backticks	90.9%
Read-before-Edit rate	93.5%
Verify-after-action rate	87.7%
Top tools: Bash (1,544), Edit (960), Read (443), Write (311), PowerShell (136)

Session-level patterns

Metric	Measured
Average turns per session	77.75
Median turns per session	38
Max turns in a session	439
Sessions with hypothesis-driven debugging	68.3%
Sessions with same-turn fix attempts	78.3%
Top per-turn step combo: ACK+SCO+VER+ITR	19.2% of turns

🛠 CLI Reference

fable5-skills init [target] [--agent=<name>] [--force] [--only=<id>] [--dry-run]
fable5-skills list
fable5-skills show <skill-name>
fable5-skills doctor
fable5-skills --version
fable5-skills --help

Flag	Description
--agent=<name>	Target agent layout: claude-code (default), cursor, cline, windsurf, continue, generic
--force	Overwrite existing skill files
--only=<id>	Install only one skill (e.g. --only=fable-debug)
--dry-run	Print what would be copied, but do not write

Run fable5-skills doctor after install to verify package integrity (checks all 5 skills present + 31/31 verification claims + Node version).

🧑‍💻 Usage After Install

Once installed, the skills live at (for example) .claude/skills/ in your project. Three ways to use them:

1. Reference from your system prompt:

When you encounter a bug, follow the methodology in
.claude/skills/fable-debug/SKILL.md.

2. Load directly into context:

cat .claude/skills/fable-think/SKILL.md
# paste the contents into your agent's context window

3. Auto-load via agent config: Some agents (Claude Code, Cline) can be configured to auto-load skills from a directory. See your agent's docs.

🔬 How It Was Built

1. Source dataset

Distilled from the public HuggingFace dataset Kuberwastaken/Fable-5-traces — 4,665 rows, 60 sessions of real Claude Fable 5 chain-of-thought traces captured from production coding sessions.

⚠️ License note: The source dataset is licensed under AGPL-3.0. Because fable5-skills is derived from that dataset, this package is also distributed under AGPL-3.0 to comply with the copyleft terms. See LICENSE for the full text. If you modify or redistribute this package, or build a network service on top of it, you must publish your source code under AGPL-3.0 as well.

2. Quantitative extraction

A deep statistical pass over all 4,665 traces extracted:

• Per-CoT structure (word count, paragraph count, sentence count, opener words)
• Pronoun & contraction distribution
• Self-correction markers (actually, however, oops)
• Verification vocabulary (should be, to verify, to ensure, to confirm, to make sure)
• Tool-to-text ratio, tool-usage distribution, read-before-edit rate
• Per-turn reasoning step coverage
• Session-level stats: turns/session, hypothesis-driven debugging rate, same-turn fix rate

The full extraction is shipped in data/DEEP_STATS.json (raw stats) and data/VERIFICATION_REPORT.json (claim-by-claim verification).

3. Skill authoring

Each SKILL.md is written from the data — not from intuition. Every percentage, count, and ratio in every skill is verifiable against data/VERIFICATION_REPORT.json.

4. Cross-model emulation benchmark

We tested whether the skills actually make a different model (MiniMax M3) reason like Fable 5. We built a 7-scenario test harness (reasoning, debug, code, verify, architect, self_correct, loop) and ran 9 iterative test rounds, driving the score from 80.82% to 100.00% through systematic diagnosis and targeted fixes.

📁 Repository Layout

fable5-skills/
├── bin/
│   └── fable5-skills.js          # CLI installer (Node.js, ESM, zero deps)
├── skills/
│   ├── fable-think/SKILL.md
│   ├── fable-code/SKILL.md
│   ├── fable-debug/SKILL.md
│   ├── fable-architect/SKILL.md
│   └── fable-verify/SKILL.md
├── data/
│   ├── DEEP_STATS.json           # Full statistical extraction from 4,665 traces
│   └── VERIFICATION_REPORT.json  # 31/31 claim verification
├── docs/
│   └── GRADE_A_FINAL_REPORT.md   # Full Grade A benchmark report
├── package.json
├── LICENSE                       # AGPL-3.0
├── CHANGELOG.md
└── README.md

⚖️ License

GNU Affero General Public License v3.0 only — see LICENSE.

This package is derived from the Kuberwastaken/Fable-5-traces dataset (also AGPL-3.0). The copyleft is therefore inherited. In plain terms:

• ✅ You may use, study, modify, and redistribute this package.
• ✅ You may build network services on top of it.
• ⚠️ If you modify or redistribute (including as a network service), you must publish your source code under AGPL-3.0.

For enterprise use that cannot tolerate AGPL, contact the maintainer to discuss a dual-license arrangement.

🙏 Acknowledgements

• Anthropic PBC — for Claude Fable 5, the model whose reasoning patterns this project distills.
• Kuberwastaken — for publishing the Fable-5-traces dataset under AGPL-3.0, without which this project would not exist.
• MiniMax Inc. — for MiniMax M3, used as the cross-model emulation benchmark.

🚫 Disclaimer

Claude, Fable 5, and Anthropic are trademarks of Anthropic PBC. This project is not affiliated with or endorsed by Anthropic. MiniMax is a trademark of MiniMax Inc. This project is not affiliated with or endorsed by MiniMax. The skill files in this package were distilled from publicly available traces under the terms of their original AGPL-3.0 license; this package inherits that license.

_{Built with mathematical precision. Tuned over 9 rounds. Verified to 100%.}

Report a bug · Request a feature · Changelog