mrtooher/fable-mode

What It Does

Fable Mode is a doc-only Claude skill that imposes a four-step execution loop on complex tasks: write a numbered stage plan before acting, delegate independent stages to subagents where the runtime allows, verify each stage against a check that can actually fail, and self-critique the output before delivery. It targets tasks that span multiple files, multiple sources, or multiple sessions. The skill appeared on June 13, 2026 as the breakout of a "Fable wave," a burst of community skills recreating a staged-execution capability that Anthropic reportedly discontinued. The repo contains two distinct SKILL.md files: the namesake execution-discipline skill, and an unrelated joke-writing skill added through a merged community pull request.

The Good

The skill is unusually honest about its own limits. The SKILL.md states plainly that the skill shapes procedure, not capability, and that "coherence across long tasks and genuine self-correction live in the weights, not in a prompt." It calls itself "a checklist, not a capability transplant." This is rare in a skill riding a hype wave. Most execution-discipline prompts oversell; this one tells you when it will not help and includes a "When NOT to use this" section advising you to skip the loop for tasks with one obvious approach.

The EXAMPLE.md is concrete and demonstrates the concept catching real errors. Four worked examples cover software engineering, research, data analysis, and multi-session refactoring. Each shows a one-shot attempt that ships a plausible but wrong result, then runs the same task through the loop where the failable check (a test, a source lookup, a data-quality assertion) catches the bug. The API endpoint example, where a missing null check causes a 500 instead of a 401, is the kind of error a visual review passes. The examples make the abstract loop tangible.

The four-step loop is well-structured and domain-general. The core loop (stage map, delegate, verify with a failable check, self-critique) stays constant while only the verification artifact changes by domain. The "verify with a check that can fail" framing is the strongest idea here. The skill explicitly rejects "I reviewed it and it looks right" as a check, on the grounds that a model that skips verification will also pass its own introspection. That is a sharp observation worth the install on its own.

The joke-writing bonus skill is grounded in real research. The bundled writing-jokes skill cites eight arXiv papers on LLM humor (Jentzsch and Kersting 2023, Mirowski et al. 2024, HUMORCHAIN 2025) and bases its method on the finding that ChatGPT recycles roughly 25 jokes, with four covering half of all answers. The craft toolkit (punch word last, specific beats generic, misdirect, rule of three, cut) is practical comedy advice, not generic filler.

The Bad

There is no license file, and the GitHub API reports the license as null. For a skill whose entire value proposition is being copied into a skills directory and shared, the absence of a license is a real gap. Without one, default copyright applies, which technically reserves all rights to the author even though the README invites you to install and use it. A repo banking on community adoption and derivative "Fable wave" skills should ship a permissive license on day one.

An unrelated skill is bundled in, diluting the repo's focus. The writing-jokes skill arrived via pull request #2 from a community contributor and sits in a writing-jokes/ subdirectory. It is a competent humor skill, but it has nothing to do with staged execution discipline. A repo named "fable-mode" that markets itself as an execution-discipline skill shipping a joke generator creates confusion about what the repo actually is. The two skills should be separate repositories or the joke skill should be removed.

Install instructions are vague. The README says to "place the fable-mode directory wherever your Claude environment loads skills from," without naming a concrete path, a slash command, or a settings reference. A first-time Claude Code user has to already know where skills live. Compared to peer skills that give an exact path and a one-line install command, this is friction.

The repo is one day old and riding a hype wave, with no release tags. Created June 13, 2026 with 231 stars in under 24 hours, the growth is driven by the "Fable wave" narrative rather than proven usage. There are no tags, no releases, and no versioning. The .gitignore references a RockyVoice/ directory and a .claude/ path that have nothing to do with the skill content, suggesting the author reused a template without cleaning it. None of this is fatal, but it signals the packaging has not been pressure-tested.

Smoke Test Results

Tested on the host against a shallow clone of the repository in /tmp/fable-mode-review. The skill is doc-only (no scripts, no dependencies, no install step beyond file copy), so smoke testing reduces to structural validation and content verification.

Run A. Structural validation (host-based)

$ python3 validate-fable-mode.py
 SKILL.md: 130 lines
 writing-jokes/SKILL.md: 123 lines
 EXAMPLE.md: 243 lines
 Description length: 485 chars
 [OK] SKILL.md exists at root
 [OK] README.md exists
 [OK] EXAMPLE.md exists
 [OK] SKILL.md has YAML frontmatter
 [OK] Frontmatter has name field
 [OK] Frontmatter has description field
 [OK] SKILL.md file count is 2 (fable-mode + writing-jokes)
 [XX] LICENSE file present
 [OK] Description under 1024 chars
 [OK] writing-jokes has frontmatter
 [OK] writing-jokes has name
 [OK] writing-jokes has description
 [OK] No gitignored directory references
 [OK] SKILL.md has section about: Core Loop
 [OK] SKILL.md has section about: When NOT to use
 [OK] SKILL.md has section about: Domain-specific
 [OK] SKILL.md has section about: What this skill doesn't do
 [OK] EXAMPLE.md has at least 3 worked examples (found 4)
 [OK] README has Installation section
 Total: 18 passed, 1 failed out of 19

Pass rate: 18 of 19. The single failure is the missing LICENSE file, which the GitHub API confirms is null. All frontmatter, section structure, and cross-reference checks pass.

Run B. Functional verification (content and schema)

$ python3 functional-fable-mode.py
 PASS: main SKILL.md frontmatter parses as valid YAML
 PASS: name='fable-mode', description is str (485 chars)
 PASS: frontmatter has only name + description (minimal, spec-clean)
 PASS: writing-jokes frontmatter parses, name='writing-jokes'
 PASS: all 4 core loop steps present and in order
 PASS: all 4 domain patterns present
 PASS: honesty disclaimer present ('not a capability transplant')
 PASS: writing-jokes cites 8 arXiv IDs
 PASS: 4 worked examples, 7 explicit failure demonstrations
 PASS: no executable script files (confirms doc-only classification)
 All functional checks passed.

Pass rate: 10 of 10. The frontmatter parses as valid YAML with only name and description fields (minimal and spec-clean). The four-step core loop is present and ordered. The writing-jokes skill cites eight arXiv IDs in valid format. No executable files exist, confirming the doc-only classification.

What the runs tell you

The skill is structurally sound and internally consistent. The content is well-organized and the worked examples hold up under inspection. What cannot be verified here is whether the execution-discipline loop measurably improves output quality on a capable model, because that requires a live agent session and depends on the model's existing reasoning ability. The skill is honest about this limitation.

Setup Walkthrough

1. Clone the repository: git clone https://github.com/mrtooher/fable-mode.git
2. Copy the directory into your agent's skills folder. For Claude Code this is typically ~/.claude/skills/, so the skill lands at ~/.claude/skills/fable-mode/. The README does not name this path, so you must know it already.
3. The skill triggers on qualifying tasks (multi-file, multi-source, multi-session) or when you explicitly ask for systematic execution. There is no slash command or manual activation step beyond naming it.

There are no dependencies, no build step, and no configuration. The only gotcha is that copying the whole repo also installs the unrelated writing-jokes skill, since it lives in a subdirectory alongside the main SKILL.md. If you only want the execution-discipline skill, copy only the root SKILL.md and EXAMPLE.md and skip the writing-jokes/ folder.

Alternatives

1. obra/superpowers -- a much larger and more mature agentic-skills framework that includes subagent-driven development and planning methodology. Far broader in scope, with an established community and years of iteration. Choose it if you want a full development methodology rather than a focused execution-discipline loop.
2. OthmanAdi/planning-with-files -- persistent file-based planning for coding agents, focusing on crash-proof markdown plans that survive context loss. Complements fable-mode's session-spanning pattern with a concrete file-based implementation rather than a procedure to follow.
3. alicicek/tale-mode -- a sibling "Fable for Opus" skill from the same wave, emphasizing verifiable receipts and adversarial self-review. Slightly different emphasis (receipts over checks) and worth comparing if you are already evaluating Fable-wave skills.