A Claude Skill that brings disciplined, five-gate reasoning (scope, evidence, adversarial review, verify, report) plus effort and model-routing judgment to any Claude model.
A skill file can't transfer a model's raw intelligence, but it can transfer its discipline: how it scopes a task, gathers evidence, attacks its own answers, verifies, reports, and decides how much effort and which model a given step actually deserves. Loading this skill makes that discipline available to Opus, Sonnet, or a cheaper worker model, regardless of which one is executing the task.
SKILL.md— the skill file itself. Drop this into your skills directory.examples/example-session.md— a worked example showing the five gates catching a real mistake mid-task.Changelog.md— version history.
Place SKILL.md in your skills directory, keeping the folder name fable-mode:
your-project/
└── .claude/
└── skills/
└── fable-mode/
└── SKILL.md
If you're using this outside Claude Code (for example, pasted into a system prompt or another orchestration setup), just include the contents of SKILL.md as a standing instruction block.
Use it proactively the moment a task has:
- Multiple dependent steps
- Unknowns that could change the approach mid-task
- Debugging where the first theory might be wrong
- Anything that needs verification before handoff
- A task that keeps failing or stalling
Or invoke it directly with phrases like "fable mode," "think like Fable," "use the Fable method," or "slow down and do this right."
- Gate 0 — Calibrate. Decide effort level and which steps deserve a strong model versus a cheaper one, before starting.
- Gate 1 — Scope. Define what done looks like and how you'll check it, before touching anything.
- Gate 2 — Evidence. Open the real file, API, or dataset. Never design from memory of what it "probably" looks like.
- Gate 3 — Adversarial reasoning. Attack your own answer as a hostile reviewer before committing to it.
- Gate 4 — Verify. Confirm at the layer of the actual claim, not one layer below it.
- Gate 5 — Report. Separate what was verified from what was assumed, out loud.
Full detail, including standing habits and failure-mode smells, is in SKILL.md.
Models come and go. Access changes, pricing changes, a specific model can be deprecated or temporarily restricted. What doesn't have to disappear is the process a strong model uses to work carefully: scoping before building, checking evidence instead of guessing, attacking its own conclusions, verifying before declaring done. This skill is an attempt to write that process down so it survives any single model's availability.
The five-gate structure here is built from ideas in Nate Herk's video How I Make Opus Think Like Fable (5 easy steps), which walks through extracting Fable 5's working discipline from its leaked system prompt and applying it to cheaper models like Opus and Sonnet. This repo is an independent rewrite built around those same principles, not a copy of his original file. If you found this useful, his channel is worth a follow: @nateherk.
MIT. See LICENSE.