Skip to content

tstanmay13/think-like-fable-5

v0.1.1MIT

Fable 5's working discipline, distilled — outcome-first reporting, act-don't-ask autonomy, faithful verification. Load at session start on a smaller model.

GitHubOpen Website

think-like-fable-5

A Claude Code and Codex agent skill written with Matt Pocock's skill-authoring method and shipped with paired evals that measure whether it actually changes model behavior.

demo: the skill, then a live paired eval on haiku

Fable 5 will not be around forever. What makes a session with a frontier model feel different is mostly not knowledge. It is discipline: leading with the outcome, acting instead of asking permission, reporting failures plainly, refusing to end a turn on a promise. Those are behaviors, and behaviors can be written down.

This skill distills that discipline into explicit, checkable rules. Load it at the start of a session on a smaller model (Haiku, Sonnet) and hold that model to the same standard.

Usage: type /think-like-fable-5 at the start of a session. The skill is user-invoked, so it costs nothing in context until you reach for it.

Claude Code and Codex

The same skill files support both agents. Install from the combined catalog:

claude plugin marketplace add tstanmay13/claude-skills
claude plugin install think-like-fable-5@tstanmay13-skills
codex plugin marketplace add tstanmay13/claude-skills
codex plugin add think-like-fable-5@tstanmay13-skills

Invoke /think-like-fable-5 in Claude Code or $think-like-fable-5 in Codex. Restart the agent after installation or updates. The repository name claude-skills is retained for existing installs; its catalog now serves both agents.

plugin.json is the portable package manifest; .claude-plugin/plugin.json and .codex-plugin/plugin.json provide runtime compatibility. All three share one version and one skills/ directory. Source changes belong here, never in an installed plugin cache. Directory publication is separate from GitHub distribution; see the catalog release guide.

Other installation options

npx skills add tstanmay13/think-like-fable-5

Or clone and symlink into ~/.claude/skills:

./scripts/link-skills.sh

Evals

Every rule in the skill is only worth its tokens if it changes what a smaller model does. The eval harness in evals/ measures that directly: each scenario is a prompt built around a trap that undisciplined models fall into, and it runs twice on the model under test, once bare and once with the skill appended to the system prompt. An LLM judge grades both responses against a binary rubric. The delta is the skill's lift.

node evals/run.mjs                          # all scenarios, haiku under test
node evals/run.mjs --scenario buried-lede   # a single scenario
node evals/run.mjs --model claude-sonnet-5  # a different model under test

Requires the claude CLI, installed and authenticated.

The harness is also how the skill was written. The first full run showed the skill making Haiku worse on one scenario: a fix request with an "open question" planted inside flipped it into menu-and-ask mode. That finding became a new rule in the skill (sub-decisions inside a task are yours to make), which took the scenario from 2/5 back to 4/5. Full-suite totals on Haiku before that fix: 32/42 bare, 35/42 with the skill. Expect single-run noise; judge verdicts land in evals/results/.

Demo

A recorded terminal session is in docs/demo.cast (asciinema play docs/demo.cast).