Fable Skill — executing-hard-tasks

A pressure-tested Agent Skill for Claude Code that changes how a frontier model decomposes hard tasks, verifies its own work, and decides what to do next.

Most "think better" prompts restate what a capable model already does. This skill was built the opposite way: agents were first run through pressure scenarios without any guidance (deadline pressure, sunk cost, tired-user pressure, no-runtime environments) to observe where they actually fail — then the skill was written to counter only those observed deltas, re-tested with the skill loaded until every baseline failure stopped recurring, and hardened by an adversarial critic hunting for loopholes a pressured agent could negotiate with.

What it enforces

Phase	Load-bearing rule
Decompose	"Done" stated as a checkable observation; thinnest end-to-end slice through the riskiest integration path first (never backend-block → frontend-block → test-block); a pasteable checkpoint per subtask, fixed at plan time; spike before building on an architecture; explicit DECISION: lines for choices users can observe
Verify	Three buckets — Verified (observation pasted inline, produced after the last edit), Believed, not verified, Not checked. "Done" is only declarable when the done-observation itself sits in Verified — otherwise the report opens with NOT DONE
Decide next	Red checkpoint → debug now; two failed fixes on one checkpoint → the diagnosis is wrong, not the patch; surprise (output ≠ planned expectation) → re-derive the plan; goal met and verified → stop, list follow-ups without executing them

How the loop works

Install

mkdir -p ~/.claude/skills/executing-hard-tasks
curl -o ~/.claude/skills/executing-hard-tasks/SKILL.md \
  https://raw.githubusercontent.com/wsbm393/Fable-Skill/main/SKILL.md

Or clone and copy:

git clone https://github.com/wsbm393/Fable-Skill.git
cp Fable-Skill/SKILL.md ~/.claude/skills/executing-hard-tasks/SKILL.md

Claude Code discovers it automatically; it triggers on multi-step tasks, on any completion claim, and on next-action decisions after failures or surprises.

How it was built (skill-TDD)

1. RED — six pressure scenarios ran with no skill loaded; judges extracted failures and the agents' verbatim rationalizations ("fine for a demo", "the dependency already exists, it costs nothing").
2. GREEN — the skill was drafted against only the observed failures, then the same six scenarios ran with it loaded: all six passed, no baseline failure recurred.
3. REFACTOR — three critics attacked the text: a loophole hunter (role-playing an agent that wants to cut corners), a format reviewer, and a dead-weight critic that removed every rule merely re-teaching what the baseline already did well.

Written for and tested against Claude Opus-class models; the rules are model-agnostic.

Tested against nine scenarios — and deliberately kept small