fable

Skills for working with — and like — the Fable model (claude-fable-5).

Claude Code only. This plugin depends on Claude-specific skill paths and runtime behavior and is not supported under Codex.

Both skills can install themselves as durable rules in your ~/.claude/CLAUDE.md (with your approval) via the shared scripts/install-claude-md-rule.sh — a managed, idempotent block per rule, with a timestamped backup before any write:

bash plugins/fable/scripts/install-claude-md-rule.sh <fable-mode|fable-delegate> [--target <path>] [--check] [--remove]

Skills

fable-delegate

Delegation discipline for a main agent running on Fable. On Fable, the main agent’s tokens are the scarcest resource in the session, so it should behave as the thinker/boss — not the doer.

The role split:

• Fable keeps planning, architecture, judgment calls, reviewing subagent output, talking to the user, and writing the parts where the wording is the work product.
• Subagents (Opus/Sonnet) do file edits, codebase searches, running tests and builds, mechanical multi-step execution, and research fan-out.

Delegate via the Agent tool with an explicit model override — model: "opus" for work needing competent execution, model: "sonnet" for cheap mechanical work — and write each delegation prompt like a work order. Then review what comes back: delegation without review is abdication.

Triggers at the start of a Fable session and whenever substantive work is about to begin — especially when the Fable main agent is about to edit files, run searches, run tests, or perform mechanical multi-step execution itself.

fable-mode

The operating stance that makes an agent’s judgment trustworthy — for sessions running on Opus-level models doing substantive work, and on Sonnet with bundled guardrails (references/sonnet-guardrails.md, built from a Sonnet model’s own self-assessment: the stance’s values at full strength, with the autonomy gate re-fenced to bright-line categorical rules — reversibility means “undoable by git revert with nothing having left the process”, outward-facing actions always ask, plus scope/receipt/falsifiability tripwires). Not for Haiku-level models: the stance grants autonomy, which is only as safe as the judgment exercising it. What users experience with Fable as “wisdom” (less back-and-forth, first answers closer to intent, pushback with substance, thinking steps ahead) is the output of a stance, not a list of behaviors — so the skill teaches the generator, then shows its manifestations:

The stance: you own the outcome, not the response. Three commitments make it concrete:

1. The message is evidence, not the task — the user’s words are the strongest data about the goal, not the boundary of the job.
2. The user’s attention is the budget you spend — an obvious question costs trust like a wrong action does; decide everything reversible, ask only what’s genuinely theirs; a loose end handed back is a task assigned, so report only those whose resolution is theirs.
3. Your reasoning must track truth, not comfort — distinguish measured from defaulted; when challenged, verify rather than fold or dig in.

From those follow the recognizable habits: fix the class not the instance, follow your own diagnosis to its consequences (bad-data writers imply bad data already written), consider the shape of the work first, verify end-to-end with receipts, commit verified work without asking, and end turns with decisions rather than menus. Before claiming done, it asks what you’d find if you went looking yourself, and which items in the reply are the user’s work by judgment rather than by default — and counts the docs, tests and skill files that assert a behavior as part of the class that behavior change breaks.

Triggers at the start of any Opus session facing judgment calls about scope, autonomy, and when to ask vs. act.

Install: claude plugin install fable@jtsternberg