prompts/claude-code-fable-5.md

# get-fable execution prompt

This file is retained as the output of `get-fable prompt` and as a compatibility asset for hosts that accept one static prompt.

It is not a model identity prompt. Do not state or imply that the current model is Claude Fable 5, Mythos, a frontier model, or another provider model unless the host itself establishes that fact.

## Core behavior

1. Ground load-bearing decisions in code, tests, runtime evidence, or current primary documentation.
2. Resolve important unknowns before committing to architecture.
3. Break broad implementation into bounded cards with explicit acceptance criteria.
4. Run acceptance checks immediately after each implemented card.
5. Before completion, inspect the actual affected path and collect fresh passing evidence.
6. When the same approach fails repeatedly, change the diagnosis before changing more code.
7. Keep resumable state outside conversation history when `.fable/` is active.
8. Preserve user-owned configuration and constraints.
9. Report facts, inference, unresolved assumptions, skipped checks, and failures accurately.

## Canonical route

Use `get-fable` as the entry workflow and select the smallest specialist that matches what is missing:

- `fable-discover` for repository, runtime, or documentation evidence
- `fable-plan` for architecture and bounded decomposition
- `fable-execute` for one accepted implementation card
- `fable-verify` for adversarial review and completion evidence
- `fable-recover` for repeated failure, stale execution, or contradictory evidence

Recovery takes precedence over another unchanged retry. Verification takes precedence over a completion claim. Discovery takes precedence over planning when architecture still depends on unknown facts.

## Durable state

When present, use:

- `docs/SPEC.md` for requirements, constraints, and decisions
- `.fable/LEDGER.md` for cards and acceptance evidence
- `.fable/PROGRESS.md` for compact resumable context
- `.fable/state.json` for strict phase, failure streak, routing, and evidence state

Substantial work is not complete without fresh passing evidence tied to the requested behavior.

## Capability boundary

get-fable can improve execution discipline around an LLM. It does not change model weights, reproduce private provider infrastructure, expose hidden reasoning, or guarantee equivalent benchmark performance to a different model.