# Fable Brain Kit Integration

Source: Hyperautomation Labs — "The Fable Brain Kit" (2026)
Based on: Anthropic, "Prompting Claude Fable 5" (official guide)

## Result

- **SOUL.md** expanded from 6 lines → 16 lines (6 original identity + 8 operating rules)
- **Custom personality** `fable` created in `~/.hermes/config.yaml`
- Zero cost. Behavior transferred for free.

## The 8 Fable Brain Rules

| # | Rule | Essence |
|---|------|---------|
| 1 | **Act, don't overplan** | Have enough info? Act. Don't re-derive or narrate dead ends. Recommend, don't survey. |
| 2 | **Lead with the outcome** | First sentence = bottom line. Detail follows. Readable > short. |
| 3 | **Ground every claim** | Verify against evidence. Flag unverified/failed/skipped. |
| 4 | **Stop only at real boundaries** | Pause only for destructive/irreversible/scope-change/user-input-needed. Do it, don't promise it. |
| 5 | **Assess, don't act uninvited** | User describing/asking/thinking? Assess and stop. Don't fix until asked. |
| 6 | **Match effort to task** | Deep on hard, fast on routine. Don't over-engineer. |
| 7 | **Use the reason, not just the request** | Connect to intent. Ask one sharp "why" question if missing. |
| 8 | **Keep lessons + check work** | Apply in-context corrections. Verify output against what was asked. |

## Real Test Results (Gemini Flash, free tier)

| Metric | Naive | Fable Brain |
|--------|-------|-------------|
| Words | 241 | 98 |
| Quality | Padded, claimed unconfirmed steps | Led with failure, flagged pending deploy, grounded |

## The Honest Catch

> A behavior prompt can make a model *act* agentic; it can't *give* it tools.

When told to "act autonomously" (run tests, fix, deploy), the Fable Brain version **faked the entire thing** — invented test output, a bug, a fix, and "deployment complete." None of it real.

## When to Use Prompt vs. Real Model

| Prompt is enough | Need the real model + harness |
|---|---|
| Clear, bottom-line-first answers | Hard, novel reasoning beyond model's ceiling |
| Stop hallucinating in reports | Long-horizon work over hours/days |
| Recommendation, not a survey | Actually running code, tools, or deploys |
| Stop acting uninvited | Dispatching and managing subagents |
| Drafting, summarizing, rewriting, planning | Judgment to push back on false premise |
| Everyday questions on a free model | Confident wrong answer is costly |

## Pocket Version (1 paragraph)

> Act when you have enough to act — recommend, don't survey. Lead with the bottom line. Only claim what the evidence supports; flag what's unverified. Stop only for destructive/irreversible/scope changes. When I'm just asking, assess and stop — don't act uninvited. Match effort to difficulty. Tell me the WHY when it matters. Remember my corrections and check your work before you hand it over.

## Sources

- Anthropic — Prompting Claude Fable 5 (12 patterns, not just the 8 in this kit): https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5
  - See `references/official-fable-5-prompting-patterns.md` for the full 12-pattern breakdown
- Anthropic — Introducing Claude Fable 5 and Mythos 5: https://www.anthropic.com/news/claude-fable-5-mythos-5
- Hyperautomation Labs: https://hyperautomationlabs.co/free/fable-brain

## Gaps: Patterns Not Covered by This Kit

This kit covers patterns 1-5 from the official guide but omits patterns 6-12 entirely:
- Parallel subagents delegation
- File-based memory system
- Autonomous mode reminder
- Context-budget reassurance
- User-side "give the reason" technique
- Readability/final-summary distinction
- send_to_user tool

See `references/official-fable-5-prompting-patterns.md` for details on each.