🧠 Claude Fable 5 Prompt Optimizer

🚀 Quick Start

Option A: Claude Project (recommended)

• Open claude.ai and create a new Project (e.g., "Fable 5 Prompt Optimizer")
• Paste the contents of project-instructions.md (everything below the divider) into the Project Instructions field
• Upload all six files from knowledge/ as knowledge files in the project

Option B: Claude API

Concatenate the instructions and knowledge files into one system prompt:

import anthropic
from pathlib import Path
client = anthropic.Anthropic()
system_prompt = "\n\n".join(     p.read_text(encoding="utf-8")     for p in [Path("project-instructions.md"), sorted(Path("knowledge").glob(".md"))] )
response = client.messages.create(     model="claude-fable-5",     max_tokens=64000,     system=system_prompt,     output_config={"effort": "high"},  # the default — and the primary dial on Fable 5     messages=[         {"role": "user", "content": "prompt: Your raw prompt here"}     ], )
print(response.content[0].text)

Fable 5 API notes:

• Do not set temperature, topp, or topk — non-default values return a 400 error
• No assistant prefill (returns a 400 error) — use system prompt instructions instead
• Do not configure thinking at all — adaptive thinking is the only mode and always on; thinking: {"type": "disabled"} and budget_tokens produce errors
• effort defaults to high — keep it there even for workloads that ran on Opus 4.8 with xhigh; step down to low/medium for routine tasks, up to xhigh/max only for the most capability-sensitive work
• Set max_tokens generously — it caps thinking plus answer text; 64k is the documented starting value at high effort
• Handle refusals: stop_reason: "refusal" arrives as HTTP 200 — branch on it and configure fallbacks to claude-opus-4-8 (beta header server-side-fallback-2026-06-01)

🎯 The prompt: Trigger

| You type… | You get… | |---|---| | prompt: <anything> | The full optimization — everything after the prefix is treated as the raw prompt, verbatim | | What does R8 say? | Meta mode — answers about Fable 5, the rules, and their evidence status, grounded in the knowledge files | | Why no temperature? | Meta mode with the documented reason (non-default sampling → HTTP 400) and the prompt-based alternative | | make it shorter / add a German version | Follow-up mode — iterates on the last optimization and re-emits the full output | | give me the API parameters after all | Switches the last result to the API variant with a full <parameters> block |

Output anatomy: <optimizedprompt> (system/user split for API targets; chat targets fold system content into the user prompt; {{DOCUMENTS}}/{{INPUT}} placeholders for bulk content) · <parameters> (API only) · <changes> (each entry cites rule numbers) · <openquestions> (missing info and risk notes — never blocking counter-questions).

Non-English prompts are optimized in their own language; only the commentary follows your conversation language.

✨ Features

• 🎛️ Adaptive target detection — API prompts get a full <parameters> block; claude.ai chat prompts stay clean (ambiguous → chat default, API offered)
• 📏 43 source-backed rules — every rule tagged [DOCUMENTED: source] or [DERIVED], nothing invented
• 🧾 Auditable change log — every edit cites a rule, every rule cites an official Anthropic doc
• 🛡️ Refusal-risk warnings — cyber/bio/reasoning-extraction domains flagged, with fallbacks: claude-opus-4-8 advice
• 💬 Meta mode — ask the optimizer about Fable 5 and its rules, no trigger needed
• 🌍 Language preservation — prompts are optimized in their own language, commentary in yours

🆕 What's New in Fable 5

Fable 5's API and behavior retire several 4.x-era prompting habits; some now produce hard errors or safety refusals instead of merely suboptimal output:

| Your 4.x habit | On Fable 5 | The optimizer… | |---|---|---| | "Explain your reasoning step by step" | Can trigger a reasoning_extraction safety refusal and reroute you to Opus 4.8 [PromptF5], [Refusals] | removes it, offers thinking.display: "summarized" instead | | temperature: 0.3 for determinism | HTTP 400 — non-default temperature/topp/topk are rejected [Migration] | deletes it; tightens the prompt instead | | Assistant prefill for format control | HTTP 400 ("Use system prompt instructions instead") [Migration] | moves format control into the system prompt | | thinking: disabled / budget_tokens | Error — adaptive thinking is the only mode, always on [Intro], [Thinking] | strips it; depth is steered via effort only | | "CRITICAL: You MUST be thorough!" | Documented overtriggering cause; prescriptive 4.x prompts "can degrade output quality" [BestPract], [PromptF5] | neutralizes trigger language, deflates micro-plans | | effort: xhigh as coding default | Outdated — high is the recommended start, even for former xhigh workloads [Effort], [Migration] | recommends high and says when xhigh is worth it | | Showing the model its token budget | Inverted: countdowns trigger premature wrap-up [PromptF5] | removes countdowns, adds context reassurance |

A generic prompt optimizer tuned for 4.x models would actively damage prompts on every one of these points. This one is built from the Fable 5 documentation up.

🔄 How It Works

Six phases (0–5), run on every prompt: message:

flowchart LR
    IN(["prompt: …"]) --> P0["0 · Target detection<br/>API or claude.ai chat?"]
    P0 --> P1["1 · Analysis<br/>anti-pattern scan"]
    P1 --> P2["2 · Routing<br/>S · M · L · XL"]
    P2 --> P3["3 · Transformation<br/>structure → cleanup → snippets"]
    P3 --> P4{"4 · Quality gate<br/>12-point checklist"}
    P4 -- fail --> P3
    P4 -- pass --> P5(["5 · Structured output"])

• Target detection — API signals (tools, SDKs, harnesses, parameters) get a full <parameters> block; everything else defaults to the chat variant, with the API block offered on request.
• Analysis — scans for reasoning echoes, banned parameters, trigger language, over-prescription, token countdowns, cyber/bio refusal risks, and missing intent.
• Complexity routing — four routes decide which rule and parameter package applies:

• Transformation — structure (XML, document placement) → explicitness (intent, boundaries) → cleanup (strip 4.x patterns) → behavior snippets per route → parameters (API only).
• Quality gate — a 12-point checklist (no reasoning echo, no banned parameters, no trigger language, structure correct, verbosity steered, language preserved, …) runs before anything is emitted; failures loop back.
• Structured output — API variant (four blocks) or chat variant (three blocks).

📏 The 43 Optimization Rules

43 rules in six groups — every one tagged [DOCUMENTED: source] (stated in official Anthropic sources) or [DERIVED] (hypothesis inferred from them). The change log cites rule numbers, so every edit is auditable back to a primary source.

| Group | Rules | Topic | |---|---|---| | A | R1–R7 | Structure & explicitness (XML, document placement, intent, role) | | B | R8–R11 | Thinking & reasoning (the reasoning_extraction hazard, effort-only control) | | C | R12–R15 | Verbosity & communication | | D | R16–R26 | Agentics, long-run, autonomy (boundaries, checkpoints, memory, subagents) | | E | R27–R34 | API parameter recommendations | | F | R35–R43 | 4.x assumptions that no longer hold |

What gets transformed in practice:

• Structure: XML tagging, long documents (20k+ tokens) moved to the top with the query at the end (documented up to +30% answer quality), multi-document scaffolding, quote grounding.
• Explicitness: intent/audience block ("I'm working on X for Y…"), success criteria, explicit boundaries against scope creep.
• De-prescription: removal of behavior enumerations, step-by-step micro-plans, anti-laziness scaffolding, "CRITICAL/MUST" language, "think step by step" relics.
• Safety-aware rewrites: reasoning-echo instructions removed and replaced with thinking.display: "summarized" guidance; cyber/bio refusal risks flagged with fallback advice (fallbacks: claude-opus-4-8).
• Behavior snippets: brevity, communication style, anti-overplanning, anti-overengineering, autonomy reminders, progress audits, memory instructions, agent-to-agent guardrails — inserted selectively by route.
• Parameters (API targets): effort, max_tokens, streaming, prompt-caching breakpoints, fallbacks, task budgets — each with a one-line justification.

📂 Repository Structure

├── README.md                          # this file
├── LICENSE                            # MIT
├── project-instructions.md            # paste into the Project's custom instructions
└── knowledge/                         # upload these six files to the Project
    ├── 01-fable5-model-facts.md       # model facts, source legend, open points
    ├── 02-prompting-rules.md          # rules R1–R43 with evidence markers
    ├── 03-workflow-and-routing.md     # phases, routing, target detection, edge cases
    ├── 04-output-templates.md         # output blocks, API/chat variants
    ├── 05-snippet-library.md          # paste-ready behavior snippets
    └── 06-worked-examples.md          # three end-to-end examples

📚 Sources

All claims trace to these official Anthropic sources (snapshot: June 9, 2026, Fable 5 launch day):

| Code | Source | |---|---| | [News] | https://www.anthropic.com/news/claude-fable-5-mythos-5 | | [Product] | https://www.anthropic.com/claude/fable | | [SystemCard] | https://www-cdn.anthropic.com/d00db56fa754a1b115b6dd7cb2e3c342ee809620.pdf | | [Intro] | https://platform.claude.com/docs/en/about-claude/models/introducing-claude-fable-5-and-claude-mythos-5 | | [PromptF5] | https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5 | | [Models] | https://platform.claude.com/docs/en/about-claude/models/overview | | [Pricing] | https://platform.claude.com/docs/en/about-claude/pricing | | [Migration] | https://platform.claude.com/docs/en/about-claude/models/migration-guide | | [Thinking] | https://platform.claude.com/docs/en/build-with-claude/adaptive-thinking | | [Effort] | https://platform.claude.com/docs/en/build-with-claude/effort | | [BestPract] | https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices | | [Refusals] | https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback | | [StructOut] | https://platform.claude.com/docs/en/build-with-claude/structured-outputs |

Core facts (model ID, context window, output limit, pricing, thinking restrictions, effort guidance) were verified by triangulation across at least four independent official sources. The only measured discrepancy: tokenizer overhead "~30%" ([Models]) vs. "up to 35%" ([Pricing]) — compatible figures (approximation vs. upper bound).

⚠️ Limitations & Open Points

The official sources do not answer everything. The optimizer says "not documented" rather than guessing on:

structured outputs / strict tool use · knowledge cutoff · vision API limits (resolution/token formula) · rate limits and tier assignment · few-shot specifics for Fable 5 · tool search / parallel tool calls · fast mode and 300k batch output (absent from the lists, no explicit negative statement) · Mythos 5 prompting differences. Full list with details: knowledge/01-fable5-model-facts.md.

The knowledge is frozen at launch day (June 9, 2026). Anthropic updates its docs; re-verify parameter recommendations against the migration guide before relying on them in production.

🤝 Contributing

Contributions are welcome — new examples, rule refinements, translations, bug reports, and pull requests.

• Open an issue
• Submit a pull request

📄 License

This project is licensed under the MIT License.

⚠️ Disclaimer

Not affiliated with Anthropic. "Claude" is a trademark of Anthropic PBC. Rule texts summarize official documentation; entries marked [DERIVED] are hypotheses inferred from documented properties, not facts. Model behavior, pricing, and API contracts are Anthropic's — verify against the primary sources above before making decisions that depend on them.

⭐ Star History

If this optimizer saves you time or improves your Fable 5 results, consider starring the repo.

⭐ Star this repo · 🐛 Report an issue · 💡 Request a feature

Made with 🖤 by cheswick.dev · Not affiliated with Anthropic