Claude Fable 5 is the flagship of Anthropic's Claude 5 family, and Claude Opus 5 sits one step below it as the more accessible option. Both are text models with a context window of roughly 1,000,000 tokens, and both can reason, analyze images, call tools and functions, search the web, generate images, and handle code very well. Both are available in GPTunneL — open a chat, pick the model, and try the examples from this guide as you read.
When to pick Fable 5 vs Opus 5
- Claude Fable 5 is for the hardest problems: multi-step reasoning, deep analysis of large documents, architectural decisions in code, long autonomous chains with tools. Reach for it when mistakes are expensive and the task is hard even to specify.
- Claude Opus 5 is the balance of quality and accessibility: everyday coding, writing, document review, structured data extraction. For most working tasks it is more than enough.
A practical rule: start with Opus 5, and if the answer falls short on a genuinely hard task, rerun the exact same prompt on Fable 5.
What to keep in mind when prompting
- XML tags are Claude's signature technique. The family is trained to understand markup like
<instructions>,<data>,<examples>, and<answer>: it separates instructions from data and removes ambiguity. See Anthropic's guide to XML tags. - Long context. About 1,000,000 tokens fit entire books and repositories. Put large materials at the top of the prompt and your question and instructions at the end, and ask the model to reference specific sections of the source — answers get noticeably more precise.
- Code. Both models are strong at code, but results improve when you specify the language, version, style, and constraints ("no new dependencies", "keep the public API intact"). Ask for an explanation of changes — it makes review easier.
- Clarity beats length. State the goal, output format, and acceptance criteria explicitly — the core principle from Anthropic's prompt engineering overview.
Techniques with examples
The structural template
A universal starting point for any non-trivial task:
<instructions>
Role: technical editor. Goal: cut the article to 800 words.
Constraints: keep all numbers and quotes. Format: Markdown.
</instructions>
<data>{article text}</data>
<answer>Return only the final text, no commentary.</answer>
Few-shot: examples instead of descriptions
When style or output schema matters, show 1–3 input → output pairs — it works better than long explanations:
Classify support tickets. Categories: billing, models, bug.
<examples>
"My payment won't go through" → billing
"Which model is best for code?" → models
</examples>
Ticket: "The submit button doesn't respond" →
Guided reasoning with <thinking>
For tricky tasks, ask the model to reason first and answer second:
First, work through the problem step by step inside <thinking>…</thinking>:
list the known facts and check edge cases.
Then give the final answer inside <answer>…</answer>.
Prefilling the answer
Dictate the first characters of the response to lock the format and cut preambles:
Return the project risks strictly as a JSON array.
Begin your answer with exactly these characters: {"risks": [
Write nothing before or after the JSON.
Self-critique chains
Split the work into "draft → critique → final", in one message or several:
Step 1. Write a draft of the function from the spec below.
Step 2. Inside <critique>, find at least 3 problems in your draft:
bugs, edge cases, readability.
Step 3. Inside <answer>, provide the corrected final version.
Ready-made prompts — copy and adapt
- Code review: "Review the code in
<data>. For each issue give the line, severity (critical/major/minor), and a fix. Finish with a summary table. Do not rewrite the whole file." - Document digest: "
<data>contains a long document. Produce a digest: the 5 key points, each with a reference to its section in the source. Then a section called 'Risks and open questions'. Format: Markdown." - Screenshot analysis: "I'm attaching a UI screenshot. Describe what is happening, spot UX problems, and suggest 3 concrete improvements — from the cheapest to the most expensive to implement."
- Refactoring: "Refactor the code in
<data>: remove duplication, improve naming, preserve behavior and the public API. In<answer>, give the code, then a list of changes with a rationale for each." - Verified web research: "Search the web for current information on topic X. Cite a source for every claim. Separate facts from opinions. If sources disagree, show both versions."
Diagnosing common failures
| Symptom | Cause | Fix |
|---|---|---|
| Vague, off-target answer | Goal and format not stated explicitly | Add <instructions> with a goal, format, and criteria |
| Instructions mixed up with data | Everything in one paragraph | Split into <instructions> and <data> |
| Filler openings like "Sure, here is…" | Output format not pinned down | Prefill the answer or say "return only X, no preamble" |
| A fact lost in a long document | Question placed before the data | Data first, question last, ask for section citations |
| Output style drifts between runs | Style described in words | Replace the description with 2–3 few-shot examples |
| Rushed answer to a hard problem | The model didn't reason | Add a <thinking> step or switch to Fable 5 |
You can verify every example in GPTunneL within minutes: run the same prompt on Claude Opus 5 and Claude Fable 5 and see where the difference actually matters for your tasks.