Claude Fable 5.2 prompting guide

Which effort level to use

Start at high, the default, and move only when your own tests say so. Anthropic reports that most of the capability gain is already present at high: the step from medium to high raised scores on most of its coding and knowledge-work evaluations, while gains above high were uneven and task-dependent. Re-run this choice for Fable 5.2 instead of copying the level you tuned on Fable 5.1 or Opus 5.5, because the amount of thinking behind each level differs between models.

Effort barely changes how long the visible reply is, so control length and tone in the prompt. You can mix levels in one job, for example planning at medium and raising effort for the long unattended part; note that setting a new top-level effort on a new request starts the prompt cache over. On apimodels.app, thinking tokens bill as output at $24 per million, so effort is also the main cost lever.

Make it finish the whole task

On long, multistep work Fable 5.2 sometimes ends a turn early: it lists what remains, offers to continue, or stops at an ordinary obstacle such as a timeout. The fix is a short “finishing work” section in the system prompt. It should ask the model not to stop while it still owes work, to carry on with anything that does not depend on the user’s answer, and to diagnose ordinary obstacles (errors, timeouts, locked files, empty results, failing tools) and work through them. It should also tell the model to leave deliberate blockers alone, such as a file marked not to be touched, withheld access or a safety rule, and to report them. In Anthropic’s replay of coding sessions at medium effort, a version of this text halved the sessions that ended with doable work left, and output tokens rose by about 30%.

Because it asks for permission less often with this text, keep your own confirmation step for anything risky or irreversible. When you do approve something, say exactly what: “yes, run the migration on the staging database” rather than “go”. Approval rules in a system prompt should name the exact actions they cover, since the model applies them as written; in one Anthropic test with nobody available to approve, a rule that required approval for messages to people outside the conversation made it hold back about half of its routine requests to colleagues.

Say who owns the code

Fable 5.2 reacts strongly to what the prompt says about ownership. When the prompt is silent, or presents the code base as someone else’s, it tends to add to the existing structure, even when that repeats code, and to report nearby problems instead of fixing them. Telling it that the code base is its own to design and keep building on changes that: in Anthropic’s small coding tasks it merged repeated code in about half of the tries, against under a tenth without the sentence.

The cost is scope. With the ownership sentence it used about 1.5 times the output tokens, made changes nobody asked for (often tests and documentation) and, apart from bug fixes, changed how an existing feature behaves in one try in six to ten. Its closing summary named the big changes and often skipped small ones, so review the diff. Leave the sentence out when the user owns the code and wants the smallest possible change.

Progress updates and final reports

On long runs of tool calls Fable 5.2 can go quiet. In your own application, the short updates it writes between tool calls arrive as thinking blocks whose text is empty by default, so a client that only renders text blocks looks silent; Anthropic’s guide covers the display setting that returns them. The prompt-only fix is one sentence asking for a one-line status every few tool calls. Where a person watches the work, also ask it to say what it is about to do before it starts and to end with a recap that stands on its own; in Anthropic’s test that raised output tokens by about 20%.

For final reports after long work, ask it to write for a reader who has not seen the tool calls or files: the key point first, plain words, no terms it invented during the work, and prose where ideas connect with bullets kept for lists.

What to delete from prompts written for older models

Fable 5.2 follows instructions closely, so lines that were harmless on Fable 5.1 or Opus 5.5 can raise cost or cause side effects. Review your system prompts, skills, agent files and memory notes before you switch. One more change matters for applications: Fable 5.2 checks that the start of a conversation has not changed when thinking blocks are sent back, so editing the system prompt, the tools or an earlier message in a stored conversation can return a 400 error. Put new instructions in the newest turn, or start a new conversation.

When it refuses

Fable 5.2 runs safety classifiers that can decline a request; the reply comes back with stop_reason refusal and a category: cyber, bio, frontier_llm (help building competing AI models), reasoning_extraction or general_harms. Legitimate security and life-science work can trip the cyber and bio classifiers. Organisations doing that work can apply to Anthropic’s verification programmes; for a single request, the practical route is another model. Anthropic’s own fallback is Opus 4.8, which is also on apimodels.app under the same key at $3 / $13 per million tokens.

Prompts for creative work in code

The Fable 5.2 work that spread after launch, from Blender films and Three.js scenes to browser games, code-painted pictures and music, came from prompts with the same few traits. The deliverable is concrete, with format, length and resolution. The quality bar is stated plainly, with permission to take the time. Everything is generated in code, with no downloaded models, textures or image models. And the model checks its own output: it renders small test frames, looks at them, fixes what is wrong, and only then renders the full piece. For web animations, a deterministic renderAt(t) function lets every frame be captured and encoded into a video.

Some of the posts behind these patterns, summarised: @techfren asked for a 3D pelican riding a bicycle, as realistic as possible, and Fable 5.2 drove Blender from the command line inside Claude Code; @RezAIT00 asked for a single three.js page rendering a voxel island of Paris in 1888 with renderAt(t) exposed; @felixrieseberg asked for a small-scope first-person shooter with graphics that look like a big-budget game; @aug5thmusic asked for a Bach-style four-part chorale written in LilyPond; Jake Eaton had it invent a painter and write a Python paint engine. The links are in the sources box below; the templates in Quick start are ours, distilled from these posts, and our use-cases article covers the full stories, time and cost.