Claude Fable 5.2 Prompting Tips: Setting Effort Levels and Completing Long Tasks
Practical prompting tips based on the official Claude Fable 5.2 prompting guide, covering effort parameter selection, preventing mid-task stops, specifying appr
This guide summarizes key configuration and prompting practices from the official Claude Fable 5.2 prompting guide, as compiled and shared by @minheo93698092. The recommendations focus on tuning the effort reasoning parameter to match task complexity, preventing long-running sessions from stalling mid-task, and defining explicit boundaries when approving model actions.

Image source: TechCrunch 쨌 CC BY 2.0 쨌 https://creativecommons.org/licenses/by/2.0 쨌 https://commons.wikimedia.org/wiki/File:Dario_Amodei_at_TechCrunch_Disrupt_2023_01.jpg
Summary of the Original Tips
Based on @minheo93698092's breakdown of the official Claude Fable 5.2 prompting guide, five core practices stand out:
- Explicitly set
efforttohighand state completion instructions first: Configure theeffortparameter tohighand place an explicit instruction at the start of your prompt telling the model to carry the task through to completion. - Test
high,medium,xhigh, andlowaccording to task needs: According to the official guide,highis the default setting, and users are encouraged to testmediumandxhighalongside it. Thelowsetting is built for fast tasks; in Anthropic's internal testing,lowon Fable 5.2 performed close to highereffortlevels on Fable 5.1. - Instruct the model to work until finished to prevent mid-task pauses: If the model pauses during a long task to ask
"Should I continue?"(or"계속할까요?"), explicitly instruct it to keep working until the task is completely finished. In the author's testing at themediumeffort level, adding this instruction cut incomplete sessions roughly in half. - Replace
"Go"approvals with specific permitted actions: When granting approval for the model to proceed, avoid vague single-word confirmations like"Go". Instead, spell out exactly what actions the model is permitted to perform. - Lower the
effortsetting instead of prompting the model not to think: Instructions such as"don't think"("생각 마") do not work wheneffortis set tomediumor higher. If the model spends too long reasoning, lower theeffortparameter directly rather than trying to suppress reasoning through prompt wording.
Why These Settings and Prompt Patterns Matter
Effective control over reasoning models relies on separating system-level compute parameters from natural-language task instructions:
- How the
effortparameter governs reasoning depth: Theeffortparameter controls how much internal reasoning (thinking token generation) the model allocates before producing its final output. Available tiers—low,medium,high, andxhigh—trade off response latency and token usage against deeper multi-step deliberation. Becausemediumand higher tiers enforce internal reasoning at the system level, telling the model"don't think"in the prompt text will not override its reasoning behavior. Adjusting theeffortparameter directly is the reliable way to shorten response times. - Differences across execution environments: As noted in community follow-up discussions on the post, default
effortsettings can vary depending on the interface—for example, API calls default tohigh, whereas agentic coding environments like Claude Code may default tomedium. Explicitly specifyingeffortensures predictable behavior regardless of where the model is invoked. - Clear finish lines and bounded approvals: Conversational models frequently pause during extended multi-step workflows to confirm whether they should continue. Stating upfront that the model should continue until the task is complete eliminates unnecessary interruptions. However, when authorizing autonomous execution across a longer workflow, replacing a generic
"Go"with a concrete list of allowed operations keeps the model within safe, intended boundaries.
Practical Selection Matrix by Workload
When applying the Fable 5.2 prompting guide to daily workflows, match the effort level and instruction structure to the nature of the task:
| Workload Type | Recommended effort | Prompting and Approval Approach |
|---|---|---|
| Quick queries and simple summaries | low | Prioritizes low latency while offering quality comparable to higher effort tiers on Fable 5.1 |
| Standard coding and iterative tasks | medium (compare with high) | Include explicit completion instructions to reduce incomplete sessions |
| Complex architecture and multi-step tasks | high (test xhigh as needed) | Put completion requirements at the top of the prompt and define exact permitted actions upon approval |
| Excessively long reasoning times | Step down one effort tier | Lower the effort parameter directly instead of adding "don't think" instructions |
Rather than relying on a single fixed prompt or leaving effort at its maximum tier for every request, benchmark the same representative task across medium, high, and xhigh to evaluate output quality against latency and token cost. Combining an explicit completion boundary with the right effort tier yields more reliable, end-to-end execution.
Original source
- @minheo93698092 (X): Claude Official Fable 5.2 Prompting Guide Summary Post