Optimizing Claude Fable 5.1 Workflows

The Shift to High-Effort Reasoning

Theo argues that for most complex development tasks, users should default to 'High' or 'X-High' reasoning levels in Claude Fable 5.1. While lower tiers save tokens, they frequently fail on complex reasoning, leading to higher overall costs when the task must be re-run. Unlike other models that struggle to scale reasoning usage, Fable 5.1 is efficient enough that 'High' does not result in excessive token waste for simple tasks, making it a reliable default for deep work.

Refined Prompting and Behavioral Control

Recent updates to Fable 5.1 have reduced the model's tendency to provide verbose, user-facing progress updates. While this keeps the interface clean, users who prefer transparency can re-enable it via system prompts. Theo emphasizes that many common 'anti-pattern' instructions—such as forbidding bold text or bullet points—are now largely unnecessary and should be pruned from agent.md or claude.md files to allow the model to function more naturally. He also highlights the effectiveness of explicitly prompting against 'mannered prose' to eliminate the 'Claude slop' characterized by flowery, jargon-heavy language.

Delegating Complex Decision-Making

Instead of acting as a manual supervisor, the developer's role is shifting toward providing the model with the context required to make autonomous decisions. By feeding the model raw data (e.g., PR comments, error logs, or screenshots) and setting clear constraints (e.g., 'don't scope creep', 'only address real shortcomings'), the model can effectively manage multi-step workflows like PR maintenance or architectural refactoring. Theo demonstrates this by having the model evaluate competing PRs, merge changes, and perform risk assessments on high-stakes code changes.

Key Takeaways

• Default to High: Use 'High' reasoning for most tasks; the cost difference for simple tasks is negligible compared to the risk of failure at lower settings.
• Prune System Prompts: Delete legacy formatting constraints in your agent.md files; modern models handle formatting better without rigid, outdated rules.
• Ask for Updates: If you miss the model's 'thinking' process, explicitly prompt it to provide brief updates before and after tool calls.
• Context is King: Rather than choosing between options yourself, provide the model with all available data (PR history, logs, screenshots) and ask it to evaluate and execute the best path.
• Beware of Distillation Guards: Be aware that editing message history in a thread will strip the model of its reasoning traces to prevent distillation attacks.