Anthropic's Fable 5.2 playbook: own the code, re-tune effort

The official migration guide is out: one line ~5x's code consolidation at 1.5x the tokens, effort now defaults to high, and stale Opus 5.5 lines backfire.

Nowline NOV 10 11:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • One sentence, ~5x the code dedup

    Adding “This codebase is yours…” to your prompt made Fable 5.2 merge duplicated code in about half of tries, versus under a tenth without it — the single biggest lever in the guide. The catch: ~1.5x the output tokens and changed behavior in existing features in roughly 1 in 6–10 runs, so review the diff, not the summary, and drop the line when you want minimal changes.

  • Effort now defaults to high — recalibrate

    Don't carry over effort settings from Fable 5.1 or Opus 5.5: start at high (most of the capability is already there), then step down to medium wherever your evals hold to cut the bill. At high and up it thinks before nearly every reply — even “hi” — and an “answer directly” instruction won't stop it at medium or above.

  • A drop-in block that halves early stops

    Anthropic's “finish the work” system-prompt snippet cut sessions that quit with work still owed by about half, at roughly 30% more output tokens. Pair it with precise approvals — a bare “continue” underperforms, and a vague gate like “messages to people outside this conversation” made the model hold back about half its routine messages to colleagues.

  • Stale cautions quietly make it timid

    Old notes pile up and suppress action: a single “never retry” line dropped retries from nearly always to about 4 in 10, and a reviewer's “stop at the first error” killed them in a third of tries. Rewrite absolute rules with scope and reason — a bare “never” got the model to act 1 in 10 times, while “get a go-ahead first” got it to 9 in 10.

  • Audit prompts written for older models

    In new conversations, delete Opus 5.5's “treat that answer as done” lines (they stop Fable 5.2 re-checking its mistakes) and strip hard token budgets baked into prompts (lower cost at high/xhigh for equal results). Also check which model runs your subagents — they can self-review at xhigh/max and quietly dominate a session's cost.