That 'never retry' line in your CLAUDE.md is burning money for the new model

Anthropic has published a prompting guide for Fable 5.2, and one set of test data in it is worth pausing over for anyone who maintains a CLAUDE.md.

The guide mentions that Fable 5.2's context contained about a dozen notes left over from earlier sessions, only one of which said 'never retry.' That single line reduced its retries after calling a busy tool from almost every time to roughly four out of ten. The old rule didn't disappear; it just found a different way to charge you.

Files such as CLAUDE.md and AGENTS.md are instruction lists in a project that are meant for AI models to read. Developers write things like 'do not retry,' 'ask before deleting,' and 'keep answers under 20 characters' into them, expecting the model to comply. Most of these rules share a common origin: earlier models weren't smart enough, so people had to mark every pitfall one by one.

The problem is that the model has been upgraded, but the rules haven't.

Fable 5.2 follows instructions more closely. When you write 'never,' it really treats that as an absolute rule. As a result, defenses set up for older models have now become stumbling blocks for the new model. The guide lists a set of things to reconsider: standing instructions that tell the model to delegate work to subagents, hard-coded style restrictions, approval rules that don't specify concrete actions, and old notes piled up in memory. Each one consumes tokens and slows down responses.

The guide's wording is restrained: some interventions designed for earlier models increase costs, some have side effects, and some simply no longer work.

It also provides a prompt you can send directly to Fable 5.2, asking it to read the official guide and then review the project's CLAUDE.md, AGENTS.md, skills, subagent files, and memory files. The things it should look for include: instructions that make you delegate work, sentences that make you think less and answer directly, hard-coded style restrictions, 'never' rules, accumulated old notes, and approval rules that don't clearly specify concrete actions. For each issue, it should quote the original line, explain the impact, and suggest the smallest possible change. Proposals involving security boundaries and reduced verification should be listed separately and left unmodified for now.

A few other behavioral changes in the guide are also worth knowing. At medium effort and above, phrases like 'answer directly, don't think' no longer keep the model in check; the right move is to lower the effort level. Telling the model 'the code is your own' raises its rate of consolidating duplicate code in small tasks from under 10% to about 75%, at the cost of 70% to 80% more output tokens, and it will also make changes nobody asked for. These numbers aren't meant to be accepted wholesale. What they show is that the old rules and the new behavior no longer line up.

Several replies added practical ideas.

@Blum said that one line of rules doing so much makes him wonder whether other notes are also burning tokens for nothing.

@Monomade's suggestion was more concrete: before deleting anything, date each 'never' rule and note what incident originally made you add it. Rules with no story behind them should be deleted first; for those tied to real failures, after changing them, rerun the same task to see whether Fable 5.2 still falls into the trap.

@Snaki pointed out that the cost is twofold: bad retries, plus the input tokens you pay for on every turn.

@Krandor's analogy had a touch of dark humor: having Fable review my old rules feels like a new employee auditing the boss.

One Chinese reply also offered a useful way to categorize things: keep security boundaries, relax style restrictions, review every hard 'never' rule one by one, and after deleting them, run a round of real tasks to compare token consumption.

Original guide: platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-2

Original tweet link: x.com/Voxyz_ai/status/2120501801313267818

Cleaning up CLAUDE.md is fundamentally a bookkeeping problem. Every rule should have a source, an effective date, and conditions under which it expires. The model has changed generations, so the ledger should be reviewed too.