Anthropic cut 80% of Claude Code's system prompt and the evals did not move

Scaffolding depreciates because of what it was built for. Anthropic says the hard rules in Claude Code's prompt were there to keep earlier models away from worst cases such as deleting files, and that the cost was knowingly accepted: without those guardrails, older models wrote comments that were incorrect in many cases ^[11]. That is a hedge against one generation's failure modes. Hedges expire when the thing they were hedging stops existing.

The running cost is the part most teams have never priced. Anthropic's own transcripts showed a system prompt, skills and a user request colliding inside a single call, one instruction asking for documentation as appropriate while another said not to add comments ^[4]. The model can usually work out intent, but it has to reason harder about the overlap before it acts ^[5]. Guidance is not free storage. It is charged on every request, and past some point each additional rule buys contradiction rather than control.

Read the headline result strictly. Removing more than four fifths of the prompt with no measurable loss on the coding evaluations ^[1] leaves the remaining prompt at under a fifth of its former size ^[13] and puts the measured value of everything deleted inside the noise floor of those evaluations at the moment it was cut ^[14]. That is not a claim that the text never helped. It is a claim that by the Opus 5 and Fable 5 generation, it had stopped paying, and the only way anyone found out was by running the comparison.

Which is where the asymmetry sits. The evidence is Anthropic's, the model is Anthropic's, and the evaluations are coding evaluations, so nothing here speaks to a claims-handling agent or a retrieval bot ^[15]. A vendor can delete four fifths of a prompt because it has a suite that tells it when quality moved. Most buyers maintaining CLAUDE.md files, skill libraries and prompt fragments cannot run that experiment, so they will either keep paying tokens and latency for guidance written against a retired model, or delete on faith.

The reversal on examples is the one that should sting. Examples were the number one rule for tool use, and Anthropic now says they pen the model into a narrower exploration space, recommending expressive interface design instead, with a todo status enumerated as pending, in_progress and completed doing the work that a worked example used to do ^[8][16]. Few-shot blocks are the artefact teams are proudest of and the most laborious to maintain. On this guidance they are a liability with a version number attached.

The other lesson is about where context should live. Claude Code once leaned on CLAUDE.md as its memory, and now has memory, artifacts and skills to load and share context across sessions ^[10], alongside progressive disclosure replacing the habit of putting everything up front ^[12]. Anthropic has packaged the pruning advice as a command, /doctor, to rightsize skills and CLAUDE.md files ^[9]. A vendor shipping a tool whose purpose is to shrink what its customers wrote is a fair signal about which side of the ledger that writing now sits on.

The durable asset was never the prompt text. It is the ability to measure what the text is worth this month.