Anthropic Cut 80% of Claude Code's System Prompt. We Audited Ours.

On July 24, 2026, Anthropic published new context-engineering guidance for Claude 5 models with a striking result: it had removed over 80% of Claude Code’s system prompt for Opus 5 and Fable 5 without a measurable drop on its coding evaluations.

The obvious response is to start deleting. We tried something slower: we opened the CLAUDE.md from this website’s own production repository and audited every section.

What we found was more useful than a smaller file. Prompt bloat does not only waste tokens. Stale instructions can make an agent confidently misunderstand the system it is working on.

First, what Anthropic actually changed

Anthropic’s argument is not simply “short prompts are better.” It is that stronger models need a different kind of context.

Older models often needed rigid rules, repeated instructions, and worked examples. Newer models can use more judgment, infer behavior from well-designed interfaces, and retrieve specialized context only when needed. Anthropic describes six shifts:

Two of these mattered most in our audit.

Progressive disclosure means the main instructions do not need to contain every deployment command, review checklist, and incident note. They need to tell the agent what specialized guidance exists and when to load it. Claude Code applies this through selectively loaded skills and deferred tools whose full definitions can be retrieved with ToolSearch.

One authoritative instruction prevents duplicated guidance from drifting. If “run tests after editing” appears in the system prompt, CLAUDE.md, a coding skill, and a tool description, one copy will eventually change while another survives.

These are not six ways to delete text. They are six ways to put information where it becomes useful and verifiable.

Always loaded

Small, universal, costly to miss

• Product identity
• Safety boundaries
• Non-discoverable gotchas

When relevant

Specialist and operational detail

• Deployment runbooks
• Regional policies
• Review skills

Source of truth

Current and executable

• Code and configuration
• Tests and schemas
• Working references

Our production instructions had become a history book

Our CLAUDE.md started as a concise guide to an Astro landing page. Over five months and 13 revisions, the site absorbed a finance suite, publishing controls, retired applications, new routes, infrastructure changes, and lessons from content failures.

The file now contains:

None of that happened because someone decided to write a long prompt. Each addition made sense at the time. A deployment failed, so the exact command was recorded. A route moved, so its history was added. A tax-content error escaped, so the failure and new safeguards were documented.

This is how prompt bloat usually happens: one reasonable addition at a time.

Three kinds of prompt bloat

Our audit exposed three different problems that are often grouped under “long prompt”:

• Harmless bloat: true but obvious information the agent could discover, such as a hand-maintained file tree.
• Misplaced context: useful details loaded for unrelated work, such as production log commands during a homepage copy edit.
• Dangerous staleness: authoritative instructions that no longer match reality.

The first wastes tokens. The second distracts attention. The third can send the work in the wrong direction.

The real cost was stale confidence

During the audit, we found instructions that no longer matched the code:

• The file tree still listed an impact-diagnostic.astro page that has been deleted and replaced by a redirect.
• An early routing section described /app and /api as active; a later section correctly said both were retired.
• The finance section said article JSON-LD was carried verbatim because it varied too much to generate. The current content schema says the opposite: JSON-LD is derived, and explains the production bug that forced the change.

A capable agent may inspect the implementation and notice the contradiction. But we have made its job harder: the file presented as project ground truth is arguing against the code.

That is the risk hidden by the 80% headline. A long prompt costs context. A stale prompt costs judgment.

Keep, move, or test and delete

We classified its eight major sections. Two—the project gotchas and financial-content governance—earned permanent context. Five—development, deployment, infrastructure, routes, and finance implementation—belonged in focused guides with short pointers from the main file. The hand-maintained layout section needed rebuilding or deletion.

That first pass also found three stale claims requiring correction. This was an editorial audit, not a completed behavioral evaluation: the placement decisions remain hypotheses until a slimmer version is tested on representative work.

Keep always loaded

Some context earns permanent space because the cost of missing it is high or because it cannot be discovered reliably from the repository:

• Never edit production servers directly; deployments are built from Git.
• Indexed tax URLs must not move.
• Time-sensitive tax claims require primary sources and an as_of date.
• Never invent a human reviewer or publish an AI as a schema.org Person.
• Fix corrupted ground truth before fixing the pages it influenced.

These are not style preferences. They protect deployment integrity, public URLs, and financial information. “Make the prompt shorter” is not permission to remove safety boundaries.

Move closer to the work

Large sections matter, but not on every task:

• deployment and log commands belong in a deployment runbook
• historical routing details belong in infrastructure documentation
• finance collection internals belong beside the finance schema
• publication verification belongs in the relevant content skill

The main instructions file can retain a short pointer: what resource exists and when the agent should read it. A homepage copy edit does not need the complete history of a retired subdomain.

Test and delete

The easiest candidates are duplicated, discoverable, or already contradicted by the code:

• a hand-maintained file tree that a search can reconstruct
• retired routes described as active
• old implementation decisions preserved after the implementation changed
• commands already defined in package.json

Deletion should follow correction. If an instruction is wrong today, remove or update it now. If it merely looks unnecessary, test its removal.

Here is what that change looks like at the instruction level:

The part the 80% headline leaves out

Anthropic’s own recent history shows why. In April 2026, a Claude Code instruction limiting verbosity passed weeks of internal testing. A broader evaluation later revealed a 3% coding-performance drop, and Anthropic reverted the prompt change.

The lesson is not that evals make every change safe. It is that your first eval set may also be incomplete.

How to audit your own instructions

You do not need to begin with a token counter. Begin with provenance.

For every section, ask:

1. What failure caused us to add this? If nobody knows, mark it for investigation.
2. Is it still true? Check the implementation, not another document repeating the same claim.
3. Does it matter across most tasks? If not, move it beside the workflow that needs it.
4. Can the agent discover it safely? Repository structure is discoverable; an unwritten production constraint is not.
5. What would detect a regression? Name the task, expected behavior, and pass condition before deleting.

Then remove one coherent group at a time and run representative tasks against both versions. Compare correctness, policy violations, unnecessary actions, token use, and latency. Restore the instruction if quality falls.

Claude Code users can also run claude doctor (or /doctor in a session), which Anthropic says can rightsize CLAUDE.md files and skills. It is a useful second opinion, not a substitute for an eval built from your own failures.

Do not begin with security rules, privacy boundaries, destructive-action safeguards, or instructions protecting customer data. Start with stale maps, duplicated guidance, and details that already have a better home.

Copy this audit worksheet

Instruction or section:
Why was it added?
Is it still true?
Is it needed for most tasks?
Can the agent discover it safely?
Better location:
Regression task and pass condition:
Decision: KEEP / MOVE / CORRECT / TEST AND DELETE

Download the Markdown worksheet and complete its “regression task” field before removing anything.

Correctness first, length last

Our audit did not earn us an “80% smaller” badge. It did something better: it exposed contradictions before they shaped future work and separated universal safeguards from task-specific knowledge.

Anthropic found that more than 80% of its prompt no longer earned permanent context. We found something worse: parts of ours no longer told the truth.

Audit for correctness first, placement second, and length last.