Anthropic Adds a Claude Code Audit for Skills That Waste Context

Anthropic has added a Claude Code audit that identifies loaded skills consuming context without being used and points developers to settings that can disable them. Released on September 4, 2026, /skill-doctor tracks whether each eligible skill was invoked and shows its context cost. Developers can use the report to disable listing entries that consume context on every turn without being invoked.

A charge on every turn

“Every skill in the skill listing adds to your context on every turn, whether or not Claude ever uses it,” Anthropic’s documentation says. The listing gives Claude the names and descriptions it uses to decide which procedures apply to a request. That means an unused entry still occupies room that could hold code, tool results or the conversation itself.

The report covers session skills other than bundled and enterprise entries. It flags listed skills that have never been invoked, tells the user where to switch them off and identifies plugins that have not been used recently. It requires Claude Code version 2.1.252 or later. Remote Control sessions opened from a phone or browser cannot generate the report.

Anthropic also caps the combined description and when_to_use text for each listing entry at 1,536 characters. The listing budget defaults to 1% of the model’s context window. Developers can raise it through skillListingBudgetFraction; the documentation’s example sets the allowance at 2%.

The earlier cleanup

The new audit follows two changes to /doctor in July 2026. Version 2.1.205 made it a full setup checkup on July 8, while version 2.1.206 added a check on July 9 that proposes trimming checked-in CLAUDE.md files by removing material Claude can infer from the codebase. The command can also deduplicate local copies, move always-loaded guidance into skills or nested files that load when needed, and ask for confirmation before changing anything.

“We removed over 80% of Claude Code’s system prompt for models like Claude Opus 5 and Claude Fable 5 with no measurable loss on our coding evaluations,” Thariq Shihipar, a member of technical staff at Anthropic, wrote on July 24, 2026. He said the team had been overconstraining the product through its system prompt, CLAUDE.md files and skills.

Internal transcripts showed instructions colliding within one request, including “leave documentation as appropriate” beside “DO NOT add comments.” Anthropic replaced a blanket rule that code should default to no comments with a direction to match the surrounding code’s comment density, naming and idiom. Shihipar also wrote that examples can constrain newer models to a particular “exploration space.”

What usage cannot prove

Anthropic has not published the coding evaluations behind its “no measurable loss” claim. No independent measurement has been published showing what happens to output quality when the system prompt is cut.

The tool records whether a skill ran during a session and how much listing space it consumed, but not whether the skill improved the result. A rarely invoked but valuable procedure is therefore indistinguishable from dead weight in the report.

One commenter on a small Ask HN thread disputed the idea that existing configurations had become incompatible. The thread had 5 points and 6 comments when captured on September 7, 2026. “we moved from opus 4 to opus 5 without changing anything and no specific conflicts,” gojkoa wrote. The commenter called Anthropic’s advice a recommendation to shorten configuration, “not a backwards compatibility issue.”

Controls stay with the user

The broader /doctor command remains available even when the disableBundledSkills setting turns off the other bundled skills. /skill-doctor provides the narrower usage report, while /doctor reviews the configuration itself.

Neither command silently edits a developer’s files. /doctor presents its findings first and asks for confirmation before making a change.

Microsoft’s David Fowler Says ‘Typing Code Is Absolutely Over’

Microsoft distinguished engineer David Fowler said, “Typing code is absolutely over,” arguing that manual entry is becoming a smaller part of software work. Microsoft’s existing Aspire workflow gives

The Implicator

Repo Radar: 5 GitHub Projects Worth Your Week

Five projects climbed GitHub this week and none of them is a model. Each sits in the layer a team can own outright: the rules an agent reads before it types, the diagrams it hands back, the workspace

The Implicator

Why AI Coding Agents Need Context Compaction Before 100 Percent

Anthropic engineer Thariq Shihipar published a session-management guide for Claude Code on April 15, pairing the assistant's 1 million-token context window with a decision table that lists which comma

The Implicator

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai