Anthropic’s prompting guides for Claude Sonnet 5.5 and Opus 5.5 name four habits that now make results worse: discouraging tool use, asking for reasoning in the response, telling the model not to think, and injecting harness text after every tool result. I ran /doctor prompt-audit on this site’s Claude Code config to find them. It found zero of the four, and 14 findings of a different kind. The worst: three of my four subagents were still briefed on a workspace this repo has never been part of.
The four habits the 5.5 guides retire
Each row below traces to a sentence in Anthropic’s own guides. The models differ, so the table says which guide covers what.
Three terms first:
- Harness is the program that runs the model and feeds it tool results - Claude Code, or your own agent loop.
- Effort is the API setting that controls how much the model thinks. Both guides say lowering it cuts thinking more reliably than a prompt instruction does.
between_toolsis Sonnet 5.5’s lowest thinking setting: no thinking up front, only short notes between tool calls.
| Habit | What goes wrong | Guide | Fix |
|---|---|---|---|
| “Only use tools when strictly necessary” / “minimize tool calls” | The model answers from training knowledge when a search would catch details that changed | Sonnet 5.5 | Remove the line; add one saying when to check current sources |
| “Include your reasoning in the response” | Can be declined as a reasoning_extraction refusal; server-side fallback won’t retry it | Sonnet 5.5, Opus 5.5 | Remove; read reasoning from display: "summarized" thinking blocks |
| “Don’t think” | Sonnet 5.5 with between_tools: more likely to write internal XML tags into visible output. Opus 5.5: thinking can’t be turned off, so remove the rule “either way” | Sonnet 5.5, Opus 5.5 | Lower effort instead of asking in prose |
| Harness text after every tool result (a token countdown, per-step instructions) | Sonnet 5.5 can treat a genuine user message typed mid-task as a possible prompt injection | Sonnet 5.5 | Put user text in the user turn, after the last tool_result; no countdown in interactive sessions |
The last row is not a prompt line. The harness adds that text, and that’s why a grep of your CLAUDE.md never finds it. The Opus 5.5 guide even points the other way for multiagent setups: there it recommends the harness append an elapsed-time line like elapsed 340s / 1200s so a team of agents paces itself. Whether harness text after tool results hurts depends on the model. The two lines most likely to sit in your own files have their own notes: don’t tell Sonnet 5.5 to minimize tool calls and why “show your reasoning” gets refused.
This is the third piece in a thread. Why agents ignore your CLAUDE.md argued from mechanism that more instructions don’t buy more control. A 129-rule CLAUDE.md replied like an empty one measured it. This one checks what an official audit tool actually catches.
What /doctor prompt-audit does
Claude Code 2.1.283 (September 25, 2026) added the command. The changelog describes it as auditing “your CLAUDE.md files, skills, agents and commands for prompting patterns written for older models”. The same release says “stale paths, stale commands and contradicting instruction files now lead the report”.
/doctor prompt-audit/checkup prompt-audit is the same command. The audit targets whatever model your session runs on - Opus 5.5 on Claude Code 2.1.285, in my case - and returns two things: a report with file:line, pattern, reason and confidence per finding, and a proposed diff (a line-by-line patch you can take piece by piece). Nothing gets applied unless you ask. Because a model does the reading, a maintainer weighing it up in issue #1500 concluded “it cannot be a CI gate: it needs the CLI and a model call”.
The before report
Scope: CLAUDE.md, AGENTS.md and their imports, 4 subagents and 14 project skills. It also scanned the 66 installed skills under ~/.claude in report-only mode. The full raw report, before and after, is in this gist.
| Group | Findings |
|---|---|
| Dated prompt text (older-model habits) | 4 |
| Config rot (stale paths, commands, contradictions) | 10 |
| Tool descriptions / request code | not applicable |
| Total | 14 (10 with a proposed fix, 4 flag-only) |
| The four 5.5 habits above | 0 (project, ~/.claude, 66 installed skills) |
The high-confidence findings were all rot:
- Three subagents described another workspace.
code-reviewer,explorerandresearcher(the fourth,content-reviewer, was written for this site) were copied from a multi-repo setup, and each still named its repos:b24,react_panel,b24depot,product_ops. None of those exist here. The reviewer was told to check PHP 8.4 conventions and GraphQL contracts on an Astro site. - The reviewer diffed against a branch that doesn’t exist. It ran
git diff master...HEAD, but this repo’s default branch ismain. - Two skills called commands that were never installed.
grill-meconsisted of a single line, “Run a/grillingsession”, and nogrillingskill exists.decision-mappingreferenced/grilling,/domain-modeling,/prototypeand/to-prd, also absent. Both arrived throughskills-lock.jsonwithout the sibling skills they call.
Medium-confidence findings covered the older-model habits the audit guide lists: line caps like “Keep the whole response under ~60 lines”, a 240-character paragraph cap, and an all-caps “You NEVER edit files” in an agent whose tool list already can’t write. One cross-file contradiction also turned up: the storytelling skill claimed humanize-copy rewrites in place, while humanize-copy itself defaults to a sibling file.
The fix
Ten findings came with a proposed hunk, and I took all ten. The largest one:
-description: Reviews a diff or files in a fresh, clean context. Use after completing an implementation, before commit/MR, to review changes in any repo (b24, react_panel, b24depot, product_ops) or cross-repo work. Read-only.
+description: Reviews a diff or files in a fresh, clean context. Use after completing an implementation, before commit/PR. Read-only.
-1. ... otherwise run `git diff master...HEAD` ...
+1. ... otherwise run `git diff main...HEAD` ...
-Keep the whole response under ~60 lines.
+The main session reads this, so report findings only - no restated diff, no preamble.Four findings stayed flag-only by design, for example AGENTS.md importing CLAUDE.md, which loads the same text twice if both files are read. The audit leaves a duplicate alone when the copies agree.
The after report
| Finding | Status |
|---|---|
| 3 subagents briefed on another workspace | fixed |
git diff master...HEAD | fixed (main) |
| 4 uninstalled slash commands | fixed (removed, or repointed to skills that exist here) |
| Line and character caps, all-caps rule | fixed |
storytelling / humanize-copy contradiction | fixed |
content-reviewer “~50 lines” cap | missed by the before run, caught by the re-scan, fixed |
| 4 flag-only items | unchanged |
content-reviewer carried the same line-cap pattern as the other three agents, and the first pass walked past it. Run the audit twice before you trust a clean report - the rule, in one note.
I ran the same audit on the leanharness template, the minimal harness from the earlier benchmark. It had zero of the four 5.5 habits too, and one real leak: the Non-negotiable scope rule named src/billing.js, a file that exists only in the benchmark’s fixture, in the CLAUDE.md every user installs. It is now a placeholder the benchmark fills in, so the measured install stays byte-identical. The template’s audit record lists what was fixed and what stays on purpose.
Three of the fixed skills - grill-me, decision-mapping and edit-article - are copied in from upstream repos and pinned by skills-lock.json. Their hashes no longer match upstream, so the next skills update will quietly put the old text back. For those three, the durable fix belongs upstream or in a fork.
What this changes about “more instructions, more control”
The 5.5 guides make the familiar case: lines written to push an older model harder now push the new one the wrong way. On this repo, nobody had written those four lines. The instructions doing damage were plain, calm and wrong, and nothing announced it:
- A reviewer told to judge PHP conventions raises no error. It reviews an Astro diff against PHP rules and reports back.
- A skill that says “run
/grilling” never fails loudly. The model improvises something grilling-shaped.
Instruction count feels like control because each new rule looks like coverage. Each one is also a claim about your repo that can go stale, and nothing re-checks it unless you run something that does. The audit’s ordering says the same thing: stale paths, stale commands and contradictions lead the report, ahead of the prompt-style patterns.
Run it yourself
- Update Claude Code to 2.1.283 or later, then run
/doctor prompt-auditin the repo. - Save the raw report before you change anything. It’s your only before snapshot.
- Take the high-confidence hunks first. They’re the ones the repository itself contradicts.
- Run it again after the fixes. Mine caught a line the first pass missed.
Then read Why agents ignore your CLAUDE.md for why the file you’re left with should stay short, and how effort and thinking settings work for the lever that replaces “don’t think”.