ImageFirm Run the audit
ImageFirm · Infographic series N°09 · Fact-check · 26 Oct 2026

Prune your Claude setup. Don't gut it.

A widely shared infographic tells you to delete most of your Claude setup and keep a handful of habits. We checked every one of its 29 claims against Anthropic's own documentation. The direction is right. The absolutes are not — and most setup bloat turns out to be misplacement, not excess.

Claims checked 29 Anthropic pages cited 20 Model in scope Claude Opus 5.5 Verified 26.10.2026
01 · Verdict

Just over half the card holds as printed. The rest needs a red pen.

The infographic is current — it names Opus 5.5 and the merged Chat/Cowork experience — and its instinct is sound. Where it fails, it fails by turning "prune and test" into "delete".

Its strongest line is almost Anthropic's own wording. Anthropic's Claude Code guidance warns that an over-long instruction file gets half-ignored because important rules drown in noise. [6]

Two lines would do damage if followed literally: emptying your instructions and deleting every reusable prompt. Anthropic's method is prune each line, test, and move recurring procedures into skills that load on demand. [6][8]

"It's all one" is half true. Chat and Cowork merged into one conversation; Claude Code still has its own tab. The Capabilities toggle the card points to is an organisation-owner setting, not a personal one. [11][12]

"Ask for HTML" predates living artifacts. Docs, Slides and Design now come back as private, shareable, editable pages that export to Word, PowerPoint or PDF. Ask for the outcome, not the format. [18][19]

Most bloat is misplacement. Facts sitting in instructions, procedures sitting in standing rules, folder maps Claude could read for itself. The fix is a placement map, not a purge. See §03

Verdict distribution 16 of 29 hold as printed · 10 need correcting · 3 mislead
    02 · Claim ledger

    Every claim, the evidence, and the move.

    Claims are paraphrased from the card. A struck-through claim should not be followed as printed; a wavy underline marks a claim that needs a correction. Open any row for the evidence and the source.

    The "delete" list

    D·01Empty your Instructions for Claude.Overstated
    What Anthropic says
    Keep standing instructions short and test each line; don't abandon them. Claude also generalises better when a rule comes with its reason. Since the merge, Cowork's global instructions live inside Instructions for Claude, so this box is now the single home for standing rules.
    Do this instead
    Keep only what Claude can't infer and would otherwise get wrong — language rules, compliance limits, format defaults. Run the line test on each. Expect a short box, not an empty one.
    Sources
    [6][2][11]
    D·02Turn off your writing skills.Nuanced
    What Anthropic says
    Opus 5.5 writes more plainly by default, putting the important information first, and follows the writing rules you give it. An installed skill costs roughly 100 tokens of metadata until it triggers.
    Do this instead
    Cut generic "write better" skills. Keep one that encodes a voice you own — brand terms, house style, claims you never make. Run your three most common writing tasks with and without it, then decide.
    Sources
    [5][7][8]
    D·03Delete the over-organised Cowork folder.Holds
    Basis
    Anthropic's context guidance favours letting the agent retrieve files just in time over pre-loading structure, and its Claude Code guidance says to let Claude fetch what it needs. No Anthropic page addresses folder depth directly — this verdict is our judgment.
    Do this instead
    Flatten, and name files descriptively — Anthropic lists clear filenames as a retrieval aid for projects. Point Claude at the one or two sources of truth.
    Sources
    [9][6][17]
    D·04Collapse folders nested inside folders.Holds
    Basis
    For skills, Anthropic recommends keeping file references one level deep from the entry file. The same logic applies to any working folder Claude has to navigate.
    Do this instead
    One level of structure, named by what is inside (contracts, brand, data), not by workflow stage.
    Sources
    [8]
    D·05Delete the about-me file you keep rewriting.Nuanced
    What Anthropic says
    Memory is on by default on Free, Pro and Max, can be reviewed and edited under Settings › Your Claude, and keeps a separate memory per project. A static profile file duplicates it and drifts out of date.
    Do this instead
    Let memory hold facts about you and review it monthly. Keep standing rules in Instructions for Claude. In Claude Code, repo-wide context belongs in CLAUDE.md, which is read every session.
    Sources
    [15][6]
    D·06Delete the anti-AI-writing file.Nuanced
    What Anthropic says
    Describing what you want works better than listing what to avoid. The Opus 5.5 guide adds that generic "avoid the AI look" instructions mostly trade one default for another. It makes that point about design; in our experience the same holds for prose. Named, specific patterns work.
    Do this instead
    Replace "don't sound like AI" with three to five positive, checkable rules. Read the first output, then add only the patterns you actually saw.
    Sources
    [2][1]
    D·07Remove the folder protocol from your instructions.Verified
    What Anthropic says
    Its CLAUDE.md guidance lists file-by-file descriptions, and anything Claude can learn by reading the files, among the things to leave out.
    Do this instead
    Keep a pointer only where the answer can't be discovered — for example, which spreadsheet is the source of truth for prices.
    Sources
    [6]
    D·08Stop choosing between Chat, Cowork and Code — it's all one.Nuanced
    What Anthropic says
    Chat and Cowork have merged into one conversation with no mode to pick. Claude Code has not: it stays in its own Code tab. Incognito chats still open in the previous experience, without file creation or code execution.
    Do this instead
    Use one conversation for questions, research and documents. Open the Code tab for repository work.
    Sources
    [11][16]
    D·09Delete your prompt library — you only need one prompt.Overstated
    What Anthropic says
    Repeatable workflows belong in skills, which cost almost nothing until triggered and can be invoked by name. One-off prompts can go; recurring procedures should not be retyped from memory.
    Do this instead
    Turn any prompt you reuse into a skill with a precise "use when" description (our rule of thumb: the third reuse). Delete the rest.
    Sources
    [6][7]
    D·10Delete the 31-page prompt guide.Holds
    Basis
    Anthropic publishes a separate prompting guide per model, and the advice changes between releases — Opus 5.5's guide, for example, retires instructions that were standard for earlier models. A saved copy ages with the model it was written for.
    Do this instead
    Bookmark the live guide for the model you use. Re-read it when you switch.
    Sources
    [1][2]
    D·11Stop producing report_v3_FINAL.docx.Holds
    Basis
    Claude Docs are living documents in your Artifacts tab: shared by permission, edited by Claude and collaborators, and exportable to Word, PDF, Markdown or Google Docs.
    Do this instead
    Ask for the document, not the file. Export to .docx only when a counterparty requires it — banks, notaries, procurement portals.
    Sources
    [18][19]
    D·12Strip projects stuffed with files, rules and constraints.Nuanced
    What Anthropic says
    Files are not the problem. When project knowledge nears the context limit, retrieval switches on automatically and a project can hold up to ten times more. Projects also keep their own memory. Stacked rules are what compete for attention.
    Do this instead
    Keep the reference files. Cut project instructions to what is unique to that body of work. Name files descriptively.
    Sources
    [17][15]
    D·13Retire skills written for older models.Verified
    What Anthropic says
    Test skills with every model you plan to use. The Opus 5.5 guide says to remove instructions that ask for reasoning in the reply or tell the model not to think, and to consider dropping "think carefully" lines. Emphatic wording written to push older models now tends to overtrigger.
    Do this instead
    After every model change, re-run each skill on a real task. Rewrite the ones you still need rather than deleting them.
    Sources
    [8][1][2]
    D·14Merge skills that overlap.Holds
    Basis
    Claude picks a skill from its description alone, potentially among a hundred or more. Anthropic's context guidance names overlapping tool sets as a common failure mode.
    Do this instead
    One skill per job. If two descriptions could fire on the same request, merge them or sharpen the "use when" clause.
    Sources
    [8][9]
    D·15Shorten instructions so long Claude ignores half.Verified
    What Anthropic says
    Close to Anthropic's own wording: an over-long instruction file gets half-ignored because the important rules are lost. Emphasising many lines makes none of them stand out.
    Do this instead
    Prune ruthlessly. If Claude already behaves correctly without a line, delete it. If a rule must hold every time in Claude Code, make it a hook.
    Sources
    [6]
    D·16Don't run thinking on high to ask about the weather.Nuanced
    What Anthropic says
    Right idea, outdated mechanics. On Opus 5.5 thinking can't be switched off; effort is the control. At lower effort Claude can skip thinking on simple problems, and Anthropic recommends the defaults for everyday tasks.
    Do this instead
    Leave effort at the default for routine work; raise it for the conversation when the task is genuinely hard. See the ladder in §05.
    Sources
    [3][14]
    D·17Drop the connectors added "just in case".Verified
    What Anthropic says
    In one documented setup, 58 tools from five servers took about 55K tokens before the conversation began, and wrong tool selection was the most common failure. Deferred tool search cut token use by 85% and improved accuracy in Anthropic's tests — but it only helps where the client supports it (figures from November 2025).
    Do this instead
    Connect what your recurring work needs. Disconnect the rest; reconnecting takes a minute.
    Sources
    [10]
    D·18Remove the plugin pack you used once.Holds
    Basis
    A plugin bundles skills and connectors, so it inherits both costs above: more descriptions to choose between and more tool definitions.
    Do this instead
    Uninstall after the task. Note the plugin's name if you'll need it again.
    Sources
    [11][10]
    D·19Delete screenshots of other people's prompts.Holds
    Basis
    Anthropic advises examples that mirror your actual use case. A prompt tuned to someone else's context, data and model rarely transfers.
    Do this instead
    Keep the idea in one line; write your own version against your own task.
    Sources
    [2]
    D·20Stop perfecting a setup you haven't used.Holds
    Basis
    Anthropic's skill method runs the other way round: run real tasks without the setup, record where Claude fails, then write the minimum instructions that fix those failures.
    Do this instead
    Start bare. Add a rule only after a real failure, and note which failure each rule fixes.
    Sources
    [8]

    The "keep" list

    K·01One Claude: unify Chat and Cowork under Settings › Capabilities.Nuanced
    What Anthropic says
    Pro and Max users have nothing to switch on: the merged experience rolls out account by account, and once an account moves it can't go back. The "Chat and Cowork unified" toggle is an organisation-owner setting for Team (beta) and Enterprise. Enterprise turns it on by default on 16 November 2026 unless an owner keeps it disabled; it needs Chat and cloud code execution on.
    Do this instead
    Individuals: check the message box — no Chat and Cowork buttons means you have it. Organisation owners: decide before 16 November.
    Sources
    [11][12][13]
    K·02Opus 5.5 + Medium + Auto as your default.Verified
    What Anthropic says
    Medium is Opus 5.5's default effort, and Anthropic reports Opus 5.5 at medium matching or beating Opus 5 at high on its coding and knowledge-work evaluations. Auto is the default permission setting on Pro, Max and Team; Enterprise defaults to Manual.
    Do this instead
    Accept the defaults, but set effort per task. Treat Auto as a convenience for reversible work, not a governance decision — see the ImageFirm position below.
    Sources
    [1][3][13]
    K·03One clear goal: what you want, and what it's for.Verified
    What Anthropic says
    Giving the reason behind an instruction helps: Anthropic's own example shows Claude generalising from a single explanation to behaviour that was never spelled out.
    Do this instead
    State the outcome and the purpose in the first two lines of the request.
    Sources
    [2]
    K·04One line of context: who it's for, what good looks like.Nuanced
    What Anthropic says
    Audience and quality bar help. For anything agentic, though, Anthropic's first recommendation is a check Claude can run — tests, a build, a screenshot to compare — so it knows when the work is actually done.
    Do this instead
    Add a done-test: "Done when the totals reconcile to the bank export." "Done when every link resolves."
    Sources
    [6]
    K·05The restate line: Claude repeats your goal before it starts.Nuanced
    What Anthropic says
    Planning first pays off on ambiguous, multi-file or unfamiliar work, and for larger features Anthropic suggests letting Claude interview you. For tasks you could describe in one sentence, it says to skip the plan.
    Do this instead
    Ask for a restatement or plan only when a wrong guess would be expensive.
    Sources
    [6]
    K·06Ask for HTML: a live link, not a frozen file.Outdated
    What Anthropic says
    Artifacts are now hosted pages that start private and are shared by link. Docs, Slides and Design are live, editable artifact types that export to Word, PowerPoint, PDF or standalone HTML. Saying "HTML" no longer buys the live link — asking for the outcome does.
    Do this instead
    Ask for "a doc", "a deck", "a dashboard". Ask for HTML when you want an interactive tool, or a page you will host yourself.
    Sources
    [19][18]
    K·07Memory on; switch it off per chat with the +.Verified
    What Anthropic says
    The + menu in the message box turns memory off for one chat; the choice locks once the chat starts, and the chat stays in your history. For a chat that is neither remembered nor kept in history, use incognito (the ghost icon) — retained 30 days by default.
    Do this instead
    Use + for sensitive one-offs you still want on record; incognito when nothing should persist. Inside a project, turn off account memory to keep it fully separate.
    Sources
    [15][16]
    K·08A clean chat: starting from scratch is often better.Verified
    What Anthropic says
    After two failed corrections on the same issue, Anthropic advises clearing the session and restarting with a sharper prompt; a fresh session with a better prompt usually beats a long one full of corrections.
    Do this instead
    Restart with what you learned folded into the first message.
    Sources
    [6]
    K·09Delete it all again in six months.Nuanced
    What Anthropic says
    The line traces to a remark by Claude Code's creator in a Y Combinator interview, not to Anthropic's documentation. Models now ship faster than that: Opus 5 (24 July 2026) to Opus 5.5 (22 September 2026) took two months.
    Do this instead
    Audit on each model change, not on a calendar. Soft-delete first — keep an undo path — and restore only what a real task proves you need.
    Sources
    [20][5][21]
    03 · Placement map

    Every instruction has one right home.

    Bloat is usually a rule in the wrong place: loaded every time when it is needed once, or written as a rule when it is a fact. Move before you delete.

    Loads · this message only

    The prompt

    Put here: the goal, why it matters, who it is for, and how you will judge done.

    Keep out: anything you would repeat tomorrow.

    Loads · every conversation

    Instructions for Claude

    Put here: standing rules Claude can't infer — language, hard constraints, format defaults, each with its reason.

    Keep out: facts about you, procedures, folder maps. Settings › Your Claude.

    Loads · when relevant

    Memory

    Put here: facts about you, your people and your ongoing work. Claude maintains it; you review it.

    Keep out: rules about behaviour. Review under Settings › Your Claude › Manage.

    Loads · inside that project

    Project

    Put here: reference files (retrieval handles volume), instructions unique to that work, its own memory.

    Keep out: your global rules, repeated.

    Loads · ~100 tokens idle, body on trigger

    Skill

    Put here: repeatable procedures and house style, with a precise "use when".

    Keep out: one-offs, and two skills doing one job.

    Loads · tool definitions

    Connector

    Put here: live data your recurring work needs — calendar, drive, CRM.

    Keep out: "just in case" apps. Customize › Connectors.

    Loads · every session in the repo

    CLAUDE.md (Claude Code)

    Put here: commands Claude can't guess, conventions that differ from defaults, known gotchas.

    Keep out: anything readable from the code itself.

    Runs · deterministically

    Hook (Claude Code)

    Put here: rules that must hold every time — formatters, blocked paths, checks.

    Keep out: judgment calls; instructions are advisory, hooks are not.

    Sources: [11][15][17][7][10][6]

    04 · The line test

    One question per line. Rewrite what survives.

    “Would removing this cause Claude to make mistakes?”Anthropic, Best practices for Claude Code [6]

    If the answer is no, delete the line. If yes, keep it — and check its wording against current guidance. Six rewrites we see most often:

    CRITICAL: You MUST ALWAYS use web search if in doubt!!! Search when the answer depends on current prices, laws, or who holds a role.

    Why. Emphasis and blanket triggers written for older models now cause overtriggering. Name the condition instead. [2]

    Never use markdown. Write in flowing prose; use a table only when comparing three or more options.

    Why. Describing the target works better than a prohibition. [2]

    Think step by step and show your reasoning before every answer. (Delete. Raise effort for hard tasks instead.)

    Why. Opus 5.5 thinks adaptively and always; effort is the control. Prompts asking for reasoning in the reply may be declined. [1][3]

    Don't sound like AI. Avoid AI slop. Lead with the answer. Plain verbs. No closing summary. No rhetorical questions.

    Why. Generic bans swap one default for another; named patterns stick. [1][2]

    Be concise. Status updates: three lines — done, blocked, next.

    Why. A checkable format beats an adjective. Be specific about the output you want. [2]

    Reply in Spanish. Customer emails go out in formal Spanish because our clients are in Paraguay.

    Why. The reason lets Claude generalise — invoices and follow-ups stay formal too, without another rule. [2]

    05 · Effort ladder

    Opus 5.5: thinking is always on. Effort sets the depth.

    The card's "Medium" is correct as a default. The professional habit is to move the dial per task — the change applies from the next response.

    1. LowRoutine lookups, short rewrites, high-volume runs. Stretches your usage furthest.
    2. Medium · defaultMost work. Anthropic reports it matching or beating Opus 5 at High on its coding and knowledge-work evaluations.
    3. HighComplex analysis, planning, detailed document review — Anthropic's balance pick for quality and speed.
    4. Extra highLong agentic and coding runs; Anthropic suggests trying it first for complex coding on recent models.
    5. MaxCorrectness-critical work only. Slower; reserve it for tasks where you have measured a gain over Extra high.

    Sources: [3][14][1]. Thinking cannot be switched off on Opus 5.5 at any level; lower effort lets Claude skip it on simple problems.

    ImageFirm position · Governance

    Auto is a default. Authority is a design decision.

    Auto lets Claude keep working while automated safety checks screen each action; Manual asks before Claude acts. Anthropic ships Auto on Pro, Max and Team and Manual on Enterprise — one product, two governance postures. A leaner setup should make Claude faster, never less accountable.

    • AutoReversible work in a contained space: drafts, research, analysis, files you can regenerate.
    • ConfirmAnything that sends, publishes, pays, deletes, or changes a live system.
    • RecordWhich rule fixes which failure — so the next pruning is evidence-led.
    Explore the ImageFirm library
    06 · Instruction linter

    Paste your instructions. See what to question.

    A heuristic first pass, not a verdict: it flags patterns Anthropic's guidance warns about, with line numbers. The line test decides.

    Runs entirely in this page. Nothing you paste is sent or stored.

    0Lines to defend
    0Words
    0≈ Tokens (words × 1.3, rough)
    • Findings appear here as you type.
    07 · Audit prompt

    Let Claude audit the setup. Keep the decision.

    Paste into a new conversation. Every cut must cite current Anthropic guidance; nothing changes until you apply it and test it on your own tasks.

    Audit my Claude setup against Anthropic's current guidance. Do not change or delete anything yourself.
    
    1. Read what you can see of my setup: Instructions for Claude, this project's instructions, my installed skills and connected apps, and any CLAUDE.md files you can open. If you can't see something, ask me to paste it.
    2. Read the live guidance first: Anthropic's prompting best practices, the prompting guide for the model you are running, the Claude Code best-practices page, and the skill authoring best practices.
    3. Go line by line. Give each instruction one verdict (KEEP, CUT, REWRITE, or MOVE to memory, a project, a skill, or a hook) using one test: would removing this line make you get something wrong?
    4. Justify every CUT with the guidance it rests on, and link the page. If you can't find support, the verdict is KEEP.
    5. Flag emphatic wording, rules phrased only as prohibitions, "think carefully" lines, references to older models, duplicates, and anything you would do correctly without being told.
    6. Return a table (line, verdict, reason, source), then the rewritten setup, then the three tasks I should run to test it before I apply anything.

    08 · Sources

    Every verdict rests on a page you can open.

    Anthropic documentation, read on 26 October 2026. Product details move quickly — recheck before acting on anything time-sensitive.

    1. Prompting Claude Opus 5.5platform.claude.com · Claude Platform Docs
    2. Prompting best practicesplatform.claude.com · Claude Platform Docs
    3. Effortplatform.claude.com · Claude Platform Docs
    4. What's new in Claude Opus 5.5platform.claude.com · Claude Platform Docs
    5. Introducing Claude Opus 5.5anthropic.com · 22 Sep 2026
    6. Best practices for Claude Codecode.claude.com · Claude Code Docs
    7. Agent Skills — overviewplatform.claude.com · Claude Platform Docs
    8. Skill authoring best practicesplatform.claude.com · Claude Platform Docs
    9. Effective context engineering for AI agentsanthropic.com/engineering · 29 Sep 2025
    10. Introducing advanced tool use on the Claude Developer Platformanthropic.com/engineering · 24 Nov 2025
    11. Claude Cowork and chat are one Claudesupport.claude.com
    12. Set up the unified Claude experience for your Enterprise organizationsupport.claude.com
    13. Get started with Claude Coworksupport.claude.com
    14. Change the model, effort, and thinking settingssupport.claude.com
    15. Use Claude's chat search and memorysupport.claude.com
    16. Use incognito chatssupport.claude.com
    17. Retrieval augmented generation (RAG) for projectssupport.claude.com
    18. Get started with Claude Docssupport.claude.com
    19. What are artifacts and how do I use them?support.claude.com
    20. System Card: Claude Opus 5anthropic.com · 24 Jul 2026
    21. Opus 5: delete your CLAUDE.mddev.to · secondary source, for the interview attribution only

    Method. Each claim on the card was paraphrased and checked against the pages above. Verified means Anthropic states it. Holds means Anthropic's guidance supports it in principle without saying it outright — our judgment, labelled as such. Nuanced: right direction, wrong detail. Overstated: harmful if followed literally. Outdated: the product has moved on. Claims come from the infographic "Delete 90% of your Claude setup" by Ruben Hassid (How to AI). ImageFirm is not affiliated with Anthropic.