Working with Claude, Tested

A guide to Anthropic’s assistant for people who direct it rather than program: which Claude you are using, how to give it rules it keeps, and which popular tips held up on a small folder of files

What works when you direct Claude to do real work: the apps and plans, the instruction file it reads at the start of every session, five habits each tested plainly and carefully on the same made-up folder, the tips that made no difference, and where your files go. Last checked November 2026.
Author

Behzad Samadi

Published

November 11, 2026

A paper-craft garden committee table: an open laptop with a chat window, a folder of meeting notes being sorted into a dated stack, a cork board with a card of house rules and an amber tick, seedlings, gloves, a mug, and the five-dot AI glyph above. Title: Working with Claude, Tested.

Last checked 11 November 2026, on Claude Opus 5.5, in Claude Code 2.1.322.

This guide is for someone who uses Claude to get things built or sorted out, not to write code themselves: a committee’s paperwork, a spreadsheet, a small app. It covers the assistant as a whole, the chat in a browser, the desktop app and Claude Code, the version that works on the files in a folder.

The practices in this guide were tried on the same small folder of made-up files: a garden club’s seven sets of committee minutes, with untidy names and one misleading date, and a members list with a fee column that has four traps in it. Each task was asked twice, once plainly and once with the habit being tested, three times each, in a fresh copy of the folder for each run. The counts are small, so the guide says “in three runs out of three”, not “always”. How the tests were set up is at the start of Practices that held up.

If you have not yet let an assistant work on a folder, start with Let It Open the Folder, which sets up Claude Desktop and explains why a folder beats pasting into a chat.

Other guides in this series: Working with ChatGPT, Tested and Working with Grok, Tested.

NoteIn short
  1. Claude is good at reading a pile of documents and getting the facts right. On the plainest questions it found every open action, the right fee total, the misdated file and the figure that was not there.
  2. Do by hand: the final read for your house style, unless the rules are written down. Plain requests broke two or three of six simple writing rules. Nothing in these tests called for another assistant.
  3. The habit that made the most difference is the instruction file: write your rules once in a file called CLAUDE.md in the folder. With it, every rule held in three runs out of three. Correcting it once and asking it to update that file worked as well as writing the file yourself.
  4. To protect your files, give it a copy of the folder and ask it to show its check. The work was right either way. Asked, it proved with a checksum that no text had changed; unasked, it said “I didn’t change what’s written” and gave no proof. In the tests it also edited a file nobody mentioned, and once wrote outside the folder it was given.

Which one you are using

Claude comes in three forms, all sold on the same subscription. They share the model but not everything else. Which plans include which is in The moving parts.

What it is Works on your files
Chat claude.ai in a browser, or the phone app Files you upload; it can also create files and run code
Desktop app The same chat on Mac, Windows or Linux, plus a Code tab and Cowork Cowork runs its sessions in the cloud, with your files kept in your Claude account; on the desktop you can also connect a folder on your computer
Claude Code The command-line version, typed in a terminal, also inside the desktop app as the Code tab The folder you start it in

Anthropic is merging chat and Cowork into one “Claude”, so the buttons you see depend on when your account was moved.

This guide was tested in Claude Code, because a script can run it the same way thirty times and keep what it wrote. The desktop app’s Code tab runs Claude Code, so what holds here should hold there; that was not tested separately. The chat and Cowork were not tested; where a practice depends on them, the guide says so and gives Anthropic’s own page instead. For running Claude Code beside other coding agents, see One Repository, Three Agents.

Giving it context

Claude starts each conversation knowing only what you give it and what its memory has kept. There are four ways to give it more, and they are not the same thing.

Files and folders. In the chat you upload them; in the desktop app or Claude Code you give it the folder, and it opens what it needs.

Projects (chat and desktop app). A workspace with its own files and its own instructions, used by every chat inside it; free accounts have a limit (see The moving parts). Anthropic’s page warns that chats inside a project do not see each other: “Context is not shared across chats within a project unless the information is added into the project knowledge base.” A new kind of project, one long conversation that splits work into parallel threads, is in beta for some paid accounts.

Memory (chat and desktop app). Claude saves short notes about you as you chat (“Mention that a deadline moved, and your next conversation already knows”). It is on by default on the personal plans. You can read and edit every note under Settings > Your Claude, pause it, or reset it. Each project has its own memory. Two things to know: deleting a chat does not delete what was remembered from it, and an incognito chat is the one switch that does two things at once: the chat is neither remembered nor used for training.

The instruction file (Claude Code, and so the desktop app’s Code tab). A plain text file called CLAUDE.md in the folder, which Claude Code reads at the start of every session. Whether Cowork reads it was not tested. Write in it what you would otherwise repeat. If the folder already has an AGENTS.md, the file other assistants read, Claude Code reads that instead when there is no CLAUDE.md. Anthropic’s advice is to keep it under 200 lines and concrete (“Use 2-space indentation” rather than “Format code properly”), and to remember it is advice to the model, “not enforced configuration”.

What the instruction file did

The task: Write the notice to members about the work day, from the committee’s notes. Save it as notice.md. The club has six house rules for anything sent to members:

- Dates in full, with the weekday and the year: Saturday 14 November 2026.
- Times on the 24-hour clock: 09:30, not 9:30 am.
- British spelling.
- No exclamation marks.
- Under 120 words, heading and signature included.
- Signed "The committee", with no personal names.

Run three times without the rules, and three times with them saved as CLAUDE.md in the folder. Marked out of 10: four facts (the date, the times, gloves, soup) and the six rules.

Run 1 Run 2 Run 3
No instruction file 7 8 7
Rules in CLAUDE.md 10 10 10

Every notice had the facts right. Without the file, every notice ran over 120 words (132 to 138), each was signed with the garden’s full name, one with a committee member’s name (“Ibrahim Solheim, for the … committee”), and two dropped the year from the date. It already wrote 24-hour times and British spelling and left out exclamation marks without being told. With the file, every rule held, in 80 to 99 words.

In an earlier round, on Claude Fable 5.2 and an earlier version of the folder, the same rules saved as AGENTS.md instead of CLAUDE.md were followed in three runs out of three.

Practices that held up

How the tests were run. On 11 November 2026, in Claude Code 2.1.322. Model: claude-opus-5-5 asked for on the command line, and each turn reported claude-opus-5-5. Effort: not set, so the tester’s own setting applied, which is higher than the default. Permission: --permission-mode auto, with no sandbox; no command was refused. Connectors and MCP servers were kept out with --strict-mcp-config with no server list, and --no-chrome. Web search: built in and left on; no reply cited the web. Instruction files: the tester’s own global instruction file, settings and hooks loaded, as they would for anyone. Each prompt was run three times plainly and three times with the practice, each in a fresh copy of the folder.

Brief it, do not just ask it

Plain: What do we still need to do?

With the practice: Goal: a list of the committee’s open actions, so nothing is forgotten before the garden closes for winter. Read every file in notes/. An action is something a named person, or the committee, agreed to do. It is open unless a later note says it was done. If an action was carried over, list it once, with its latest date. Write actions.md with one table: action, who, due date, the note it was last mentioned in. Under the table, list anything you were unsure about and why. Do not change any other file.

Marks out of 9 Run 1 Run 2 Run 3
Plain 9 9 9
Brief 9 9 9

Both found the five open actions in three runs out of three each, including the one whose owner changed from Nell to Dov partway through. The difference was in what came back. The plain question gave a page that mixed the actions with fee problems, a doubt about members’ rates and things nobody had agreed to do (closing the garden in all three; once, arranging the soup). The brief gave exactly the five in a file, the finished ones listed apart, and the doubtful cases (who pays the roofer once the $80 is explained; who finalises the checklist) in a separate list. On this model the brief saves you reading, not mistakes.

Plan first when it is going to change files

Ask for the plan, read it, then say go.

With the practice, first message: I want the notes folder tidied: every note named for the date of its meeting, as YYYY-MM-DD-committee-meeting.md, and a notes/index.md that lists each note with one line on what was decided. The text of the notes must not change. Do not change anything yet. Tell me your plan: what you will rename to what, how you will work out each date, and how you will check afterwards that no text was lost. Then stop.

Second message, same conversation: Go ahead with that plan. When you finish, tell me what you checked.

The plain version was Tidy up the notes folder: consistent file names, and an index. One note is a trap: the file called minutes 25 aug.md records a meeting moved to 26 August.

Run 1 Run 2 Run 3 Said how it knew nothing was lost
Plain passed passed passed no, in all three
Plan first passed passed passed yes, in all three

All six runs did the job and all six caught the 26 August meeting. (The plain runs pass the test’s check for the plain prompt, which accepts any file name that starts with the right date and any name for the index. Its stricter check, which wants the names only the careful prompt gives, fails them on the file or index names alone.) The plain runs said “I didn’t change what’s written in any of the minutes” or the like, without saying how they knew. The planned runs replied after the first message that they had changed nothing (the folder was not checked between the two messages, so that rests on their word), flagged the 25/26 August note in the plan, and afterwards gave a checksum, a fingerprint of each note’s exact contents, before and after.

Outside what was asked. All three plain runs also edited the folder’s README.md, which nobody asked for. One planned run kept its checksums in a check/ folder one level above the folder it was given, and said so. Give it a copy of the folder, not the original.

Ask for the evidence, not just the answer

When the answer is a number, ask what could make it wrong and how it checked.

Plain: How much have we collected in fees this year?

With the practice: How much have we collected in fees this year? Before you answer, look through members.csv for anything that would make a simple sum wrong: duplicates, blanks, refunds, odd formats. List each one you find and say how you treated it. Then give the total, the number of members, and how you checked the total.

The right answer is $855.00; a plain sum of the column gives $900.00.

Marks out of 5 Run 1 Run 2 Run 3
Plain 5 5 5
With the practice 5 5 5

The answer was right in all six runs. Every plain reply found the duplicate member, the blank fee and the refund, and pointed out that the plain sum of $900 only looks right because the duplicate and the refund cancel out. What the careful prompt added was a second check: in all three runs a short script that grouped the rows by email, and a hand count (“19 × $45 + 2 × $22.50 = $900.00, and $900 − $45 = $855.00”), and a member count with its reasons (22 listed, and 20 or 21 paid, depending on how the refunded member was counted, each explained). On this model the practice does not make the sum more right; it gives you the means to see that it is.

Say “only from these files”

Plain: What did the club pay for water in 2025?

With the practice: Answer only from the files in this folder. If the answer is not in them, say so, and say what the files do contain on the subject. What did the club pay for water in 2025?

The figure is not in the files. The minutes give 2026 only, and one gives $96.20 for electricity in 2025, a near miss. Marked right if it says the 2025 water figure is not there, wrong if it gives any amount.

Run 1 Run 2 Run 3
Plain right right right
With the practice right right right

All six runs said the 2025 water figure is not there, gave the 2026 figures, and named the $96.20 as electricity. None made up a number. With no connectors, the plain question was as safe as the careful one. In the earlier round, with the tester’s account’s connectors switched on, one plain run went looking in a connected service from that account; none of the “only from these files” runs did.

Correct it once, and have it write the rule down

When Claude gets something wrong, tell it what was wrong and ask it to record the rule in CLAUDE.md. The next session reads it.

The test, three times, in a folder with no instruction file: the notice task as above; then, in the same conversation, “That notice breaks our house rules … [the six rules] … Fix notice.md, then update your CLAUDE.md so you don’t make that mistake again.”; then the notice task once more in a new conversation, with the old notice deleted.

Run 1 Run 2 Run 3
First notice 8 8 7
After the correction 10 10 10
New conversation, reading the file it wrote 10 10 10

In all three runs the file it wrote said more than the rules it was given. For example, run 1 added what not to do (“Never write ‘14 Nov’ … or a date without its weekday and year”) and a way to check the length (“Count with wc -w before saving”); run 2, whose first notice had named Wren, turned that mistake into an example (“Say ‘the water will be turned off’, not ‘Wren will turn off the water’”). The same test passed in three runs out of three on Claude Sonnet 5.5 in an earlier round.

What did not hold up

These leads were tested in the earlier round, on the first version of the folder, not on Opus 5.5; the model is given for each.

Capitals and IMPORTANT (Fable 5.2). A common habit is to write rules as “IMPORTANT: YOU MUST…” and “NEVER, EVER…”. The same six house rules written that way scored 10, 10 and 10, exactly as the plain wording had. Anthropic’s own Claude Code guide now lists the habit under “What you can stop doing”: current models “follow a plainly worded line, and piling on emphasis mostly makes Claude over-apply the rule.”

“Claude ignores AGENTS.md” (Fable 5.2). True of older versions of Claude Code. Here the house rules saved as AGENTS.md were followed in three runs out of three, and each reply named the file. Anthropic’s page says Claude Code now reads it when there is no CLAUDE.md (from which version: The moving parts).

“ultrathink” or “think harder” (Sonnet 5.5): inconclusive. Adding ultrathink to the plain fees question scored the same in three runs out of three as the plain question had on Fable 5.2; but these runs were on another model at the tester’s own effort setting, with no plain run on that model to compare. Anthropic’s guide says phrases like “think hard” are “ordinary prompt text”, that the control for how hard it thinks is the effort setting, and that “ultrathink still works: it asks for deeper reasoning on that single turn”.

“Put your project structure in CLAUDE.md.” Often repeated; not tested here. Anthropic advises against it: “Claude reads the repo itself. Only write down what isn’t in there.” Put in the file what Claude cannot see: the house rules, the decision made at a meeting, the check that must pass.

“Claude ignores CLAUDE.md.” Many reports say so. On these tasks the rules in an instruction file held in twelve runs out of twelve in the earlier round and six out of six in this one. Anthropic’s own position is that the file is advice the model reads, not a lock; for a rule that must never be broken (never delete originals, say), it points to settings that the app enforces, called hooks and deny rules.

What the tests do not show

  • The chat, the desktop app and Cowork were not tested. Projects, memory and the apps’ own limits come from Anthropic’s pages.
  • One model, at the tester’s effort setting. The marks above are for Claude Opus 5.5 in Claude Code, with the tester’s own instruction file loaded and no connectors. No effort level was passed, so the tester’s own effort setting applied, which is higher than the default; a new subscriber on the default setting may see different results. The earlier round, on Fable 5.2 and Sonnet 5.5 with an easier folder, gave the same picture.
  • Web search and web fetch were still available. They are built into Claude Code, not connectors, and the no-connectors setting leaves them on. No answer cited the web.
  • The tasks were easy for this model. Plain and careful scored the same on four of five tasks; task 2, the instruction file, differed by two or three marks a run. The practices showed up mostly in what Claude told you about its work and in the house style, not in whether the facts were right. A harder task, or a smaller model, might separate them more.
  • Three runs each. Enough to see that something happened in three runs out of three, not to say how often it fails.

Where your data goes

Everything Claude reads is sent to Anthropic to be processed, whether you upload it in the chat or let Claude Code open it on your computer. Claude Code’s documentation is plain about it: it sends “all user prompts and model outputs”.

Training. On the personal plans, Anthropic uses your chats and coding sessions to train future models when the model improvement setting is on. The pricing page calls this “Opt-out”; Anthropic’s pages do not say in words whether the setting starts on for a new account, so check yours at claude.ai/settings/data-privacy-controls. The business plans are not used for training by default. Incognito chats are never used for training, whatever the setting.

How long it is kept. Longer if the model improvement setting is on than if it is off. A chat you delete leaves your history at once and Anthropic’s systems later. Chats flagged by Anthropic’s safety systems, and conversations you rate with the thumbs up or down, are kept longer. The periods are in The moving parts.

On your own computer. Claude Code keeps a plain-text transcript of each session in a hidden folder in your home directory (~/.claude/projects/), so you can resume it, and deletes it after a while. Its own memory notes are there too, as files you can read and delete.

Settings worth changing:

  1. The model improvement switch, if you do not want your work used for training.
  2. Memory: read what it has saved (Settings > Your Claude > Manage), and use an incognito chat for anything you would not want remembered.
  3. Connectors: turn off the ones you do not need while working on private files, since Claude may look in them.
  4. In Claude Code or the desktop app, keep the folder small: give it a copy of what it needs, not your whole Documents folder.

The moving parts

As read on Anthropic’s pages on 10 and 11 November 2026. Prices are as the page shows them, in US dollars; Anthropic shows no Canadian price, and tax is extra.

Free Pro Max Ultra
Price a month US$0 US$20, or US$17 paid yearly (US$200) US$100 US$200
Usage Limited 1× (the base) 5× Pro 10× Pro a week; no 5-hour limit (verified accounts)
Claude Code, Cowork No Yes Yes Yes
Projects Up to 5 Yes Yes Yes
Fable No Paid with usage credits Up to half the weekly limit Up to half the weekly limit

Team is US$25 a seat a month (US$20 paid yearly); Enterprise is US$20 a seat plus usage. Chat is on every plan; Claude Code and Cowork in the desktop app on the paid ones.

  • Models: Claude Fable 5.2 (added to Claude Code on 9 November 2026), Opus 5.5, Sonnet 5.5 and Haiku 5.5. Claude Code starts new subscribers on Opus 5.5 at medium effort; Fable is chosen with /model fable.
  • Limits: a five-hour limit and a weekly limit, shown as bars under Settings > Usage. No page gives them as a number of messages.
  • Claude Code 2.1.322, 10 November 2026. It reads AGENTS.md from 2.1.277; Fable 5.2 needs 2.1.321 or later.
  • Retention on the personal plans: 30 days with model improvement off, up to 5 years with it on (“in a de-identified format”); a deleted chat leaves Anthropic’s systems within 30 days; chats flagged by the safety systems up to 2 years; conversations rated with the thumbs up or down up to 5 years. Claude Code’s local transcripts: 30 days by default.
  • What the tests used: the thirty runs on Opus 5.5 came to about US$7 at API list prices, about US$0.22 a run; the earlier round on Fable 5.2 cost about US$1 a run. That is not charged on a subscription, but it comes out of the plan’s limit: the earlier round stopped after three runs on “You’ve hit your 5-hour limit”.

What changed

  • 12 November 2026. Made consistent with the ChatGPT and Grok guides: the practices in one order, the task 3 table and task 5 table in one shape, one paragraph on how the tests were run, and what the runs touched outside the request.
  • 11 November 2026. Corrected after a check: the effort setting used in the tests is stated; the task 3, task 4 and lead quotes and counts are fixed; the AGENTS.md and ultrathink leads carry the model they ran on; volatile facts moved to The moving parts; the training default is no longer implied.
  • 11 November 2026. Tests run again on Claude Opus 5.5, with no connectors, on a harder version of the folder; marks and claims now rest on those runs. The first draft had been tested on Fable 5.2 and Sonnet 5.5.
  • 11 November 2026. First published.

Sources

Anthropic’s pages, read on 10 and 11 November 2026. The research notes, the test log with every mark, and each run’s output are in the accompanying folder. The made-up folder, prompts and answer key shared by all the assistant guides are in the common folder.