Working with Grok, Tested

A guide to xAI’s assistant for people who direct it rather than program: grok.com, the phone apps, Grok in X and Grok Build, how to give it rules, and which popular tips held up on a small folder of files

What works when you direct Grok to do real work: the apps and the new subscription, the instruction file Grok Build reads, five habits each tested plainly and carefully on the same made-up folder, the tips that failed, and the three switches that decide whether your words train its models. Last checked November 2026.
Author

Behzad Samadi

Published

November 11, 2026

A paper-craft garden committee table: an open laptop listing files in order, a magnifying glass over a desk calendar with one day circled, meeting notes fanned in date order, a trowel, herbs, a mug, and the five-dot AI glyph above. Title: Working with Grok, Tested.

Last checked 11 November 2026, on Grok 4.7 (reported as grok-4.7-build), in Grok Build 1.0.70.

This guide is for someone who uses an assistant to get things built or sorted out, not to write code themselves. It covers Grok as a whole: the chat at grok.com and in the phone apps, Grok inside X, and Grok Build, the version that works on the files in a folder.

Grok is made by xAI. Its pages now give the company’s name as SpaceX Super Intelligence LLC, “SpaceXSI”; this guide calls it xAI. What sets Grok apart from the other assistants is its tie to X: it can search public posts on X, it comes with some X plans, and what you say to it inside X falls under X’s rules, not xAI’s.

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.

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

NoteIn short
  1. Grok 4.7 got the facts right on every task, asked plainly or carefully: the open actions, the right fee total with its traps, the meeting held a day late, and a figure that was not there. Thirty runs of the five tasks, no wrong answer.
  2. Do by hand: the check of what it hands you. What the careful prompts changed was what it showed you, not whether it was right: a file instead of a chat reply, a list of what it was unsure about, a check you can repeat. Ask for those when you need to rely on the result. Nothing in these tests called for another assistant.
  3. Ask for a plan, and ask how it will check. Asked in words, with the names wanted and a check, it planned, stopped, then proved with checksums that no text was lost, in three runs out of three. The plain request said how it checked in none of three; started in plan mode, it did in two of three.
  4. Do not count on a .grokignore file to keep a file private. It read the file it was told to ignore in three runs out of three. Give it a copy of only what it needs, and check the three training switches in Where your data goes.

Which one you are using

What it is Works on your files Comes with
grok.com and the phone apps The chat, in a browser or on iPhone and Android Files you upload (limits in The moving parts) Free, with limits; more on paid plans (see The moving parts)
Grok in X The Grok button in the X app and on x.com Posts and what you type X accounts; governed by X’s terms, not xAI’s
Grok Build A coding agent you run in a terminal, in a folder, with grok The folder you start it in Paid plans (see The moving parts)

x.ai’s download menu lists only iPhone and Android apps; on a computer, the chat is grok.com. (Grok Build’s own guide mentions a “Grok Desktop” that this guide has not seen.) There is also a “Build mode” on grok.com and Grok Bot, an agent that works on a computer in xAI’s cloud. Neither is covered here.

Check which grok you have. Typing grok can start one of two programs: xAI’s Grok Build, installed from x.ai/cli/install.sh, or an unrelated open-source tool that uses the same name. grok --help tells you: xAI’s starts with “Grok Build”. This guide means xAI’s.

This guide was tested in Grok Build, because a script can run it the same way thirty times and keep what it wrote. The chat was not tested; where something depends on the chat, the guide gives xAI’s own page instead. For Grok Build beside other coding agents, see One Repository, Three Agents.

Giving it context

Files. In the chat, attach them with +. Grok Build opens what it needs in the folder you start it in; you can also point at a file in your message with @, as in @notes/aug11.md.

Projects, memory and custom instructions in the chat: xAI’s help pages read on 11 November 2026 do not describe them. Its FAQ mentions Projects once, in passing. Third-party pages disagree about what exists. Until this guide has checked in the app, it says nothing more.

Memory in Grok Build is “experimental and disabled by default”. When switched on, it keeps notes as ordinary text files on your computer, for all your work and for each project.

The instruction file (Grok Build). A plain text file called AGENTS.md in the folder, read at the start of every session. xAI’s description: “Put coding conventions, build and test commands, and architecture notes in an AGENTS.md at your repo root, and Grok follows them without being told each session.” It also reads CLAUDE.md, AGENT.md and a few other names, and every file in a .grok/rules/ folder. There is no length limit, but “short, specific instructions are followed more reliably than long ones.”

Two things the documentation says, both seen in the tests:

  • It reads Claude Code’s files too. Unless you switch it off, Grok Build also loads the instruction files, skills, permission rules and MCP servers (connections to other services) that Claude Code uses on the same computer. On a computer that also has Claude Code set up, grok inspect listed Claude’s instruction file, skills and connections before any Grok file had been written. They were switched off for the tests, except Claude Code’s permission rules, which have no off switch.
  • It ignores the file in a folder it does not trust. In a new folder, grok inspect did not list AGENTS.md, and Grok, asked to quote its project rules, answered “None.” Once the folder was trusted, it quoted them. xAI’s guide says you trust a folder by answering yes when the app asks; then run grok inspect to see the file listed.

grok inspect, typed in the folder, lists every instruction file and connection Grok found. Run it once in any folder you care about.

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 AGENTS.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 10 10 10
Rules in AGENTS.md 10 10 10

Grok kept all six rules without being told them, in three runs out of three: the full date with the year, 24-hour times, no exclamation marks, no American spellings (though no word in these notices would have tested it), a short notice, signed “The committee”. With the file the notices were shorter, 46 to 50 words against 63 to 83, and dropped the extras (“Dear members”, “We hope to see you there”). The same rules saved as CLAUDE.md instead also scored 10, 10 and 10.

So on this model the instruction file did not make the notice more correct; the rules it was given were ones it already kept. Its use is for rules Grok would not guess, such as your club’s name for itself, or which committee member must never be named.

Practices that held up

How the tests were run. On 11 November 2026, in Grok Build 1.0.70. Model: grok-4.7 asked for on the command line, and each run reported grok-4.7-build. Effort: not set. Permission: bypassPermissions, because a scripted run cannot answer a question, with --sandbox workspace, which keeps its writes to the folder (one run still used the system’s temporary folder). Connectors and MCP servers were kept out with environment variables that stop Grok Build reading Claude Code’s and Cursor’s set-ups, --deny MCPTool, and the MCP tools removed with --disallowed-tools. Web search: off (--disable-web-search); Grok’s own X search tools stayed available, and no reply used them. Instruction files: the tester’s Claude Code instruction file and skills were kept out, but Claude Code’s permission rules, which have no off switch, still loaded; the folder-trust check was turned off (GROK_FOLDER_TRUST=0) so that AGENTS.md loaded. Each prompt was run three times plainly and three times with the practice, each in a fresh copy of the folder.

In each case below the answer was right both ways; what held up is what the practice added.

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

The plain question found all five open actions, with the right people and dates, in three runs out of three, in a chat reply. It also added things the notes do not record as agreed actions: all three listed a member’s unpaid fee beside the actions, and one counted the work day itself as a sixth open item. The brief gave the same five in a file you can keep, with a list of honest doubts under it: whether holding the roof payment is its own action, whether Tessa owes comments on her own checklist draft. It took about three times as long.

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.

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

Both did the job. The plain runs named every note by its date (2026-08-26.md), kept the text, wrote an index, and spotted that the file called “25 aug” was the meeting held on the 26th. They said the text was “unchanged” without saying how they knew. (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 planned runs stopped as asked with a table of old and new names, then afterwards gave a fingerprint (a SHA-256 checksum) and the size in bytes of every note before and after. For a folder you care about, that is the difference between trusting it and knowing.

Outside what was asked. One planned run copied the notes to the system’s temporary folder for its check, and deleted the copies afterwards. One of the plan-mode runs below also edited the folder’s README.md, as its plan had said. Give it a copy of the folder, not the original.

Ask for the evidence, not just the answer

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: the total, and each of the duplicate member, the blank fee and the refund named.

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

The plain question already got $855.00 and named all three traps, in three runs out of three, and two of its replies showed the breakdown (18 full fees and 2 half-season fees). The careful one added what it asked for: 22 members, explained, and two or three independent checks (“18 full fees × 45.00 = 810.00, plus 2 × 22.50 = 45.00”). It took about twice as long.

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, plain and careful, said the 2025 figure is not there, gave the 2026 ones, and offered no number for 2025. Four of the six pointed out the electricity bill as the only 2025 figure, and said it was not water. The careful wording made no difference here.

Start in plan mode (mixed)

xAI’s launch post recommends it: “For complex tasks, start Grok Build in plan mode.” Tested with the plain tidying request in plan mode, then “Go ahead with that plan”, three times. All three renamed the notes by date, kept the text and passed the plain-prompt check; two of three said how they checked the text, against none of three without plan mode. That part held. But each did what its own plan said, not necessarily what you would want: two named the index notes/README.md, and one also changed the folder’s own README.md. Both were in the plan it had been told to go ahead with. Read the plan before you say go; the version asked for in words, with names and a check, did all of that in three runs out of three.

What did not hold up

.grokignore. A community best-practice list says a .grokignore file keeps secrets and other files away from Grok. No such file appears in xAI’s documentation. With members.csv listed in it, Grok read the file and gave the fee total in three runs out of three. To keep a file away from it, leave it out of the folder.

“Ask for short answers in AGENTS.md” (mixed). A tip from a forum: Grok is wordy, so say in the instruction file how long a reply should be. With Keep every reply to ten lines or fewer in AGENTS.md, the plain “What do we still need to do?” came back in eight or nine lines, 123 to 152 words, against 215 to 266 without it. The length held in three runs out of three. But each short answer lost a mark, 8, 8 and 8 against 9, 9 and 9: two left out who owes the checklist comments, one listed a refund already made as still to do. Short is fine for a reminder; for a list you will act on, ask for the file instead.

“Tell Grok the rules in each prompt.” Not needed on this model for the house-style rules tested here; see What the instruction file did.

What the tests do not show

  • The chat was not tested. Projects, memory, custom instructions and plan limits are what xAI’s pages say, or are marked as not found.
  • One model, one date. Grok 4.7 in Grok Build on 11 November 2026. The other models the app offers were not tried.
  • The tasks were easy for this model. Plain and careful scored the same on five of five tasks; no task differed. The practices changed what Grok showed you about its work, not whether the work was right. A harder task or a smaller model would separate them more.
  • A cleaner set-up than yours may be. Claude Code’s files and connections, which Grok reads by default, were switched off, except its permission rules, which have no off switch; all MCP connections were blocked. If yours are on, Grok may behave differently.
  • 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

What you type and upload goes to xAI to be processed. Grok Build’s model also runs on xAI’s servers, so what it reads in your folder goes there too, unless you set it up with a model of your own; xAI’s pages do not say this in so many words. What happens next depends on where you used Grok.

grok.com and the phone apps. xAI “may use your content and interactions with Grok … to train our models”, and “You control whether your data is used for training Grok”. The switch is Settings > Data > Improve the Model on grok.com, and Settings > Data Controls > Improve the model in the phone app. Switching it off covers your new chats. Three things it does not cover:

  • Rating a reply with thumbs up or down: “that feedback may be used for training purposes” even with the switch off.
  • Using Grok without signing in: in some regions outside the EU and UK, there is then no way to opt out.
  • Build mode on grok.com: build sessions are kept so you can carry on, and Private Chat is not available there.

Private Chat (the ghost icon at the top right) is not saved to your history, not used for training, and deleted after a short time. A chat you delete, or your whole account, also takes a while to delete, unless kept for safety, security or legal reasons. The periods are in The moving parts. xAI says a limited number of its staff “may review your conversations” for purposes that include improving the model. Business and enterprise accounts are not used for training.

Grok in X is X’s, not xAI’s. X “may share with xAI your public X data as well as your user interactions, inputs and results with Grok on X to train and fine-tune Grok”. Public data means your public posts and their metadata, public Spaces and your public profile. The switch is in X’s settings: Privacy & Safety > Data sharing and personalization > Grok & Third-party Collaborators > Data Sharing. Protecting your posts (making the account private) also keeps them out. Adding @Grok to an encrypted X chat gives X and xAI every message sent after it joined.

Grok Build. When you first use it, it asks whether xAI may keep your data “to improve the product and model”; type /privacy to change the answer later (“Coding data, retention, and training: Opt in or Opt out”). Its transcripts are also kept on your own computer, in a hidden .grok folder in your home folder. Usage statistics are a separate switch.

Settings worth changing:

  1. Improve the Model on grok.com and in the phone app, if you do not want your chats used for training.
  2. X’s Data Sharing switch, if you use Grok inside X. Grok Personalization, a separate switch beside it, is worth a look too.
  3. /privacy in Grok Build.
  4. For anything private, use Private Chat in the chat, and do not rate replies.
  5. In Grok Build, give it a copy of the files it needs, not your whole Documents folder, and run grok inspect to see what else it loaded.

The moving parts

As read on xAI’s pages on 11 November 2026. Prices are quoted as x.ai shows them, with “$” and no currency named: the currency is not stated. They are before tax, and no Canadian price is shown.

The X Subscription, xAI’s single plan for Grok, Cursor and X. Older SuperGrok and X Premium plans “continue unless you change or cancel them”; switching is optional and cannot be undone.

Free Plus Premium Super Ultra
A month, “$” (currency not stated) $0 $8 $30 $100 \(200 | | A year, "\)” (currency not stated)
Usage Limited Base 4× Plus 15× Plus 50× Plus
Grok on the plan’s page Limited chat Chat, Voice, Imagine and Build adds Grok Bot, latest models higher usage (not listed)
X benefits Standard X Checkmark, fewer ads X Premium+ X Premium+ X Premium+

Teams: from $30 (currency not stated) a seat a month.

  • Grok Build on the plans: paid plans from Plus. The pricing table also shows a “Limited” Build entry whose column was not clear.
  • Uploads in the chat: up to about 100 files at once on the web, 150 MB each.
  • Usage is one weekly pool shared by every Grok product and Cursor; the share each product used is shown under Settings > Usage. When it runs out, paid features pause until the reset, or you buy credits (from $5, currency not stated, web only).
  • Grok Build was launched on 25 May 2026 as “an early beta for all SuperGrok and X Premium Plus subscribers”. Version 1.0.70 was the one installed for this guide.
  • Models. Signed in, Grok Build 1.0.70 offered grok-4.7 (the default), grok-4.7-build-fast, grok-4.6 and grok-4.5; the tests asked for grok-4.7 and each run reported grok-4.7-build. xAI’s documentation calls grok-4.7 its “latest model”, and its example settings suggest a model called grok-build for coding sessions.
  • Retention: Private Chat deleted within 30 days; a deleted chat or account “up to 30 days”.
  • What the tests used: the thirty runs of the common tasks came to US$4.12 at the API prices Grok Build reports, about US$0.14 a run; the twelve lead runs to US$2.07. On a subscription that is not charged, but it comes out of the weekly usage pool.

What changed

  • 12 November 2026. Made consistent with the Claude and ChatGPT guides: the practices in one order, the plan-mode and short-answer leads placed and labelled as mixed, the task 5 table, one paragraph on how the tests were run, what the runs touched outside the request, and every price labelled; plan availability and retention periods moved to The moving parts.
  • 11 November 2026. Corrected after a check: task 3’s plain runs and the plan-mode lead re-marked under version 2.1 of the answer key; counts in place of absolute claims; Claude Code’s permission rules noted as still loading; upload limits and the model list moved to The moving parts.
  • 11 November 2026. First published.

Sources

xAI’s and X’s pages, read on 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.