GROK BOT / 9 TRICKS · ONE PASTE-IN PROMPT

Paste This Into Grok Bot.
Never Run Out of Credits Again

$550 on credits in one weekend. Here's how to cut it in half.

Usage limit exceeded

Weekly usage100% used

ILLUSTRATION

THE FIX9 TRICKS

010203040506070809

THE PROOF / OCT 31 TO NOV 1 · MY CREDIT CHARGES

The receipts

22 automatic credit top-ups of $25, $550 total

THE SURPRISE / WHERE THE CREDITS ACTUALLY GO

Most of your credits go to re-reading.

MESSAGE 20 OF A JOB: WHAT YOU TYPE VS WHAT GETS RE-READ

What you typed1x

What the bot re-reads~800x

ESTIMATE · REAL NUMBERS VARY

ONE USER'S REAL EXPORT78%

of one user's Grok Bot usage was re-reads

@GrokBotRadar on X · his own export (965M tokens)

COLD OPEN. I thought the credits went on the work. Most of your credits go to re-reading. What you type is a tiny fraction. Every message re-sends the bot's whole conversation, and by message 20 the bot re-reads about 800x what you typed (estimate). HOW MY ESTIMATES WORK: I measure in credit units. 1 unit is roughly 4 characters of text the bot has to read (what AI models call a token). Every estimate in this deck is shown as a multiple (1x vs 10x), because the real credit price per unit isn't published. Chart assumptions: 20 messages where each adds ~2,000 units, so message 20 re-reads ~40,000 units. A typed line is ~200 characters, about 50 units. 40,000 / 50 = 800x. Estimate. Sourced: @GrokBotRadar exported his usage: 965 million since Oct 1, and 78% was re-reading (cache reads). Careful: his share is measured in tokens, and he notes token share isn't exactly cost share. x.com/GrokBotRadar/status/2115489802682585477

STEP ZERO / MEASURE BEFORE YOU FIX

First, find your leak

EXPORT YOUR USAGE

grok.com/dashboard›Usage›Export CSV›Give it to your Chief of Staff bot

THEN PASTE

“Check this usage file and analyze my bot behaviours and tell me how we should optimize for usage.”

CLAIMED SAVING10 to 15%

a week

Claimed by @tejas3732 · 56K views · 505 likes

THE LIST / 9 BEGINNER MISTAKES I WISH SOMEBODY TOLD ME ABOUT

9 tricks, half the credits

01

One bot per kind of job

02

Batch your fixes

03

Brief it once, with a done list

04

Text, not screenshots or PDFs

05

Big outputs to files

06

Stop it browsing

07

Reset a bloated bot

08

Script it, reuse research

09

Stop bots chatting

01 / TRICK 01 OF 09

One bot per kind of job

01

~4xless history re-read over 20 messages

BEGINNER MISTAKE

One do-everything bot for decks, emails and research

SIMPLE FIX

One bot per kind of job. Unrelated work? Give it its own bot.

CREDITS USED OVER 20 MESSAGES (FOCUSED BOT TOTAL = 1x)

Do-everything botFocused bot, same work

ESTIMATE · REAL NUMBERS VARY

TRICK 1. What people do: One do-everything bot for decks, emails and research. Why it burns credits: In Grok Bot each bot is one long, persistent conversation, and every message re-reads that history. A do-everything bot drags its old decks, email threads and research into every message. A focused bot only re-reads history that's relevant to the job. The number: Over 20 messages, a do-everything bot uses roughly 4x the credits of a focused bot doing the same work (estimate). Assumptions: Assumes each message plus reply adds ~2,000 credit units (about 8,000 characters), so message n re-reads ~2,000 x n of its own job: 2,000 x (1+2+...+20) = 420,000 for the focused bot. The do-everything bot also carries ~60,000 units of unrelated history (old decks, emails, research) on every message: 20 x 60,000 = 1,200,000 extra, 1,620,000 total, about 3.9x. Conservative and illustrative; real numbers vary with how the app trims or summarizes old history. The fix: One bot per kind of job. Unrelated work? Give it its own bot. Demo: Show your bot list with one bot per kind of job, then ask your busiest bot: "Which of the jobs you do should really be its own bot?"

02 / TRICK 02 OF 09

Batch your fixes. One message.

02

10xless re-reading

BEGINNER MISTAKE

Sending 10 small fixes one at a time

SIMPLE FIX

Write every fix in one list, send it once

CREDITS FOR 10 FIXES IN A LONG CONVERSATION (BATCHED = 1x)

10 fixes, one at a time~10x

1 batched list1x

ESTIMATE · REAL NUMBERS VARY

03 / TRICK 03 OF 09

Brief it once, with a done list

03

5xfewer credits per job

BEGINNER MISTAKE

Vague ask, then “double check your work”

SIMPLE FIX

Goal, why, what to avoid, and a done list. Pass, fail or unknown, then one fix pass.

Skip the ALL CAPS “you MUST”. Say it once, calmly.

CREDITS FOR ONE JOB, INCLUDING REVISION ROUNDS (BRIEF = 1x)

Vague prompt + 6 revision rounds~5x

Brief + done list + 1 round1x

ESTIMATE · REAL NUMBERS VARY

TRICK 3. What people do: Vague ask, then “double check your work”. Why it burns credits: A vague ask gets a guess, and "double check your work" is open-ended, so the bot keeps re-reading and re-checking. Every round of "no, not like that" re-reads the growing conversation and rebuilds the work. A done list turns checking into one pass: each item is pass, fail or unknown, and only the fails get fixed. The number: A vague prompt plus 6 revision rounds uses about 5x the credits of a full brief with a done list plus 1 correction round (estimate). Assumptions: Assumes the first build sits on ~20,000 credit units of history and each round adds ~5,000. Vague: 7 turns at 20K, 25K, 30K, 35K, 40K, 45K, 50K = 245,000. Brief: the brief adds ~1,000, then 2 turns at ~21K and ~26K = ~47,000, rounded to 50,000. The done list (list what must be true, mark pass/fail/unknown, one correction pass) is from @Greshamtrade's full course on X (1.65M views, 1,463 bookmarks, x.com/Greshamtrade/status/2112214741103825314). Grok Bot has no plan mode: this is a habit, not a feature. Also skip ALL CAPS "you MUST": say it once, calmly. I brain dump with Wispr Flow (https://ref.wisprflow.ai/jordan-rafealov-tvka) and let the bot turn it into a brief. The fix: Goal, why, what to avoid, and a done list. Pass, fail or unknown, then one fix pass. Demo: Brain dump a job with Wispr Flow, then ask: "Turn this into a brief: goal, why, what to avoid, and a done list. When you finish, mark each item pass, fail or unknown and fix only the fails."

04 / TRICK 04 OF 09

Text, not screenshots or PDFs

04

5x to 9xless usage per screenshot, ~5x per PDF

BEGINNER MISTAKE

Screenshots of text and 20 page PDFs dropped in as is

SIMPLE FIX

Paste the words as text. Convert PDFs to text first.

CREDITS: PICTURES OF WORDS VS THE WORDS (TEXT = 1x)

ONE FULL-SCREEN SCREENSHOT

Screenshot~5x to 9x

Same words as text1x

A 20-PAGE PDF

Read as pages~5x

Converted to text1x

ESTIMATE · REAL NUMBERS VARY

TRICK 4. What people do: Screenshots of text and 20 page PDFs dropped in as is. Why it burns credits: An image is billed as an image, whatever is in it, and a PDF read as pages can come in as text plus a picture of every page. Text is just the words. And the bot can edit text, it can't edit a picture. The number: A full-screen screenshot uses roughly 5x to 9x the credits of the same words pasted as text. A 20-page PDF read as pages uses roughly 5x the credits of the same PDF converted to text (estimates). Assumptions: Image costs depend on resolution and how the model tiles images; a 1080p screenshot typically lands around 1,500 to 2,700 credit units. A screen of text is ~1,200 characters, about 300 units. PDF: ~3,000 to 4,500 units per page as text plus page image vs ~600 to 800 per page as plain text, 20 pages each way. Whatever you drop in stays in the bot's history and gets re-read on later messages too. The fix: Paste the words as text. Convert PDFs to text first. Demo: Paste an error as a screenshot, then as text, and show the bot fixing the text version. Then ask: "Convert this PDF to a text file first, then answer from the text."

05 / TRICK 05 OF 09

Keep big outputs out of the conversation

05

200xless re-reading

BEGINNER MISTAKE

Letting an 800 line log land in the conversation

SIMPLE FIX

Save long outputs to a file, reply with one line

CREDITS OVER THE NEXT 20 MESSAGES (FILE + ONE LINE = 1x)

800-line log in the conversation~200x

Saved to a file + one-line summary1x

ESTIMATE · REAL NUMBERS VARY

06 / TRICK 06 OF 09

Stop it browsing

06

~4xless usage, script vs clicking

BEGINNER MISTAKE

Letting the bot click around websites

SIMPLE FIX

Connector first, then a script. The browser is the last resort.

ONE PRACTICE SITE, SAME JOB, USAGE (APPROX.)

Clicking around (computer use)~213K

A script in the browser~51K

A plain script, no browser~0 while it runs

SOURCED: @Michael_Fenech_, HIS OWN APPROXIMATE TEST (MEASURED IN TOKENS)

TRICK 6. What people do: Letting the bot click around websites. Why it burns credits: Browsing is screenshots, scrolls and clicks, over and over, and every screenshot is read by the model. A connector gets the same data in one call. A script reads the page without the model looking at it. The number: @Michael_Fenech_ ran the same job on one practice site: clicking around used about 4x more than a script inside the browser (213K vs 51K, as measured by @Michael_Fenech_), and a plain script used essentially nothing while it ran. Assumptions: His own test, approximate, one site (x.com/Michael_Fenech_/status/2115421573524148479; 67K views, 357 likes, 895 bookmarks). The plain-script run still uses the model around the run (writing and reading the result), just not during it, so the chart shows a sliver, not zero. His ladder: 1 connector, 2 plain script, 3 script in the browser, 4 clicking only for what's left. The fix: Connector first, then a script. The browser is the last resort. Demo: Ask your bot: "For this task, use a connector first. If there isn't one, write a script. Only open the browser if neither works, and close it when you're done."

07 / TRICK 07 OF 09

Reset a bloated bot

07

~4xless usage per turn (one report)

BEGINNER MISTAKE

One bot dragging months of history into every turn

SIMPLE FIX

Export its memory, put must-keep rules in its description, then duplicate it.

USAGE PER TURN, ONE USER, BEFORE AND AFTER DUPLICATING

Old bot, full history158.8K

Duplicated copy, clean history35.7K

REPORTED BY A FORUM USER (IN TOKENS), CITED BY @GrokBotRadar, NOT VERIFIED

TRICK 7. What people do: One bot dragging months of history into every turn. Why it burns credits: Grok Bot has no fresh-start button, and a bot's history only grows. Every turn re-reads it. A duplicate keeps the bot's setup but starts with a clean history. The number: One forum user's usage per turn dropped from 158.8K to 35.7K after duplicating, about 75% less. Same source: one conversation grew from 10.1K to 254.6K per turn in just 7 turns. Assumptions: Both numbers are one forum user's report, cited by @GrokBotRadar (x.com/GrokBotRadar/status/2115828008061653058, 18.7K views; and x.com/GrokBotRadar/status/2105065358038360334). Not verified, one report, not a benchmark. The 3 steps: 1) ask the bot to write everything it remembers about you, your preferences and its standing rules to a file, 2) put the must-keep rules in the bot description, 3) duplicate the bot and have the copy read the file. Hide the old bot while you check the copy. Jordan: confirm where Duplicate lives in your app before filming. The fix: Export its memory, put must-keep rules in its description, then duplicate it. Demo: Ask an old bot: "Write everything you remember about me, my preferences and your standing rules to a memory export file." Then show the description with the must-keep rules. (No click-path for Duplicate on screen.)

08 / TRICK 08 OF 09

Script the repeats, reuse the research

08

Under 1%of usage for the scripted part

BEGINNER MISTAKE

Redoing mechanical steps and the same research

SIMPLE FIX

Repeat it 3 times? Script or skill. Save research with an expiry date.

ONE IMAGE WORKFLOW, SHARE OF USAGE

Model parts: generate + check images8%

The scripted part: upscalingunder 1%

SOURCED: @stupiz VIA @GrokBotRadar. DIFFERENT STEPS, SAME WORKFLOW

TRICK 8. What people do: Redoing mechanical steps and the same research. Why it burns credits: When the model does a mechanical step, it reads and reasons through every item. A script does it for almost nothing. Same with research: if the bot already found it last week, re-researching it pays twice. The number: One Grok Bot user tracked an image workflow step by step: the scripted part (upscaling, run as commands on the bot's computer) used under 1% of his usage. The model parts (generating and checking images) used 8%. Assumptions: Source: @GrokBotRadar's article on X citing @stupiz (x.com/GrokBotRadar/status/2115489802682585477). Two different steps in the same workflow, not the same step done two ways, so say "the scripted part used under 1%". Research reuse is from @Greshamtrade's course: save research with an expiry date and check for a saved result before researching again. The fix: Repeat it 3 times? Script or skill. Save research with an expiry date. Demo: Ask the bot: "You've done this step 3 times. Turn it into a script and save the instructions as a skill. And save this research with a date, so you check it before researching again."

09 / TRICK 09 OF 09

Stop bots chatting with each other

09

~7xfewer credits for one task

BEGINNER MISTAKE

Status pings and bot-to-bot back and forth

SIMPLE FIX

One complete brief per bot. In groups, message one bot directly.

CREDITS FOR ONE TASK (ONE DIRECT BRIEF = 1x)

Manager + 3 bots chatting back and forth~7x

One complete brief to the right bot1x

ESTIMATE · REAL NUMBERS VARY

TRICK 9. What people do: Status pings and bot-to-bot back and forth. Why it burns credits: Every message a bot receives makes it re-read its own history. A manager bot pinging 3 bots for status, and them pinging back, multiplies that. And one message in a group wakes every bot in it. The number: A manager plus 3 bots chatting about one task uses roughly 7x the credits of one complete brief to the right bot (estimate). Assumptions: Assumes each bot carries ~40,000 credit units of history. Chatter: each of 3 bots gets 3 messages (brief, clarification, status ping) = 9 x 40K = 360K, plus the manager reads 6 replies and pings = 6 x 40K = 240K, total ~600K. Direct: one complete brief, 1 to 2 turns of the right bot = ~80K. Group chats: one message in a group of 4 bots wakes all 4 (4 x 40K = 160K) vs 40K messaging one bot directly. Estimate. The fix: One complete brief per bot. In groups, message one bot directly. Demo: In a group with several bots, show one message waking all of them, then message the one bot that owns the job directly with a complete brief.

MYTHS / QUICK HITS

3 myths that don't save a thing

MYTHShorter prompts save credits

TRUTH

Your typing is tiny. One creator measured his typed prompts at 0.01% of his usage.

MYTHDeleting old bots saves credits

TRUTH

A bot with nothing running costs nothing. Check it has no routines first.

MYTHPassing one job between bots saves credits

TRUTH

Every hand-off makes another bot read the context. It moves the cost and usually adds more.

QUICK TIPSUnplug connectors and skills you don't useHiding a bot doesn't stop it

MYTHS. 1. Writing shorter prompts barely matters. One creator measured his typed prompts at 0.01% of his usage (his figure, I didn't measure it). 2. Deleting old bots saves nothing by itself. A bot with nothing running costs nothing. Check it has no routines first. 3. Passing one job between several bots doesn't save credits. Every hand-off makes another bot read the context, so it moves the cost and usually adds more. (Different from trick 1: one bot per kind of job is good, one job bounced between bots is not.) QUICK TIPS: Unplug connectors and skills you don't use (their descriptions can load every session; how much is not documented for Grok Bot). Hiding a bot doesn't stop it: @GrokBotRadar reports hidden bots keep running routines, and one archived bot was still running weekly. Delete idle bots you don't need.

CHECK YOUR NUMBER / TWO MINUTES

Check your number every week

IN GROK BOT

Settings›Usage & Billing›Usage by bot›Usage by day

BY BOT

Which bot eats the most?

BY DAY

Which day spiked?

THEN ASK

What was I doing that day?

THE PROMPT / PASTE THIS INTO GROK BOT

Paste this. Lean credits mode.

Turn on Lean Credits mode. Save the rules below as a skill called "lean-credits", plus a memory note to keep it on for all my work.

1. Before any big job, brief me back in one message: the goal, your plan, what you will avoid, and a done list of what must be true at the end. Wait for my yes.
2. Show me a cheap preview first (an outline or a sample) before you build the full thing.
3. Batch fixes. If I send them one at a time, ask if there are more, then apply them in one pass.
4. When you finish, mark each done-list item pass, fail or unknown, and do one correction pass on the fails only. Max 2 tries on a failing step, then stop and ask me.
5. Save long outputs to a file and reply with one line: what it is and where it is.
6. Ask me for text instead of screenshots, and convert any PDF to text before you work with it.
7. Use a connector or a script before the browser, and close the browser when you're done.
8. If you do the same step 3 times, turn it into a script or a skill. Reuse saved research if it's still fresh.
9. Don't re-read, re-render or re-deploy anything that hasn't changed.
10. When I give you work unrelated to your main job, suggest a dedicated bot for it.
11. When you hand work to another bot, send one complete brief. No back and forth.

Reply "Lean credits mode is on".

AI SYSTEMS SOCIETY / SKIP THE TRIAL AND ERROR

All my bots live in AI Systems Society

WHAT'S INSIDEMy bots plus 35 AI systems

Skool Bot

Cooper

Stitch

Byte

X Growth Bot

$47/monthusually ~~$67/month~~SAVE 30%

www.skool.com/aisystems/about

HOW I BRAIN DUMP MY BRIEFSWispr FlowLink in the description

Paste This Into Grok Bot.Never Run Out of Credits Again

The receipts

Most of your credits go to re-reading.

First, find your leak

9 tricks, half the credits

One bot per kind of job

Batch your fixes. One message.

Brief it once, with a done list

Text, not screenshots or PDFs

Keep big outputs out of the conversation

Stop it browsing

Reset a bloated bot

Script the repeats, reuse the research

Stop bots chatting with each other

3 myths that don't save a thing

Check your number every week

Paste this. Lean credits mode.

All my bots live in AI Systems Society

Paste This Into Grok Bot.
Never Run Out of Credits Again