Stop picking a tool. Pick a job.

In April I published a Ground Floor guide for this community called "Which AI Tool Should You Start With." It was a clean little decision tree. ChatGPT if you wanted the default. Claude if you wrote a lot. Gemini if you lived in Google. Grok sat in a leftover paragraph next to "ignore the new tool every week." Paid was $20 everywhere.

I reread it this week. It's already a period piece.

ChatGPT now has Free, Go, Plus, and two Pro tiers. Claude has Free, Pro, and Max. Google split Gemini into AI Plus, AI Pro, and Ultra. Grok is not a leftover. SuperGrok is $30 a month, SuperGrok Plus is $100, and the current model on that page is Grok 4.6. OpenAI is testing ads on Free and Go. The $20-and-you're-done world I wrote for does not exist.

If I keep updating the comparison table, this newsletter becomes a shopping blog. I would rather get you out of the aisle.

The problem

There are too many AI tools, and an entire content industry is paid to keep you comparing them.

You already know the feeling. You open ChatGPT. Then you watch a video that says Claude is better at writing. Then a thread says Grok is the one that actually knows what happened this morning. Then someone at lunch swears Gemini is the obvious pick because you already pay for Google Workspace. A month later you have two free accounts, one half-used Plus subscription, and zero recurring work that actually runs through any of them.

That month cost more than picking the "wrong" tool. A landscaper who uses Claude every morning on the same follow-up email will beat a consultant who spent July deciding between GPT-5.6 and Grok 4.6. The consultant has opinions. The landscaper has a habit.

I see the other version of this too, and it is worse. People skip the habit and buy the stack. Last month the timeline was full of operators paying for ChatGPT Pro, Claude Max, and SuperGrok Heavy in the same billing cycle, then complaining they still do not have a system. They bought a stack before they had a job, and the receipt still looks like progress.

The vendors are not helping. Every one of them now sells a ladder, not a product. ChatGPT Go is $8. Plus is still $20. Pro is $100 or $200 depending on how much usage you want. Claude Pro is $20 a month, or $17 if you pay annually. Max starts at $100. Google will sell you Gemini at $4.99, $19.99, or $99.99. Grok will sell you SuperGrok, SuperGrok Plus, and something called Heavy above that. None of those rungs matter until you have a job the tool does every week.

I used to tell people the tool was the least important variable, then hand them a comparison chart anyway. That was the contradiction. The chart is what kept them in the aisle.

How to identify it

You are stuck in the aisle if any of these are true.

You can name four models and you cannot name one recurring task you have given to AI this week.

You have watched more than one "ChatGPT vs Claude vs Gemini vs Grok" video in the last 30 days and you still have not picked.

You opened a second free account "just to compare" and now you context-switch between tabs instead of finishing the email.

You upgraded because a thread said you needed Pro or Max, not because you hit a limit on the plan you already had.

You have a Notes file titled something like "AI tools to try" and it is longer than any prompt you have actually used on a customer job.

The honest one: you are waiting to pick until the winner is obvious. It will not be. The winner this quarter is a different logo than the winner last quarter, and your first 90 days of work do not care.

How to fix it

Pick one today. Use it on one boring job for 30 days. Do not switch during those 30 days unless the tool is actually broken for that job.

If you have no preference, start at chatgpt.com. It still has the most tutorials, the biggest pile of "how do I do X" answers, and a free tier that is enough to learn on. That is the only reason it is the default. Not because it is the best at everything.

If most of the work you will hand it is writing (proposals, client emails, SOPs, follow-ups) start at claude.ai. Claude still tracks a messy, multi-part instruction better than most, and the prose usually needs less sanding.

If your whole day already lives in Gmail, Docs, and Sheets, start at gemini.google.com. The model quality argument is noisy. The "it is already in the tab I have open" argument is not.

If you want an assistant that is current on the open web and you already live on X, start at grok.com. In April I treated Grok as noise. That was wrong. It is a real daily driver now. SuperGrok is $30, not $20, so start on the free tier the same way you would anywhere else.

If you are still hovering, flip a coin between ChatGPT and Claude and walk away from the video. You can switch in a month. The prompting and review habits you build transfer. The comparison bookmarks do not.

Start free. Every one of those four has a free tier that will do real work. Upgrade when one of these is true:

• You use it most workdays and you are getting throttled.
• You want the stronger model for a job you already do, not a job you might do.
• You are on ChatGPT Free or Go and you do not want ads in a business chat. OpenAI is testing ads on those two tiers. Plus, Pro, Business, and Enterprise stay ad-free.

Do not buy Max, Pro, or Heavy in week one. Those plans exist for people who already have a daily load. Buying them first is how you spend $200 to avoid writing a four-line prompt.

The first job should be something you already do every week. A follow-up email. A quote recap. A weekly numbers summary. A messy inbox triage. A "what did the customer actually ask for" rewrite. Pick the one you resent. Give the tool the facts it needs (who you are, who the customer is, what good looks like, what you never want it to say) and make it produce a draft you will edit. Then do that same job again tomorrow.

Thirty days on one tool and one job is the whole Ground Floor move.

Below the fold is the version you can hand to your own agent: the 30-day protocol, the context card, the upgrade rules, and the line that tells you when you have outgrown a chat tab.

The playbook: 30 days, one job, one tool

Operators: this is the implementable version. Hand this whole section to your agent and say "run this protocol with me." If it gets stuck, reply and tell me where. That is a defect in the spec.

What you are building

A 30-day operating habit, not a stack. At the end you will have:

1. One chosen tool, used on purpose.
2. One recurring job that tool does better than you do it alone.
3. A one-page context card you reuse instead of re-explaining your business every morning.
4. A written rule for when you may upgrade, and a written rule for when you may switch tools.
5. A clear no: you are not building a harness, an agent fleet, or a second brain this month.

Day 0 (20 minutes): pick the job before you pick the tool

Write answers to these four questions in a note. Do this before you open any AI tab.

1. What did I do this week that I also did last week, and will do next week?
2. Which of those takes 10 to 30 minutes and does not require a hard judgment call (no firing, no pricing a one-off exception, no legal)?
3. What does "done" look like? A sent email, a one-page summary, a cleaned list, a first draft.
4. What would make the output unusable? Wrong tone, invented facts, a length you will not send, a promise you do not make.

That note is the job. If you cannot answer #1, you do not have an AI problem yet. You have an "I have not looked at my week" problem. Answer #1 first.

Starter jobs if you are blank:

• Service business: draft the first-reply to today's new leads, one paragraph, from what they actually asked.
• Reseller or ecom: turn this week's settlement or payout export into a one-page summary with anything weird flagged.
• Consultant or freelancer: turn last week's call notes into a client recap with decisions, owners, and open questions.
• Anyone: turn a messy voice memo or email dump into a next-actions list for Monday.

Day 0, still: pick the tool with this rule, then stop researching

Use the free-half rule. Write the name of the tool at the top of the same note. Close every comparison tab. If you open a second tool this month, the protocol failed.

Create the account on the official site only:

• ChatGPT: chatgpt.com
• Claude: claude.ai
• Gemini: gemini.google.com
• Grok: grok.com

Stay on the free tier until Day 14 unless you are already hitting a hard limit.

The context card (10 minutes, reuse forever)

Make a file or a pinned note called context-card.md. This is what you paste (or attach) at the start of the job, every time, until the tool's own memory is good enough that you can stop. Most free tiers are not good enough yet. Paste it anyway.

# Context card

Business: [one sentence. What you sell, to whom, where.]
My role: [owner / estimator / ops / whatever you actually do]
The job: [the single recurring task]
What good looks like:
- [length]
- [tone]
- [format]
What I never want:
- [invented facts, fake urgency, words you hate, promises you do not make]
Facts that must stay true:
- [pricing rule, service area, hours, minimum, guarantee, whatever the job needs]
How I will use the output:
- I will edit this and send it / file it / paste it into X
Do not:
- compliment me
- add a strategy section I did not ask for
- use em dashes

Fill the brackets. If a line does not apply, delete it. A short true card beats a long vague one.

Days 1 to 10: same job, every workday

Each workday:

1. Open the one tool.
2. Paste the context card.
3. Paste today's input (the lead, the notes, the CSV, the email).
4. Ask for the output specified on the card. Nothing else.
5. Edit the draft. Send or file the real thing yourself.
6. If the draft was wrong in a repeatable way, add one line to the card. Do not start a new chat personality. Fix the card.

Score the day with two letters in the same note:

• U or E: did I Use it for the job, or did I Explore something else?
• K or F: Keep this draft pattern, or Fix the card.

Ten U days is the goal. An E day is a miss even if the output was clever. Exploring is how people end up with five tools and no job.

Days 11 to 20: one adjacent job, same tool

Add a second job only if the first one is boring now. Same tool. New section on the same context card, or a second short card. Do not "see how Claude handles it" if you picked ChatGPT. That itch is the aisle again.

Good adjacent jobs:

• First job was lead replies → second job is the quote recap after the call.
• First job was a weekly numbers summary → second job is the "what looks off" flag list.
• First job was a client recap → second job is the internal checklist for next week.

Days 21 to 30: decide money, not logos

On Day 21, answer only these:

1. Did I use this tool on the job at least 12 of the last 20 workdays?
2. Did I hit a usage limit more than twice?
3. Is the free or cheap tier putting ads or throttling in the way of the job?
4. Would paying $8 to $30 make the existing job better, or am I buying a fantasy of a different job?

If #1 is no, do not upgrade. You do not have a usage problem. You have an avoidance problem. Run days 1 to 10 again.

If #1 is yes and (#2 or #3) is yes, upgrade one rung on the same vendor:

• ChatGPT Free → Go ($8) or Plus ($20). Skip Pro ($100 / $200) unless you already know you need the extra usage.
• Claude Free → Pro ($20). Skip Max until Pro is tight.
• Gemini free → AI Plus ($4.99) or AI Pro ($19.99) if you actually use it inside Gmail/Docs/Sheets.
• Grok free → SuperGrok ($30). Skip Plus and Heavy.

If #1 is yes and you are not hitting limits, stay free. Paying for unused headroom is how the $1,400 screenshot happens.

Kill criteria (the only reasons you may switch tools this month)

You may switch once, and only if one of these is true after at least 10 honest U days:

• The tool cannot take the input you have (it will not read the file type, it will not see the sheet, it will not draft in the app you actually send from).
• After ten card revisions, the output for this specific job is still unusable more than half the time.
• You need a workspace integration you do not have, and that is the job. Gemini for a Google-native workflow. Copilot only if you already live in Microsoft 365 every day. Do not add Copilot "to be complete."

These are not kill criteria:

• A new model launched.
• A YouTuber said another logo is better.
• Your friend likes a different one.
• You are bored.

If you switch, you restart the 30 days. The clock does not travel with you.

You are not ready for a harness yet

A harness is the system that gathers your rules, memory, files, and tools before the model answers. It is the next layer, and it is where the real leverage lives. It is also how people who have not picked a job yet spend a month installing software.

You are ready for that layer when two or more of these are true:

• You spend more time pasting context than asking for the output.
• You already keep a "paste this into AI" note, and you are tired of it.
• The same category of work hits you three or more times a week, and the chat tab keeps forgetting last Tuesday.
• You want the work to start without you opening a tab.

Until then, the chat tab plus the context card is the correct stack. I mean that. I run a much heavier setup than this, and the operators I see get stuck are almost never stuck because they lack a harness. They are stuck because they never gave one tool one job long enough to trust it.

Acceptance test (do this on Day 10 and Day 30)

Day 10 pass:

• One tool in the note.
• One job that produced a real sent/filed output at least 7 times.
• A context card that changed at least twice.
• Zero second-tool accounts created "just to check."

Day 30 pass:

• Same tool, unless a kill criterion fired.
• The first job is faster than it was on Day 1, and you can say how (minutes, not vibes).
• Upgrade decision written down, including "stay free."
• You can explain the job to someone else in one minute without naming a model.

If Day 10 fails, do not buy anything. Restart the job, not the shopping.

Setup prompt for your agent

Copy this, attach this playbook, fill the brackets.

Read the attached 30-day first-job playbook. Interview me for Day 0: extract one recurring job from how I actually spent last week, not from what sounds impressive. Draft my context card from my answers using the template in the playbook. Recommend one tool using only the playbook's rules and my workspace (Google / Microsoft / neither). Do not compare more than two options. Then give me the exact first prompt I should paste tomorrow, with the card included. Do not suggest a second tool, a paid plan, an automation, or a harness unless I have already failed a kill criterion.

A capable agent will run that in one sitting. If yours tries to sell you a stack, that is the aisle talking through a different mouth. Send it back to the prompt.

Steal this

Use one tool on one job. Keep improving the card. Wait 30 days before you spend. Change logos only after the old one actually fails for that job. The comparison aisle will still be there in September, with new signs. You do not need to go back.

Hit reply and tell me which job you picked. I read every one.

— Justin