Tom Hooper

I run my business with Grok Bot and build with Cursor. Custom AI dashboards for Business Intelligence.

[Tom Hooper @HooperRules]

If I ask my agent for the same thing twice, it gets a template.

Inputs, steps, output format.

The third time, I just say "run the template." Results stay consistent and I stop re-explaining myself.

What request do you keep re-explaining?

[Tom Hooper @HooperRules]

For any big job, I make my agent show me the plan first.

What it'll read. What it'll change. What it'll report.

Fixing a plan costs me one sentence. Fixing a finished job costs me an afternoon.

What's the last job you wish you'd asked for a plan on?

[Tom Hooper @HooperRules]

@deepakguptacse Good stuff!

[Tom Hooper @HooperRules]

I keep my agent's queues and notes in plain text files.

Not a fancy tool. Files.

I can read them, diff them, and back them up in seconds. So can the agent.

When something looks off, I just open the file.

Where does your agent keep its working memory?

[Tom Hooper @HooperRules]

I don't trust a fact from my agent unless it shows me where it came from.

The email. The file. The page.

With a source, checking takes seconds. Without one, I'm doing the research myself.

So every summary I get now comes with links.

Does your agent show its sources?

[Tom Hooper @HooperRules]

I used to write agent rules as bare commands.

Now every rule gets a reason: "Back up first, because edits can go wrong."

The difference surprised me. With the why, my agent handles edge cases I never wrote rules for.

Do your instructions say why?

[Tom Hooper @HooperRules]

I don't want status essays from my agents when they get stuck.

I want one question I can answer from my phone:

"Approve this draft?"
"Which account should I use?"

A clear ask gets a fast answer from me.

How does your agent tell you it needs something?

[Tom Hooper @HooperRules]

When my agent's work gets worse, I don't blame the model first.

I check what I gave it.

Almost every time, it's a missing file, two instructions that conflict, or a stale note.

What's the last bad result you traced back to your own input?

[Tom Hooper @HooperRules]

I found my agent following a preference I'd dropped months ago.

My fault. The old note was still in its files.

Now I clean the notes on a schedule: ended projects, changed preferences, people who moved on.

When did you last clean up what your agent remembers?

[Tom Hooper @HooperRules]

My agent has a priority order, written down:

1. Don't post anything wrong.
2. Don't miss a post.
3. Keep the voice consistent.

When two rules collide, it knows which one wins.

Before I wrote that list, it guessed.

What's #1 on your agent's list?

[Tom Hooper @HooperRules]

I stopped asking my agent for "a draft."

Now I ask for three.

Picking the best of three takes me seconds. Editing one mediocre draft into shape used to take me much longer.

I decide. Grok Bot does the typing.

What do you ask your agent for more than one version of?

[Tom Hooper @HooperRules]

What do you ask your agent for more than one version of?

[Tom Hooper @HooperRules]

Nothing that pings me goes out after dinner or before coffee unless it's urgent.

Routines can still run. The interruptions wait.

Automation should protect my attention, not spend it.

When does your agent need to stay quiet?

[Tom Hooper @HooperRules]

"Tell me how it went" gets a wall of text.

"Report the titles, the counts, and anything you skipped" gets something I can read in ten seconds.

The format of the report is part of the job.

What do you want your agent to report every time?

[Tom Hooper @HooperRules]

When I tune a routine, I change one instruction and watch a few runs.

Change five things at once and I won't know which one helped or which one broke it.

Slow tuning ends up faster.

How do you tune your agent's instructions?

[Tom Hooper @HooperRules]

My agent instructions live in files, and I keep old versions.

When output suddenly gets worse, I can see what changed and roll it back.

Instructions are code for agents. Treat them that way.

Can you tell what changed in your prompts last month?

[Tom Hooper @HooperRules]

Every routine I run can be paused in one step.

If something looks wrong, I want to stop everything first and investigate second.

A kill switch is cheaper to build before you need it.

Could you stop all your automations in under a minute?

[Tom Hooper @HooperRules]

I used to hand agents one tiny step at a time.

Now I give the whole job: the goal, the files, the rules, and what to report at the end.

Fewer round trips. Better results.

Are you still micromanaging your agent one step at a time?

[Tom Hooper @HooperRules]

"Looks under 280 characters" isn't a check.

Counting each post with code is.

When the agent can verify its own work with a real test, I stop double-checking by hand.

What rule could your agent test instead of eyeball?

[Tom Hooper @HooperRules]

If a routine isn't sure, I'd rather it skip a post than publish something off.

A late post costs a little reach. A wrong one under my name costs trust.

I write that tradeoff into the instructions.

Which of your automations should fail safe?