Ankesh | Thinking AI

Running my work on a team of Grok @Bot bots. Breaking down how small businesses can do the same, one copyable setup a day.

[Ankesh | Thinking AI @ankeshdev]

https://t.co/DlVs4YqN9e

[Ankesh | Thinking AI @ankeshdev]

Your bot doesn't need longer instructions. It needs to see one finished result you were happy with.

Wrote up the five pieces that let a workflow move to a new bot, or a new person, without you re-explaining the bar.

https://t.co/5LQ4cwawbL

[Ankesh | Thinking AI @ankeshdev]

https://t.co/yAjY1hwAlJ

[Ankesh | Thinking AI @ankeshdev]

@Henryf1w Big accounts shall answer this small question...😅

[Ankesh | Thinking AI @ankeshdev]

@NickpxJ Yes

[Ankesh | Thinking AI @ankeshdev]

@mattyp @bot I have one @bot exactly doing this...

[Ankesh | Thinking AI @ankeshdev]

@Henryf1w That's a "big" question for "small" accounts...😅

[Ankesh | Thinking AI @ankeshdev]

@Kai_muses That's the way to go...

[Ankesh | Thinking AI @ankeshdev]

Silence is not approval. Your AI agent should know that.

The risky moment in any agent workflow is the checkpoint where it should ask you, hears nothing, and carries on with its best guess.

- Write the stop points into its instructions instead:
- Never move past a checkpoint without an explicit yes.
- Any assumption needs a yes in words.
- No silent changes to anything already approved.

An agent that waits costs you a few minutes. An agent that guesses can quietly undo a decision you already made.

[Ankesh | Thinking AI @ankeshdev]

@robj3d3 This is really encouraging...love @X for this...

[Ankesh | Thinking AI @ankeshdev]

@leo_mercerr What are you building? Curious to know...

[Ankesh | Thinking AI @ankeshdev]

@melvindvivas I just name my CoS...rest follow their role as their names...

[Ankesh | Thinking AI @ankeshdev]

@Henryf1w Claude...yes, it has!

[Ankesh | Thinking AI @ankeshdev]

@sumelabs What's your Moat for Sume?

[Ankesh | Thinking AI @ankeshdev]

Apparently 37 replies a day is spam, so I'm guessing my ideal number is somewhere around "mysterious". What's yours?

[Ankesh | Thinking AI @ankeshdev]

If you could hand one weekly task in your business to your Grok @bot starting Monday, which one would it be?

The boring ones count. Especially the boring ones.

[Ankesh | Thinking AI @ankeshdev]

@bot Wow...that would be such a relief!

[Ankesh | Thinking AI @ankeshdev]

@ericzakariasson @bot Fewer approval cards for things you already asked for is the quiet winner here. Vault gets the headlines, this gets the hours back.

[Ankesh | Thinking AI @ankeshdev]

@Michael_Fenech_ Every prompt ending in "get my approval before…" is the real win here. That's what makes people comfortable handing over bank statements.

[Ankesh | Thinking AI @ankeshdev]

@mattyp This is awesome...any update on releasing for individual plans?

[Ankesh | Thinking AI @ankeshdev]

@milesdeutscher The $999 audit is the smart bit… most owners don't need more bots, they need someone to tell them which 3 steps to kill first.

[Ankesh | Thinking AI @ankeshdev]

Your @bot keep picking up the old draft because nothing tells them which one is current.

Give them one front door: a single index doc they open before any task.

Three rules make it work:
1. Every file name carries its status and date: APPROVED, PENDING or SUPERSEDED.
2. When a new version lands, the old one is renamed SUPERSEDED and moved to an archive folder the same day.
3. Bots decide what's current from the index only, never from search results.

Give your bot this:
"Create one index doc for this project and open it before every task. Put status and date in every file name. When a new version lands, mark the old one SUPERSEDED and move it to Archive the same day. Only treat files marked APPROVED in the index as current."

How do you keep your AI from working off stale files?