First Last Word @FirstLastWord_

Systems, not sentences. Rules for real work with AI — whichever model you use. Joined September 2026

Tweets

54
Followers

5
Following

47
Likes

13

First Last Word @FirstLastWord_

2 weeks ago

Sam Altman, 29 September: “Dots are remarkably capable, always-on agents built to handle really anything you can think of.” As a message, it cannot be wrong. As a brief, it is worse. No one is named as the reader. The job never ends. Whatever you put in comes back covering anything, and you cannot check it on Friday. Name the job and the date it is done, or bin it. The machine does not. cnbc.com/2026/09/29/ope…

0 0 0 11 0

View Details

First Last Word @FirstLastWord_

3 weeks ago

It’s 8:50. I need an update I can send by 9. This is for: [me / my manager / the group] What moved: [one thing] What is stuck: [one thing] Next step: [action] Who owes it: [me / a name] ← skip this line if it’s only you I won’t send this if: [next step missing / nobody on the hook / the number isn’t up top] Write the update. Short. Don’t add tone, a role, or extra headings.

0 0 0 13 0

View Details

First Last Word @FirstLastWord_

3 weeks ago

You’re about to ask ChatGPT to write the update so you can send it in ten minutes. If you don’t say what usable means, it will guess. If you pile on tone and extra headings, it will still guess. Either way it will sound finished. That’s how a bad update gets sent. Write the missing pieces first. Then ask.

1 0 0 29 0

View Details

First Last Word @FirstLastWord_

3 weeks ago

@sflorimm They built something that answers every question, and it turns out the hard part was deciding which questions it should answer.

0 0 0 32 0

View Details

First Last Word @FirstLastWord_

3 weeks ago

@chamath AI can hand back the three minutes. It cannot decide they were worth handing back.

0 0 1 168 0

View Details

First Last Word @FirstLastWord_

3 weeks ago

@techspence A workflow can finish. A decision can be logged. That still is not judgement. Reviewing agreeable output is the layer most people skip naming.

0 0 2 16 0

View Details

First Last Word @FirstLastWord_

3 weeks ago

@Reniec7 That’s a different fail, and a real one. The task was yours. The source wasn’t current. Completing it on old data still didn’t move the need-to-have.

0 0 0 16 0

View Details

First Last Word @FirstLastWord_

3 weeks ago

@moneet If an agent can finish it and the goal didn’t move, it wasn’t work. Take it off the list.

0 0 0 13 0

View Details

First Last Word @FirstLastWord_

3 weeks ago

@niketrajdwivedi If two names are on the line, stop the task until one name owns the miss.

1 0 1 22 0

View Details

First Last Word @FirstLastWord_

3 weeks ago

It’s 8:40. The note has to be in by 9. You open the chat because you have to. You’ll get something fluent fast. Write one discard line into the brief first, so the finished-looking draft has something outside it that can still kill it. Mine for a status note: the draft must name who pushes the fix and by when. If either is missing, “the team will resolve,” or any other soft stand-in, discard. If who and when still aren’t in the brief, write them in before you generate again. Sharp prose does not save the failed draft. Kill date / metric: Sun 28 Sep 2026. I used this discard on one real status note. Pass if I binned or rewrote a draft because who or when was missing. Fail if I only published the advice.

0 0 0 39 0

View Details

First Last Word @FirstLastWord_

4 weeks ago

Worked example: iPhone Duo, three models, none said “both.” firstlastword.substack.com/p/give-every-t…

0 0 0 13 0

View Details

First Last Word @FirstLastWord_

4 weeks ago

Verdict: Metric: Wrong by: Against it: Could a stranger tell you were wrong without asking you?

1 0 0 35 0

View Details

First Last Word @FirstLastWord_

4 weeks ago

“Time will tell.” “It depends on execution.” “Real upside and real risk.” A take that survives every outcome hasn’t said anything. Ask a model “will it work?” and that’s what you get back. Give the question something it can lose.

1 0 0 63 0

View Details

First Last Word @FirstLastWord_

4 weeks ago

Ask a model to check something and it will find a way to say yes. A check that can't come back empty hasn't checked anything. Claim: one sentence you could check Pass / fail / not enough: the only answers allowed What would count: the evidence that settles it, set before you look Nothing is an allowed answer. Reply with your "what would count" for one claim you're about to trust this week.

0 0 0 19 0

View Details

First Last Word @FirstLastWord_

4 weeks ago

@mardehaym Is the PE portfolio one a call you were on, or something you heard secondhand?

0 0 0 104 0

View Details

First Last Word @FirstLastWord_

4 weeks ago

@0xsael @beamnxw What's a retrieval case a good harness still can't handle without someone tuning it by hand?

0 0 2 16 0

View Details

First Last Word @FirstLastWord_

4 weeks ago

Worked example, recovered from a six-round churn brief: firstlastword.substack.com/p/every-not-th…

0 0 0 8 0

View Details

First Last Word @FirstLastWord_

4 weeks ago

Audience: Must include: Must not: Throw-away test: Write that before you ask the model to write anything.

1 0 0 17 0

View Details

First Last Word @FirstLastWord_

4 weeks ago

Most AI drafts get finished like this: “Good, but drop the intro.” “Better. No ‘headwinds.’” “Perfect.” You never said what good was. Every “don’t say that” was a scorecard line, written late. Your last chat is the wreckage. Here’s how to recover it.