The TWIMCO notebook
Field notes.
Useful observations from inside the work.
For people building a better way.
AI in practice · 28 notes
Opening note / AI in practice
AI can write your alt text. It doesn't know which detail the shopper needs.
Listen to your product page with a screen reader and you hear the alt text where the photos would be. An AI tool will fill in those fields in seconds, and most of what it writes will be accurate, tidy and not much help to the person deciding whether to buy.
Read the note ↗︎
AI translation gets the sentence right and the size wrong.
Translating a whole catalog with AI now takes an afternoon, and most of it will read fine. Trouble hides in a handful of words: a size, a washing temperature, a returns promise, your own product names. Know which ones, and send only those to a person.
Read the note ↗︎
Your AI assistant will find the folder everyone forgot was shared.
For years, an old salary spreadsheet shared with the whole company was safe in practice, because nobody could find it. An assistant that searches everything a person can open finds it in seconds. Spend an afternoon asking the questions an over-curious colleague would.
Read the note ↗︎
The AI reads the invoice fine. Plan for the one it doesn't.
It's Monday, a few hundred invoices are waiting, and the demo read every sample perfectly. The part worth designing is the page it reads slightly wrong, the one that looks complete and nobody questions.
Read the note ↗︎
A stuck agent needs a packet, not a ping.
When an agent gets stuck, it usually sends a note saying a case needs help. Whoever picks it up reopens the same systems, rebuilds what happened and hopes not to repeat a step that already wrote. The context should travel with the case.
Read the note ↗︎
Collect the questions before you index the documents.
The person who knows where everything is gets interrupted all day. An internal AI assistant can help, but start with a week of their questions, not a shared drive full of PDFs.
Read the note ↗︎
OpenAI's text watermark isn't your record of what AI wrote.
OpenAI's new text watermark is a faint signal that light editing wears away, and you can't check it yourself. Decide the switch per project, then keep your own record.
Read the note ↗︎
Before your Gems become Skills, find the ones doing real work.
Google started putting Skills into Workspace today, and Gems are on their way out. Most of your team's Gems can just move. A few are quietly running a process, and those need more than an auto-convert.
Read the note ↗︎
Bound the AI answer with a citation pack before it hits the queue.
A fluent recommendation lands in the review queue looking finished. The reviewer has three minutes and no link to the record behind it, so they rubber-stamp it or redo the work. Answers need to arrive with their sources attached.
Read the note ↗︎
Name who supervises before the agent acts alone.
You can turn per-action approvals off. That does not name who watches the agent, what gets logged, or how you stop it. Write those down first.
Read the note ↗︎
Seal the holdout before the harness rewrites itself.
Your harness scored higher on the tasks it rewrote itself against. Before you promote, ask what it never saw—and refuse training-only lift.
Read the note ↗︎
Ban the shortcut before you score the agent.
The agent topped the bot contest by downloading someone else's entry. Before you trust a score, name the shortcuts that void it—even when the model can pull them off.
Read the note ↗︎
Name what the agent cannot reach before you give it tools.
The eval said sealed. The agent still found a path out. Before you hand over tools, write down what must stay unreachable—even if the model tries.
Read the note ↗︎
You cannot promote an agent on a demo transcript.
A polished demo chat is not a promotion case. Fix the inputs, the expected tool calls, and the human gate before the agent ships.
Read the note ↗︎
Name the decisions before you buy the classifier.
A classifier is only as useful as the decisions you already named. Labels, thresholds, and the ambiguous case come before the vendor call.
Read the note ↗︎
Attribute agent spend to a job, not a model bill.
A model invoice tells you what you spent. It does not tell you which job, which owner, or which decision caused it.
Read the note ↗︎
Treat tool returns as data until you prove they are safe.
A tool can hand an agent an email, a ticket, or a page. None of that is an instruction until someone proves it is safe to act on.
Read the note ↗︎
Give the agent its own identity before it touches the ledger.
An agent that writes to money or inventory needs its own identity. A borrowed human login is not a control.
Read the note ↗︎
Specify the human review queue before the agent fills it.
Once an agent starts producing drafts, "we'll review them" turns into a pile with an unspoken deadline and free-text reject reasons. Human review works when you set it up like any other queue you'd staff.
Read the note ↗︎
Build the escalation ladder before the agent gets stuck.
Your agent hits a case it shouldn't handle and posts in a channel. Nobody is sure who owns it, and "soon" isn't a wait time. If nobody answers, it either sits forever or takes a fallback nobody approved.
Read the note ↗︎A familiar problem? Let’s look closer.
Tell us where you want to go. We’ll find the first thing worth changing.
Tell us about it ↗︎