If you're running several sessions

This page is for a particular situation.

You are building something by having an AI write it. You keep two or three windows open and give each one a different job. The output is decent. The trouble is everything around it.

• You lose track of what each window is doing. Two of them might be editing the same file and there is no way to check.
• You can’t explain today what you built yesterday. Scrolling back through the conversation is the only way to find out.
• The AI says “fixed it” and you have no way to know whether it was.
• So showing anyone feels risky, because you have no list of what works and what doesn’t.

It feels like a problem caused by not being able to read code. It usually isn’t. It is a problem caused by nothing being written down. The decisions and the progress live inside chat windows, and chat windows close.

First: a conversation is not a record

The conversations are still there, you might say. They are. But three windows means the record is in three pieces, and answering “what is in flight right now” means reading all three. Worse, a new window can’t see the other windows. Only you can, and only by copying things across by hand.

Removing that copying is the whole point. However many windows you have, they all write to one board.

Three things this fixes

1. One page shows what’s in flight

Open the board and the cards are there. Before you open a new window, look at the board, pick a card, and start from that. Nothing overlaps because you chose what each window gets.

Assigning cards makes it firmer still. If two windows are holding the same card, you can see it.

2. Yesterday is written on the card

Open a card and there is a running log (Activity) underneath. What was tried, what didn’t work, what got done instead, in the AI’s own words as it went. No scrolling back.

It stops being something you have to remember and becomes something you read.

3. You get a list of things to check

The right-hand column of the board is Ready to ship. That means the AI finished every step it planned. It does not mean the work is good. What sits there is your list of things to go and click on yourself.

tuckit doesn’t check anything for you

It doesn’t test your code or verify that anything works. It gives you the list of what to check. That is still much better than no list: you move from “I think it’s all done?” to “there are five things to click”.

The button for when you don’t know what to type

This may be the most useful thing on the page.

Open a card and there is a Hand to an agent button near the top. Press it and you get text like this:

ACME-42 — Signup fails on Safari

Read it first: get_slice("ACME-42")
Stage is needs_design — brainstorm the design and write the spec into the slice.

Copy that, paste it into your AI, done. You don’t need to understand it. The wording changes depending on which column the card is in, so it always asks for the thing that card actually needs next.

Which column the card is in tells you what to ask for:

Column the card is in	What to ask the AI for
Needs design	Work out what to build, and write it into the card
Needs steps	Break it into an ordered list of steps
In progress	Do the next step and tick it off
Ready to ship	Nothing. It’s your turn to go and click on it

Three rules for running several windows

One. Make it read the board when you open a window.

What's the state of this project?

With an agent attached it usually does this on its own. If it doesn’t seem to, ask directly. If it answers by digging through files or commit history, it isn’t connected.

Two. One card per window.

Give one window several jobs at once and you are back to not knowing which window did what. Finishing one card before starting the next feels slower and isn’t.

Three. Make it write before you close the window.

This is the important one and the one people skip. With the plugin installed:

Reconcile the board with what we just did

With any other AI, say it plainly:

Write what we just did back to the tuckit board, and create cards for
anything we said we'd do next.

Skip this and the board starts lying. Once it lies you stop opening it, and a board nobody opens may as well not exist.

When you’re nervous about showing it to anyone

“I don’t know what works and what doesn’t” is usually three separate piles. If you have been using the board, you can go through them in order.

The Ready to ship column. Everything the AI called finished. Click through each one. When something is broken, don’t drag the card backwards. Write what broke into the log and add a step.

Each card’s Constraints box. Sentences like “faked for now” or “this isn’t real payment processing” live here. If there is something you need to know before going public, this is usually where it is. If the box is empty, ask now:

Is anything in what we built faked, stubbed, or not actually working?
If so, write it into that card's Constraints.

The Inbox. Things nobody has touched yet. The count alone tells you how many things you know about and haven’t done.

Read those three and you have a list you can say out loud. What works, what’s temporary, what isn’t started. You can say all three without reading a line of code.

Where to go

• Try it in ten minutes: if you haven’t set one up yet.
• What’s on the screen: what each card and box means.
• The daily loop: the three habits above, in full.