The agent finishes the feature. You review it, it’s good. Now the other job starts: open the board, find the card, drag it to “Review”, type a note about what changed, update the checklist, set the next task to “In progress”.
Five projects, several sessions each, every day. Part 4 of Working for the AI — the board you keep for a robot that could keep it itself.
The board becomes a second job
A separate task board asks for one thing: that you tell it what happened. Every time. With agents doing more of the work, “what happened” changes more often, so the board asks for more of your time.
And the day you’re too busy to update it, it starts lying. A week later it describes a project that no longer exists. You stop trusting it, and it becomes one more tab you don’t look at.
The agent already knows the status
At the end of a task, the agent knows exactly what it did, what’s left and what went wrong. The status exists — it’s just in the wrong place, the chat.
So ask for it in the right place. End each task with:
“Update
progress.json: set the phase, write the next step, and add a note if anything is stuck.”
Now the status is written by the one who did the work, at the moment it happened, in the repo.
Read the board, don’t write it
If the status lives in a file in each repo, the board doesn’t need to be typed. It can be read from the files. That flips the relationship:
| Hand-kept board | Board read from the repo |
|---|---|
| You update it after the work | The agent updates the file during the work |
| Drifts when you’re busy | Is exactly as current as the repo |
| Lives in a separate tool | Lives next to the code, in git |
| One more thing to keep in sync | Nothing to keep in sync |
This is the same principle as memory living in files: the source of truth sits where the agent can read and write it.
Keep the file small and structured
The status file doesn’t need to be clever. One entry per task:
{ "name": "store listing", "phase": "5/6", "next": "captions for 3 screenshots", "stuck": null }Phase, next step, and a stuck reason when there is one. Structured enough that a tool can read it, short enough that you can read it too.
Your part: a quick check
You’re still the reviewer. Once a day, glance at the phases and next steps. If one looks wrong, fix the file or ask the agent to. That’s a minute, not a second job.
Where ShotMatic fits
ShotMatic is a board that reads instead of asking you to type. Your agent writes progress.json in each repo;
ShotMatic shows every task as one line with its phase, next step and stuck status. The first-run wizard writes a
starter file and a prompt so your agent keeps it up to date.
