AI Project Template

Why most AI project boards quietly become a lie

Most AI work doesn't ship. That's not a failure rate to hide. It's close to the normal shape of the work, and a board that pretends otherwise stops being useful fast. A generic task board assumes most things started will eventually finish and go live; AI experimentation doesn't work that way, and forcing it into that shape produces a board that either lies about progress or gets abandoned. Three habits cause that.

Every idea is tracked as if it's heading to production. A hypothesis and a shipped feature get the same card type, so the board can't tell you, at a glance, whether something is a two-day experiment or a committed roadmap item, and everyone plans as if it's the latter.

Experiments that don't pan out just vanish. Nobody wants to write "we tried this and it didn't work" on a board that looks like a delivery tracker, so the card gets archived, deleted, or simply stops being updated. Three months later, someone tries the same thing again, because there was no record it had already been tried.

"Evaluated" gets skipped entirely. An experiment runs, the numbers come back mediocre, and it either ships anyway because momentum carries it, or it dies without anyone writing down why the results weren't good enough.

The cost isn't abstract. A team that treats every experiment as a future feature ends up either shipping things that shouldn't have shipped, because stopping feels like admitting failure, or quietly burying the ones that don't work, which guarantees somebody tries the same idea again in six months with no memory that it was already tried and found wanting.

This template treats "shelved" as a legitimate, tracked outcome, as normal as shipped, with a place to log what was learned. Nothing here integrates with a model registry or an experiment-tracking platform; it's a board for the human decisions around AI work, not a replacement for the tools that run the training.

How to use it

1. 01Start a workspace. It opens with a sample project already on the board, so the flow from hypothesis to shipped-or-shelved is visible before you add real work. Free for up to five people, permanently.
2. 02Phrase every new card as a question, not a plan. If you can't state what you're unsure about, it probably isn't an experiment.
3. 03Set the Experiment column's WIP limit low. Two or three running at once is usually enough for a small team to actually evaluate properly instead of half-watching five.
4. 04Write the evaluation before deciding the outcome. Record the number first, decide shipped or shelved second, not the other way round. Deciding the outcome first and backfilling a rationale is how mediocre experiments quietly ship.
5. 05Log every shelved idea in the wiki, briefly. A few sentences on what was tried and why it stopped is enough to save someone a wasted week later.

A team that ships one out of every four experiments and logs the other three properly is doing better work than a team that ships one out of every four and lets the other three disappear without a trace. The ratio isn't the problem. The silence is, and the shelved column with its wiki entry is what keeps the team's memory intact instead of resetting every quarter.