Machine learning and AI teams rarely lose weeks to a hard algorithm. They lose them to Monday mornings that begin with a question like “where did I leave that experiment?” When the next move is unclear, even a capable engineer burns the first hour of the day reconstructing context. The eight habits below are aimed at one outcome for any artificial intelligence project: whoever opens the repository next, including your future self, knows the exact first action to take.
1. End the day with a failing test, not a finished one
Writing a test that fails for a specific, nameable reason is the cheapest possible bookmark. When the model’s evaluation helper needs to handle an empty batch, write the assertion for it and leave it red. The red line tells you precisely where to resume, and the failure message doubles as a to-do item that cannot go stale.
2. Write tickets that begin with a verb and a file
“Improve the retrieval step” invites a long stare. “Open retrieve.py and replace the fixed top-k value with a config entry” invites typing. A ticket that names a verb and an object removes the decision about where to start. If a ticket cannot be phrased that way, it is not yet a ticket; it is a topic, and it needs one more breakdown pass.
3. Keep one scratch note per experiment, with a final line titled “Next”
Training runs generate plenty of metadata, yet people seldom record what they intended to try after the run finished. Reserve the last line of the experiment note for a single instruction, such as “rerun with the cleaned label set and compare validation loss against run three.” Overwrite it every time rather than accumulating a list. Only the most recent line matters.
4. Pre-stage the environment before you leave
Half of the friction in resuming AI work is mechanical: activating the right environment, mounting the dataset, finding the notebook. Before stopping, open the notebook to the exact cell you will run first, and leave the terminal in the directory you need. Resuming then costs one keypress instead of ten small chores that each give procrastination a chance.
5. Let the pipeline name its own next step
When an automated job fails, the message should say what to do, not merely what broke. A log line such as “validation set missing; regenerate with the split script” beats a stack trace in ordinary use. Teams working on operations for machine learning can add a short remediation hint to every failure they have seen twice. The second occurrence is the signal that a hint is worth writing.
6. Separate “decide” tasks from “do” tasks
An item that mixes choosing and executing, such as “pick an embedding approach and wire it up”, tends to stall at the choosing half. Split it into two tickets. The decision ticket ends with a sentence in the notes; the doing ticket starts with that sentence. Neither ticket is ever vague, because each has only one verb. For a taste of how analytic choices feed practical ones, look at this short discussion of how machine learning supports everyday business choices, then ask which of your own tasks are really decisions in disguise.
7. Make the smallest possible first commit
A fresh branch with one commit that only adds a stub function and a docstring is a perfectly good place to begin. It converts a blank page into a document with a shape. The stub states the name, inputs and expected output in two lines, and the body can raise “not implemented.” The next session picks up the stub and fills it in. Systems that process images face the same temptation to plan everything at once, as this piece on scaling fast image-handling systems without creating bottlenecks shows; starting from a small, concrete component keeps the plan honest.
8. Read something short to reset your head, then act immediately
Sometimes the stall is emotional rather than logistical: the project feels large and your attention is scattered. A brief reading break can help, provided it ends with an action, not another tab. A blog-style website that arranges short articles in a news layout, like this learning-minded blog with trending lists and a search box, is the sort of place where a five-minute read fits between tasks. Set a timer, read one piece, then return and execute the line from your notes. Reading is a runway, not the flight.
Putting the eight habits together
You do not need all eight. In practice, three do most of the work for a small AI team.
- One red test left at the end of each session.
- One “Next” line in the experiment note, overwritten daily.
- One verb-first ticket at the top of the board, with a file name in it.
Together they answer the only question that matters at nine in the morning: what do I type first?
A short example from a model evaluation task
Suppose a team is checking a text classifier against a refreshed validation set. On Thursday afternoon the engineer discovers that two label names were renamed upstream. Rather than fixing everything before leaving, they do three small things. First, a failing test asserts that the label mapping includes both new names. Second, the experiment note’s final line reads “update mapping dict in labels.py, then rerun evaluation script on the refreshed split.” Third, the board’s top ticket is reworded to “Add two renamed labels to the mapping and rerun the evaluation.” On Friday the first twenty minutes are productive from the start, because nothing needs to be rediscovered.
Set that beside the more typical situation: a ticket reading “Fix evaluation issues” and an unremarkable branch. The second version is not lazier, merely less specific, and specificity is the entire difference between a smooth restart and a murky one.
The same pattern scales to a data-labelling sprint or a prompt-tuning afternoon. Whatever the activity, the closing question is identical: if I vanished for a week, what single line would let a colleague carry on without calling me? Writing that line is a ninety-second job that pays back hours.
Common mistakes when adopting these habits
- Writing the next step too grand. “Refactor the training loop” is a project. “Move the optimizer setup into its own function” is a step.
- Keeping several next steps. A list of five is a menu, and menus invite dithering. Pick one.
- Storing the note in a different place each time. Choose one location and stay loyal to it.
- Skipping the habit on busy days. Those are the days it pays most, since interruptions are likelier.
A closing instruction
Before you finish today, pick the habit that feels easiest and apply it once. Write the red test, or the “Next” line, or the verb-first ticket. Then close the laptop. Tomorrow’s version of you will open the project and know, without effort, what to do first. That small gift, repeated across a whole team, is what keeps AI work moving.