Guide
Write down what you meant to do before each AI session, review your commits and diffs at the end of the day, and note what you decided, not just what changed. If you want numbers on editor time and AI usage per project, WakaTime wins. If your day spans the browser, docs and chat as much as the editor, Dayflow's whole-screen timeline on a Mac covers more.
It is 6:40 in the evening. You have been at your desk since nine, the agent has been busy all day, and there are eleven commits on your branch. Someone asks what you got done and you say “a lot of auth stuff” while you scroll through the history.
The odd part is that you were never idle. You wrote prompts, read plans, approved diffs, rejected two approaches, chased a flaky test through three browser tabs and answered a review comment in chat. But the day does not come back as a story. It comes back as a blur of terminal output and “looks good, continue.”
This article covers why AI-assisted days are so hard to remember, six things to try that mostly need nothing more than paper and a timer, and where an automatic record of your screen fits. Dayflow, a free Mac app, publishes this guide, and we are clear below about where it is the wrong tool. Product details come from vendor documentation only: WakaTime’s own pages as read on October 4, 2026, and Dayflow’s 2.2.0 release. Nothing here was benchmarked.
A few things happen at once, and most people who code with an agent will recognize them.
Typing used to be the memory. When you wrote every line yourself, the act of writing left a trace. You remember the function you struggled with because your hands were in it. When an agent writes the lines and you review them, the work still happens, but it happens in your judgment, and judgment leaves fainter marks than effort.
Much of the day is waiting. You send a prompt, the agent runs, and you have thirty seconds or ten minutes. Many people fill that gap with something else: a message, a tab, a different ticket. Each gap is small, but by evening the day is made of fragments that never connected.
Work spreads across more surfaces. The editor, a terminal agent, a browser chat, docs, the pull request, Slack. Commit history captures only the part that landed in the repository. The research that ruled out a library, the hour spent reading a stack trace and the plan you threw away leave no commit at all.
“Approve” is not a memorable verb. Accepting a diff takes a second. Deciding whether to accept it may have taken twenty minutes of reading. Your memory tends to keep the second and lose the twenty minutes.
None of this means you were unfocused. It means the shape of the work changed and the old memory cues stopped working. The fix is to add new cues on purpose.
What to do: Keep a sticky note or index card next to the keyboard. Before each session with an agent, write one line: the outcome you want and how you will know it is done. Not the prompt. The goal behind it.
Illustration: “10:15 - Retry logic on token refresh. Done when the expired-token test passes and no duplicate refresh calls show in logs.” Later: “1:40 - Figure out why CI is slow. Done when I know which step, not when it is fixed.”
This does two jobs. It keeps you from drifting while the agent runs, and it leaves a timestamped trail of intent that is far easier to recall than terminal output. At the end of the day, the card reads like a table of contents.
How to tell after a week: Look at the five cards. If you can tell someone what you worked on each day just by reading them, it is working. If the lines are vague (“auth stuff”), push yourself to write the “done when” part every time.
What to do: Keep a running decision log in a plain text file or a notebook. One line each time you approve a plan, reject an approach, change direction or verify something by hand. Write the reason in a few words.
Illustration:
A change log tells you what moved. A decision log tells you what you contributed, which is the part that matters in a review, a standup or a conversation with your future self three months from now. The /templates/daily-work-log/ template gives you a simple structure if a blank file feels like too much. For a longer take on keeping this kind of record, see how to remember what you worked on.
How to tell after a week: Pick any decision from Monday. Can you explain why you made it without opening the code? If yes, the log is earning its place. If you skipped logging on the busiest days, make the line shorter rather than dropping the habit.
What to do: Before you write any summary, spend ten minutes reading the day’s diffs top to bottom, the way a reviewer would. Not to fix anything. Just to see the work.
Illustration: You run your usual git log for today, open each commit and skim. You notice that three commits all touched the same error handler, which tells you the “small fix” in the morning actually ate the whole afternoon. You also spot a commit message that says “wip” and realize you never finished the test it was meant to add.
Reading diffs with fresh eyes often surprises people. Code you approved at 11 a.m. can look unfamiliar by evening, which is exactly why the review is worth doing while it is still recoverable.
How to tell after a week: You should be catching at least occasional surprises: a forgotten “wip”, a change bigger than you thought, a fix you do not remember making. If every review feels like old news, you can shorten it to five minutes. If the surprises are frequent, keep it at ten.
What to do: Get a cheap kitchen timer. When you send a long task to an agent, set the timer for a sensible check-in, say eight minutes, and do one planned thing in the meantime. Decide what that thing is before you start, and write it on the same card as your intent.
Illustration: “Agent running the test suite rewrite. Timer: 8 minutes. Meanwhile: answer the two review comments on the billing PR.” When the timer rings, you go back to the agent whether or not you finished, and you note it.
The point is not productivity for its own sake. It is that unplanned waiting turns into unrecorded wandering, and unrecorded wandering is what makes the day vanish. Planned waiting leaves a trace. If you find your guesses about how long tasks take are often off, the /tools/time-estimation-test/ is a short exercise comparing your sense of time with a clock.
How to tell after a week: Count how often the timer rang and you could name what you did in the gap. If you usually can, the wait time is now part of your day instead of a hole in it. If you keep ignoring the timer, try a louder one or a shorter interval.
What to do: At the end of the day, explain what you did out loud in under two minutes. To a friend, a partner, a coworker on a call, or a rubber duck on the desk. Three parts: what moved, what you decided, what is next.
Illustration: “Token refresh retry is done and tested. I said no to the new cache layer because the bug was one line. CI is slow because of the Docker build step, and tomorrow I start by caching that layer.”
Speaking forces you to turn fragments into a sequence. If you get stuck, that is useful information: it usually means one block of the day has no record at all, and you can check the card, the decision log or the diffs to fill it. If you prefer writing, the /templates/end-of-day-report/ template follows the same three parts, and the same notes make automatic standup updates much easier to put together.
How to tell after a week: Time it. If you can do the recap in two minutes without opening anything by Friday, it is working. If you still need to dig, the gap is probably in the morning or right after lunch, so put an extra intent line there.
What to do: For one week, at the top of every hour, write a single word on a sheet of paper for what the last hour mostly was: prompting, reviewing, waiting, debugging, meetings, reading, chat. Do not judge it. Just mark it.
Illustration: Tuesday reads “meetings, prompting, waiting, waiting, reviewing, chat, debugging, reviewing.” Two hours of “waiting” before lunch is the kind of thing nobody would guess from the commit log.
A week of hourly words gives you a rough shape of where AI-assisted time actually goes, which is often different from where it feels like it goes. The /templates/time-audit/ template has a grid ready to print, and the /tools/where-does-my-week-go-calculator/ helps you add up the categories at the end.
How to tell after a week: At the end of this one, you have your answer. The question is whether the shape surprised you. If it did, repeat the audit in a month after you change one thing. If it matched your guess, you probably do not need to run it again.
Everything above depends on remembering to do it. On heavy days, the card stays blank and the decision log has two lines. That is when a record you did not have to make is useful.
Dayflow records your screen on your Mac and has AI write the day into a readable timeline. You do not start a timer or tag anything. It sees the browser research, the chat thread and the docs, not just the editor, so the hour spent ruling out a library shows up next to the commit it led to. This is the idea behind where did my day go: look at the day instead of rebuilding it from memory.
Full screen recordings and the timeline are stored on your Mac. You choose how screenshots are analyzed: Local AI runs a supported model on your Mac through Ollama or LM Studio and keeps analysis on-device, while your own cloud provider or Dayflow Pro processes selected screenshots off-device. The details are on the local and cloud processing page. You can export a date range as Markdown from Settings → Export and use it for a Codex work summary or a weekly report.
AI summaries can miss or misinterpret work, so check the timeline before you share it.
Dayflow is our product, and we publish this article. It is $0 and MIT licensed, with an optional Pro plan at $20 a month.
You want coding metrics, not a timeline. If what you want is time per project, language and editor, or how much AI you are actually using, WakaTime wins. Its open source plugins cover 100+ editors, IDEs and terminals and send coding time and AI usage to a web dashboard, including model usage and spend and AI-generated versus human-written lines across tools such as Claude Code, Cursor, Copilot and Codex. It also has team dashboards. Note that it captures only activity inside supported editors and terminals, the free plan keeps 1 week of dashboard history, and data goes to WakaTime’s hosted dashboard. For plan prices, see the vendor’s pricing page. Our WakaTime vs ActivityWatch comparison covers it in more depth.
You are not on a Mac. Dayflow needs macOS 14 or later. There is no Windows or Linux version.
The work happens off-screen. It sees only the screen. A whiteboard session, a phone call or thinking on a walk will not appear.
You need a coach or a blocker. Dayflow is a record. It does not block websites, nudge you back to work or tell you what to do next.
You need business features. It has no invoicing, payroll approvals or team reports. For those, start with our roundup of Mac time trackers, or the developer-focused list.
AI-assisted coding moves your contribution from typing to judging, and judgment is easy to forget. The fix is cheap: one line of intent before each session, a decision log while you work, ten minutes with your diffs at the end, and a two-minute recap out loud. Try two of those this week and see which ones stick. If your days still vanish, a WakaTime dashboard can show your editor and AI numbers, and an automatic record of your whole screen can show the rest.
Much of the day was spent waiting, reading and approving rather than typing, and those moments leave few memory hooks. Work also moves in short bursts across the editor, terminal, browser and chat, so no single thread feels like the day's story. A short note of intent before each session and a decision log during it give your memory something solid to hold on to.
Log what you decided and checked, not who typed the lines. For each task, write the goal, the approach you approved or rejected, what you verified by hand, and what is still uncertain. That record shows your judgment, which is the part a reviewer or your future self actually needs. A plain text file or a daily work log template is enough.
For coding metrics, yes. WakaTime's open source plugins cover 100+ editors, IDEs and terminals and record coding time per project, language and editor, plus AI usage such as model usage, spend and AI-generated vs human-written lines. Its free plan keeps 1 week of dashboard history. Dayflow is the better fit when you want a readable record of your whole screen, not editor metrics.
No. Dayflow's timeline describes what was on your screen across the day, written up by AI from screenshots. It is not a cost report or a line-count report. If you need AI adoption, model usage and spend, or AI-generated versus human-written lines across tools like Claude Code, Cursor, Copilot and Codex, WakaTime tracks those through its editor and terminal plugins.
It depends on the mode you choose. With Local AI, a supported model runs on your Mac through Ollama or LM Studio and screenshot analysis stays on-device. With your own cloud provider or Dayflow Pro, selected screenshots are processed off-device. Under the privacy policy, Dayflow Pro deletes screenshots after processing and may retain Activity Data for sync. Your own provider has its own terms.
Dayflow is free at $0 and MIT licensed. There is an optional Dayflow Pro plan at $20 a month, which manages cloud analysis for you. The free app does not include paid API usage, so if you connect your own cloud provider, that provider's charges apply. Dayflow runs on macOS 14 or later and has no Windows or Linux version.
Three short parts work well: what shipped or moved, what you decided and why, and what is open for tomorrow, including the exact next step. Add anything you verified by hand, since that is easy to forget on AI-heavy days. Keep it under ten lines so you actually write it. Five minutes before you close the laptop is usually enough.
Pull from three places: your commit messages, your decision log and the intent notes you wrote before each session. Rewrite them as outcomes, not activity, such as 'auth retry fixed and tested' rather than 'worked with the agent on auth'. Name one blocker if there is one. If you keep an automatic record of your screen, skim it to catch work that never reached a commit.
Yes. You can export a date range as Markdown from Settings → Export, then paste or attach it wherever you like. Keep in mind that local storage does not make a cloud assistant conversation local: any assistant you share the export with receives that data under its own terms. AI summaries can also miss or misread work, so check them before sharing.
Dayflow is $0 and open source. Full recordings stay on your Mac, with local AI available for on-device analysis.
Download for MacOS macOS 14+ · MIT licensed · no account required