Visibility
Tracking Where My Time Goes as an Engineering Director, With Zero Manual Logging
An engineer asked me what I do all day, and I couldn't give a straight answer. So I built the answer, a daily recap that reconstructs my day from the trail I already leave behind, and a chart that told me something I could feel but couldn't prove.
The situation
One of my engineers asked me a question with no malice in it: "What do you actually do all day?" I'm an engineering director, and I could not give a straight answer. My work is spread across Slack threads, Jira tickets, code reviews, meetings, hallway decisions, and a surprising amount of hands-on building. None of it adds up to a story I could point to. The honest answer was "I don't precisely know," which is a strange thing to admit about your own job.
A director's day is legible to almost no one, including the director. Individual contributors have artifacts: commits, merge requests, closed tickets. Most of what I do produces no artifact at all, or produces one in a system nobody cross-references. Ask me on Friday what happened this week and I'll give you the two or three dramatic things and lose the other forty hours to static.
That's not just an awkward moment in a 1:1. If I can't see where my time goes, I can't manage it, and my time is the main lever I actually control. The felt sense is unreliable in a specific direction: interrupt-driven work (Slack, meetings) feels smaller than it is because it arrives in fragments, and focused work feels bigger than it is because it's what I remember enjoying.
The constraint
Every conventional fix (time-tracking apps, journaling, calendar audits) requires manual logging, and I know exactly how long I'd sustain that. About a week. So the design rules were:
- Zero logging, ever. The system reconstructs my day entirely from exhaust I already produce. The moment it asks me to record anything, it's dead within a sprint.
- One picture across every surface, including the building time nothing else sees. Calendar audits miss the most interesting category of my work: hands-on engineering done alongside an AI agent, which lives in session transcripts, not in any calendar or ticket.
- A standing instrument, not a one-off report. It lives as a page on the dashboard I already look at, keeps a running structured log as the source of truth, and regenerates everything from that log.
- My data only. This is explicitly not a surveillance pattern to point at anyone else. It reads my messages, my calendar, my sessions. I left email out entirely, a deliberate line on what the system is allowed to read, not a technical limitation.
What I did
Working conversationally with an AI agent, describing what I wanted in plain language and never writing the integration code myself, I built a recap that sweeps six surfaces for a given date (Slack, Jira, GitLab, calendar, Google Drive, and my own AI-session transcripts) and turns the raw trail into one structured entry per day. Session transcripts become active building hours, with gaps over twenty minutes treated as not-working so a session left open overnight doesn't count. Calendar events become meeting hours, classified by nature into three buckets: mentorship (1:1s, coaching), process (standups, rituals, triage), and strategic (leadership, planning). Slack volume becomes an estimated communication load. Jira and GitLab contribute what actually shipped. The entries roll up into a stacked-bar trends chart. The first working version took about two hours.
It was not a clean two hours. I refined it the way you'd refine a junior analyst's first draft: "pull in the work I did with you too," "show me the last two weeks," "this number looks off." That last one mattered. The first pass under-counted my Slack activity roughly eight-fold, and it only got corrected because I challenged a number that didn't match my lived experience. The tool did the data wrangling. I supplied the skepticism. That division of labor turned out to be the whole method, and it set the working relationship: AI-assembled data gets treated like a draft from a smart junior analyst, genuinely useful, checked against lived experience, corrected out loud.
Two decisions I'd defend:
- Measured-only, no synthetic fill. The weekly totals looked low, because the logged signals genuinely can't see reading, thinking, planning, or review time. The obvious fix was to pad each day with an estimated "other work" block. I tried it, twice, and rejected it twice. The flat-topped bars looked exactly as fake as they were. The rule that survived: the chart shows what was measured, and a quiet bar means quiet tools, not a quiet day. The chart says so, in writing, right on the page.
- Splitting meetings by nature, not by name. "Four hours of meetings" is useless information. "An hour and a half of mentorship, an hour of process ritual, ninety minutes of strategic work" is a management document. It tells me which slice to defend and which to shrink. This one classification decision is most of what makes the chart worth reading.
What's happened so far
Once a few weeks of my time was sitting in the chart, a pattern jumped out that memory had been quietly editing. Meetings and communication were consuming the prime hours. Meetings actually edged out building time during the working day, and roughly 130 Slack messages a day fragmented whatever remained. My hands-on engineering hadn't shrunk to fit. It had moved, into evenings and weekends, in quantities I would have estimated at maybe half their measured size. One stretch of the log showed a string of consecutive days all running past 7pm.
None of this was a scandal. Some of it was even reasonable, and a crisis week looks like a crisis week in the data too, which is itself a form of validation. But the aggregate was a decision made by nobody. I never chose to do my building at 10pm. It accreted. The chart converted a vague guilt into a specific, dated, countable pattern, which is the only form in which a pattern can actually be fixed.
So I had the agent add a watch: a banner at the top of the chart that goes red when I've worked past 7pm, with a marker on every offending day. I asked for it to be blunt. The temptation with a self-built tool is to make it gentle with yourself, and gentle was exactly the problem. The banner led directly to a restructured week: designated maker days, designated meeting days, and 7pm as when the laptop closes.
Two honest flags. First, the communication column is modeled, not measured: it's derived from message counts at an assumed cost per message, because Slack doesn't tell you how long you spent. The knob is explicit and tunable, and I treat that slice accordingly. Second, the experiment has no verdict yet. The measured window is a handful of weeks, and I deliberately declined to automate the refresh, so the after-hours trend is only as current as my last run. If the red banner stops appearing because I stopped running recaps, that's failure wearing success's clothes, and I know it.
What I'd do differently
- Extend data retention before you need the history. The session transcripts only reach back so far, so the trend window can never see the early months of the year. I've since extended retention, but that data is simply gone. Retention is cheap. Reconstruction is impossible.
- Challenge the first number immediately, every time. The eight-fold Slack under-count got caught because it contradicted lived experience. A subtler error would have sailed through. Now every new signal gets one deliberate "does this match a day I remember?" check before I trust the trend it feeds.
- Instrument before the embarrassing question, not after. I spend my job asking teams to be measurable, and I'd never applied the standard to the one system only I can see. My first instinct when I couldn't answer was mild embarrassment. The right response was curiosity, about a year earlier.
Take what helps
This was the discipline I'd insist on for any team (measure before you optimize, and don't argue with the data) turned inward. It took an afternoon, cost almost nothing, and found a real, correctable problem that a year of self-reflection had politely stepped around. Pointed at a team instead of at myself, the same instinct needs a much harder look at consent, which is its own guide, and the version of this that got an audience is a different build entirely.
The leverage in working with AI this way isn't just building things faster, though two hours is fast. It's that questions that used to cost more discipline than anyone has (where does my time actually go?) drop below the threshold where you can no longer justify not knowing. My engineer got his answer. So did I, and mine came with a red banner attached.
Take what helps, skip what doesn't.