AI-powered engineering workflows
The previous two posts in this batch were specific. Why keep a journal at all; how to wire one to Slack. This one is the bridge between them and what I’ll be writing about for the rest of the series — what changes, structurally, when an AI agent enters the daily workflow loop.
I want to be careful here. AI doesn’t make the underlying work better. It doesn’t make my judgment sharper, my code cleaner, or my decisions more correct. Those parts are still on me. What it changes is the friction around the work. And the friction was, it turns out, a much bigger drag on quality than I’d have admitted.
This post is about which parts of the daily loop AI handles well, which parts still need human judgment, and the small reorganization of effort that follows when the friction drops.
The loop, in plain terms
Most working engineers run some version of a daily loop. The pieces have different names in different places, but the shape is roughly:
- Capture — what did I do? What did I notice? What’s still open?
- Plan — what am I doing tomorrow? What’s the priority order?
- Execute — the actual work. Code. Reviews. Conversations.
- Report — what did I do this week, and how do I tell the relevant audiences?
The middle of that loop — the actual work — is where the value lives. The other three are supporting infrastructure. They’re how the value becomes legible to your future self, your team, and your stakeholders.
For most of my career, the supporting infrastructure was undertaxed. I’d capture maybe twice a week, plan only when I felt anxious enough about the queue, and report mostly under deadline pressure. The pattern was: skip the supporting steps, then catch up at the worst possible moment. It worked, sort of. It also burned a meaningful amount of energy on each catch-up, and the catch-ups were the lowest-quality versions of the work.
The thing AI has changed for me, more than anything else, is making those supporting steps cheap enough to do daily. Not better-done — just easier to start.
Where AI is unambiguously good
Three jobs in the loop where I now reliably hand work to the agent without second-guessing.
Structuring loose input
If I dump a stream-of-consciousness paragraph at the agent — “today I worked on the auth thing, also looked at the deploy issue, talked to the client about the rate-limit feature, started exploring whether we even need rate limiting at the gateway” — the agent turns it into a structured journal entry without losing anything. Three categories: what I did, what I noticed, what’s open. Bullet form. Concrete enough to be searchable later.
I could do that conversion myself. It would take ten minutes. The agent does it in fifteen seconds. The difference isn’t that the agent’s structuring is better than mine — it’s that I’ll do it, because the cost dropped from ten minutes to fifteen seconds.
The same applies to taking a chat-thread context file and pulling out the three things relevant to my work. Or taking a long pull-request description and summarizing what changed. Structuring loose input is the easiest case for AI assistance and the one I lean on most.
Reformatting for different audiences
I used to write the same week’s worth of work three different ways: one version for my manager, one for the team channel, one for my own self-assessment. Each version emphasized different things. Each one took twenty minutes. Two of those twenty-minute blocks were essentially the same content rephrased — and yet I had to do all three because the audiences want different things.
The agent does that reformatting cleanly. One source — the journal entries — multiple outputs. Manager report leads with outcomes and risks. Team report leads with what’s in flight and where help is needed. Self-assessment leads with patterns and friction. Same week, three lenses.
The reformatting isn’t creative work. It’s translation between registers. The agent is good at translation between registers; I am also good at it but find it boring. That’s a great division of labor.
Searching the archive
Six months of journal entries is a lot of plain text to grep through manually. “When did I first try X? What did I conclude about Y? Did I ever resolve the thing with Z?” These are questions a human can answer with twenty minutes of clicking around or thirty seconds of asking the agent.
The agent isn’t doing magic — it’s reading the entries faster than I can and pattern-matching against the question. But the speedup is the difference between “I’ll look that up later” (i.e., never) and “I’ll look that up now” (i.e., useful). The archive becomes interactive. Which is the only reason an archive is worth keeping in the first place.#
Lowering the cadence threshold
The fourth thing AI does well is the meta-effect of the first three: it lowers the cadence threshold of the entire loop.
Before, capture was a five-to-ten-minute task. I’d do it some days. Plan was fifteen minutes. I’d do it some weeks. Report was an hour or two. I’d do it under deadline.
Now, capture is two minutes — I dump the day’s notes, the agent structures them. Plan is three minutes — the agent reads the last few entries, suggests priorities, I edit. Report is five minutes — same source, three lenses, click to send.
The wall-clock cost dropped enough that the entire loop moves daily instead of weekly, and weekly instead of monthly. The compounded benefit is that decisions, problems, and patterns are now visible while they’re still actionable. A pattern I’d have noticed six months later, I now notice within a week.
That’s the structural change. It isn’t about doing the same work faster. It’s about being able to do the work at the right cadence — which I couldn’t, before.
Where AI is unambiguously not yet good
Three jobs in the same loop where I do not hand work to the agent. These are still mine.
Deciding what matters
The agent is good at structuring what I tell it I did. It is bad at deciding which thing I did matters most. That decision depends on context the agent doesn’t have — what my manager will care about next quarter, which problem on the team is politically charged, which technical debt is about to bite us, which conversation in the hallway shifted my read of a project.
I’ve tried letting the agent prioritize. The result is plausible-but-wrong prioritization. The bullets are real; the ranking is generic. “Auth refactor — high priority” sounds right and is wrong, because the agent doesn’t know that the auth refactor is downstream of an architectural decision that hasn’t been made yet.
So when I plan, the agent suggests bullets. I rerank. The reranking is the actual work; the bullets are the input. Subtracting the human-driven reranking would lose the part that matters.
Editorial judgment
When I write a manager report, I’m making editorial choices the agent cannot make for me. Which two of the eight things I worked on are worth highlighting? Which of the open problems is worth flagging? What tone matches this manager, this week, given what’s currently on their plate?
The agent will produce a perfectly competent manager report on its own. It will also be slightly off. It will highlight the wrong two things, hedge the wrong problems, miss the political weight of a particular phrasing. The result is a report that reads polished but lands flat — and worse, sometimes lands wrong.
What works for me is to use the agent to produce a draft, then spend five minutes editing the draft with my own judgment baked in. The five minutes is the irreplaceable part. The draft is the cheap part the agent saved me.
Telling the truth about myself
The journal entries that have the most long-term value are the ones where I admit something uncomfortable about my own week. I made a worse decision than I’d thought. I procrastinated. I was wrong about how a conversation went. The agent will not produce those sentences. It can’t — the only person who knows the uncomfortable truth is me.
So when I capture a day, the agent’s structured output is the scaffolding. Inside that scaffolding, I write the one or two lines that are actually honest. Those are mine. They’re the part that pays back, six months later, when I read the entry and find the small piece of self-knowledge that the agent’s polished version would have hidden.
I want to be explicit about this: the agent is not a substitute for the introspective work. It’s a substitute for the clerical work around the introspective work. Confusing the two would erode the journal’s value entirely.
What changes when the loop runs daily
Once the supporting steps are cheap enough to do every day, the day itself feels different.
You start each morning with yesterday’s notes already structured and a plan that took three minutes to draft. You execute against a clear target. At day’s end, you capture in two minutes and head home. Each Friday, the report writes itself out of the week’s entries; you spend five minutes editing.
The aggregate is something like an extra hour a week, distributed in small chunks. But the bigger effect is invisible: you stop carrying the supporting work in your head. You’re not thinking “I should write that down later”; you wrote it down already. You’re not thinking “I need to draft that update”; the draft exists. The cognitive overhead of unfinished metawork drops, and the freed attention goes back into the actual work.
That’s the trade I keep coming back to. The agent doesn’t make me a better engineer. It clears the small administrative weight that was sitting on top of my engineering. The engineering itself gets a quieter background.
The honest version of this
I want to close with something I find easy to overstate.
The setup I described isn’t magic. There are days when the agent’s outputs are noisy and I throw them away. There are weeks when I skip the loop anyway because I’m tired or busy or the work has been chaotic. The system isn’t a productivity hack; it’s a small, quiet infrastructure that mostly works and sometimes doesn’t.
The reason I think it’s worth describing is that the direction of change has been consistent. Every iteration of this setup has produced more useful entries, sharper plans, and reports my audiences actually read. The friction keeps dropping. The cadence keeps tightening. The journal keeps becoming a more useful artifact instead of a less useful one.
That’s the whole win. Not “AI changed everything.” Just: a thing I knew I should be doing daily for years, that I now actually do daily, because the cost dropped enough that the habit could form. Some of that is discipline; most of it is the cost.