Working with LLMs, as a daily craft
Most of what's written about working with LLMs falls into one of two piles: breathless claims that your job disappears next quarter, or dismissals that it's all glorified autocomplete. Neither pile helps you on a Tuesday morning when real work is due.
At Heivol we've had LLMs in the loop for most of the working day for about two years now — writing, planning, code, infrastructure. The honest summary is less dramatic than either pile: it's the biggest change to our ways of working since version control, and it rewards the same thing every serious tool rewards. Craft.
What follows is the small set of patterns that survived contact with daily use.
Work the loop, not the prompt
The mistake almost everyone makes at the start — we certainly did — is treating the model like a vending machine: perfect prompt in, finished work out. That framing produces disappointment and worse, distrust. What actually works is a loop.
Plan — you decide what good looks like before the model types a word. What's the artifact? Who is it for? What must be true when it's done? Draft — the model produces a first version fast, and speed is the point: v1 exists to be reacted to, not admired. Critique — a conversation, not a re-prompt. "Tighten it. Numbers first. You buried the risk." Push back the way you'd push back on a colleague's draft. Verify — you check what's checkable. And ship — you put your name on it.
Notice who owns each step. The model owns exactly one of five. Everything that makes work good — intent, judgment, standards, accountability — stays with you. That's not a limitation of today's models. It's the design.
Context in, quality out
The single biggest quality lever isn't a clever prompt. It's what you hand the model to work from.
"Write a project update" with nothing attached gets you generic filler — it can't be anything else, because the model knows nothing about your project. The same request with your rough notes, the previous update, the audience and the tone comes back close to ship-ready. A few habits that pay for themselves every single day:
- Hand over sources, not summaries. Raw notes beat your compressed memory of them.
- Name the audience and the tone in one line. "For the board, direct, numbers first" is the cheapest quality upgrade there is.
- Show one previous example of the thing done well. The model matches shape and register better than it follows abstract instructions.
- Keep a context file per project — what it is, the decisions so far, the vocabulary. Start every session with it, and every session inherits everything the last one knew.
Trust, but verify
Models are at their most fluent precisely when they're wrong. Confidence in the tone tells you nothing about truth in the content. So between draft and shipped there's a gate, and the gate is yours:
Facts get checked against the source. Numbers get recomputed. And the voice has to actually be ours — if a paragraph doesn't sound like something we'd say, it doesn't go out. When something fails the gate, it goes back with a sharper instruction. That's not the tool failing; that's the loop doing its job.
The non-negotiable part: nothing goes out under your name that you haven't read and would defend line by line. The moment you ship unread output, you've stopped working with the model and started forwarding its mail.
Habits that compound
- Work in small batches. One section, one function, one decision per exchange. Small batches are easier to critique and far easier to verify.
- Ask for options before answers on anything hard. "Give me three approaches with tradeoffs" beats one confident answer, because choosing is your job anyway.
- Guard your voice. Models sand the edges off prose. Put the edges back in critique, or everything you write starts sounding like everything everyone writes.
- Know when to leave it out. Decisions that need taste, values, or accountability are yours alone. And don't automate the parts of the work you love — those are why you do it.
- Write down what works. Good prompts and context files are assets. They compound; one-off cleverness doesn't.
Don't change alone
Everything above is individual craft. But if you're trying to shift how a team or a company works, individual craft hits a wall fast, because people move at very different speeds with this. Some colleagues already run a model all day. Some are curious but haven't started. Some are skeptical. Some have quietly decided it isn't for them. A change this size doesn't get lifted by one enthusiast — it needs more hands.
The cheapest structure we've found is an interest group — and we run one ourselves. Not a committee, not a mandate: a handful of people who care, a thirty-minute slot every other week, and one rule — only real work gets shown. Someone walks through a workflow they actually used that week: the prompt, the context they attached, what came back, what they had to fix. No slideware, no vendor demos. A few things ours has taught us:
- Local proof beats external hype. A colleague showing this morning's real work moves more people than any keynote we could stage.
- Everyone joins at their own speed. Watchers watch, triers try, and skeptics ask the hard questions out loud — exactly where we want the hard questions asked.
- The lifting distributes. Each member carries the practice back into their own corner of the work. Champions everywhere beat one exhausted evangelist.
- Practice converges. The group is where our prompt library, our context-file templates, and the "what's actually OK to paste in here?" guardrails got settled — one shared way of working instead of five private ones.
Keep it voluntary. The fastest way to kill an interest group is turning it into a mandatory meeting with an attendance list. Ours runs on interest — that's the engine.
Where to start
Not with twenty tools. Pick one recurring writing task — the weekly status, release notes, meeting summaries — and run the loop on it for two weeks. Plan first, attach context, critique hard, verify before shipping. It will feel slower for the first week and faster forever after.
And if the change you're after is bigger than your own desk, don't wait until you're an expert. Find two or three colleagues who care and book the first half hour. The group can start before anyone has the answers — ours did.
The gains were never in typing speed anyway. They're in the quality bar you can hold at speed once the loop is muscle memory.
The images in this article were produced with the workflow it describes: planned by us, drafted by a model, critiqued, verified, shipped. This working style is also what we bring to consulting engagements. Disagree with any of it? [email protected] — a real person reads replies.