When people describe what AI coding assistants do, they reach for the wrong word. They say "genius." They say "brain." They say "it knows so much." That's not it. The best mental model is older and more practical.
Your AI assistant is a midwife.
Midwives don't give birth. They help you do it. They know hundreds of patterns you don't. They suggest positions, breathing, timing. They recognise distress. They intervene precisely. They document. They stay calm while you are not. But the baby is yours.
This framing changes how you use Claude Code, Cursor, GitHub Copilot, and everything like them. And it changes what you stop expecting of them.
The genius-in-a-box problem
The "genius in a box" metaphor is everywhere. Marketing encourages it. The model answers confidently. It writes code that runs. You start to treat it like a peer who happens to know more than you.
Then you ship something it confidently wrote. Production breaks. You read the code you "wrote together" and realise you didn't write it, and neither did it — nobody understood it. The pattern was right; the context was wrong.
Meanwhile, the midwife metaphor sounds demeaning to the AI. "Just a helper?" We feel cheated. We want Claude Code to be the genius. That's the product pitch.
But a midwife is not a demotion. A senior midwife is someone you want in the room for the hardest hour of your life. They have three thousand deliveries behind them. They're not the parent. They're not the obstetrician. They are exactly what they are, and that is enough.
What they actually share
Midwives and AI coding assistants share four load-bearing traits:
Both know patterns you don't. A midwife has seen breech births, shoulder dystocia, cord prolapse, and a thousand variants of "completely fine." They pattern-match faster than any first-time parent ever could. Claude Code has seen millions of React components, Rust borrow-checker errors, Postgres migrations, broken CI configs. Neither of them will teach you anatomy or distributed systems in the moment. But when you need the next move, they have options ready.
Both hand back agency. Good midwives don't grab the baby. Good AI assistants propose and wait. The moments when a midwife overrides are deliberate, documented, and rare. Most of the time, they're narrating what you're doing back to you, which is how you learn. The same is true for coding assistants: the best sessions aren't the ones where AI did all the work; they're the ones where AI narrated the options and you chose.
Both respect the room they're in. A midwife doesn't walk in with a playbook from another hospital. They watch, read, adapt. They don't tell you what epidurals are best; they know the ones your clinic offers. The best way to use Claude Code is to give it your repo conventions, your types, your tests — the room — and let it respect them. When it doesn't, the code feels like a graft.
Both document in real-time. Midwives fill in charts as they work. Not after. AI assistants should do the same: commit messages, docstrings, migration notes. Not because it's sexy, but because the next person in the room (including you, in a week) needs it.
Four shifts
If you take the midwife metaphor seriously, four things change about your day:
Own the name. You are the parent. If you ship code you don't understand, the AI was not your midwife — it was your ghostwriter. Read what gets committed. If a line surprises you, that's your signal to either learn it or reject it. Midwives never ship a baby the mother didn't push.
Don't ask for the whole thing. Ask for the next contraction. "Make me a full payments module" is not a midwife question. "What's the safest next step here?" is. Use AI for the next 20 lines, not the whole product.
Set the scene before the work. Before the session: paste your conventions, link the types, run the tests once. It costs 90 seconds and it's the difference between an assistant that respects the room and one that brings in hospital-brand protocols you don't use.
Keep your own notes. A midwife's chart is the client's medical record, not the midwife's. Same applies to Claude's output: you own it. Keep the commit messages, the PR descriptions, the decision logs in your own voice. The AI can draft them. You approve them.
The real test
The best part about a midwife is that a good one makes you feel more capable, not less. You did the hard thing; they helped. That's the mark of a senior practitioner of any kind — human or otherwise.
If your AI assistant leaves you feeling like a spectator in your own code, switch the framing. Not a genius. A midwife. See if the next session goes better.
Related reading
- Prompts are recipes, not spells — why spells belong to the caster and recipes belong to the team.
- Your AI agent should plan like a kitchen brigade — division of labour for multi-step agents.
- MCP servers are USB-C for AI — the boring protocol ending the integration drawer.
We build AI-enabled software and help teams put AI to work. If you're integrating AI assistants into your engineering workflow, we'd love to hear about it. Get in touch.