# Shipping with AI

> I use prompts to set the direction and skills to make the useful parts repeatable.

The first prompt asks for research, a plan, phases, and subagents. Skills take care of the repeated work after that: review, simplification, UI checks, releases, and PRs. I keep those on [Agent Skills](/agent-skills).

Canonical URL: https://blode.co/guide

## Chapters

1. [The kickoff prompt](/guide/agents-write-the-first-draft)
2. [Taste is the job](/guide/taste-is-the-job)
3. [Turn prompts into skills](/guide/install-the-habit)
4. [The skipped edit](/guide/the-skipped-edit)
5. [Ship at 80 percent](/guide/ship-at-80-percent)
6. [How I actually use AI](/guide/how-i-get-shit-done)

## Principles

Ten things I’ve learned about shipping, across two startup exits and 90M+ users at Linktree. The tools come last.

### 01. Ship fast or die

Ask what’s slowing you down and get rid of it: blocked tasks, too many meetings, fuzzy ownership, PR reviews. Find the 10% that gives you 90% of the value, and get it in front of real people. Nobody knows what works up front. The quickest way to find out is to put it in someone’s hands.

### 02. Don’t get attached

Now that building is cheap, throwing code away costs nothing. Your first version is just a guess, so don’t get attached and keep iterating.

### 03. Call every customer

At Fingertip we called every user. An actual phone call, then another one after they’d had a go with it. You hear the hesitation before someone works out how to say it politely. Ten people who need the thing will teach you more than a thousand who quite like it. Ring them and shut up while they talk. Skip it and you’ll ship a year of features nobody asked for, faster than anyone ever has.

### 04. Own it, then hand over the keys

Kill off “that’s not my job” wherever it shows up. One person owns a feature the whole way through. If you can’t yet say what “done” looks like, sort that out first. Ownership doesn’t scale by doing more yourself. It scales by handing someone real authority and the accountability that goes with it, both at once. One without the other isn’t fair on anyone.

### 05. Build lots, keep it simple

Solve your own problem, give the tool away, and show people how you did it. Find the little crux of each project and get it as simple as you can. Once it clicks, it’s like a ball rolling downhill.

### 06. Risk scales with blast radius

An early MVP can ship fast and loose: push to main, get it in front of users. A platform with 90M+ people on it earns feature flags, staged rollouts, and proper review. The scariest code, anything touching money, tax, or permissions, earns stronger tests even when AI wrote it. CI going green doesn’t mean the product works.

### 07. Stay at the frontier

Put yourself where you’ll learn the fastest, and keep learning the newest thing. Speed is a skill, and you only build it under pressure, when real people need a real answer and the deadline isn’t moving.

### 08. You can sense carelessness

It’s the sum of all the small details that people pick up on, even the ones they’d never consciously notice. So generate fast, then go and sand down the rough edges, and keep sanding. The thing that still sets you apart is knowing what’s good and standing behind it.

### 09. Get in the room

Play the long game and optimise for opportunities. Do the talks, build in public, help people when there’s nothing in it for you. You can’t plan luck, but you can give it more ways to find you.

### 10. Get it out of your head

Don’t carry everything in your head. Put it in a system you trust, so you can be fully present on the one thing in front of you.

When AI can write anything, your edge is the stuff it can’t do: taste, ownership, being in the room.

## Who this is for

People using Claude Code or Codex to ship real products. It assumes the agent can already write useful code.

## Who this is not for

This isn’t a prompt library or a vendor comparison. It’s how I work after two exits and shipping AI at Linktree.

## Read next

Start at [chapter 1](/guide/agents-write-the-first-draft), or jump to [how I actually use AI](/guide/how-i-get-shit-done) for the short version.
