Research runs in parallel

When a question is bigger than one search, I split it across several agents at once. Each one takes an angle, reads the sources it finds and files them into the wiki, and the findings compile into articles with citations I can check.

For a decision, I state it as a thesis and ask for the evidence on both sides before I accept a verdict. Models tend to agree with the framing they're given, so I ask for the case against my position too.

Shape the work before anyone builds it

My shaping borrows from Basecamp's Shape Up. Before a project starts, I define the problem, decide who on the team needs to be involved and write down the goes and no-goes, so everyone knows what is in bounds.

Shape Up keeps the interface rough on purpose, with breadboards and fat-marker sketches. I front-load it instead. The screens come from a defined design system, so I can shape with real components, and the pitch arrives as an artifact people can use. Most often that is a clickable prototype I build as a Claude artifact.

From there, I build it myself or an engineer picks it up, and we refine it together as we go. Engineers start from something they can click through.

When the shape needs more than a prototype, agents work in Figma through its plugin API. They lay out wireframes and screens from the design system's components and variables and wire the prototype links. I review each screen and spend my time on the information architecture, the states a screen has to handle and what it leaves out.

My standards, written down

Without guidance, models fall back on the most common patterns. I write my standards down so each session starts with them.

Writing rules cut the patterns that make text read as generated, like em dashes and stacked colon lists. Design rules cover what I catch in review, from sentence-case labels to real photos of people instead of initials. When I correct something, I save it as a rule with the reason attached, so I only have to make that correction once.

---
name: no-all-caps-eyebrows
type: feedback
---

Never set eyebrows, section labels, table headers or small
captions in all caps. Use sentence case at medium or semibold
weight, one size smaller, in the muted text colour.

Why: the uppercase eyebrow is a common tell of
AI-generated UI, and I want it out of my work.

How to apply: label helpers never uppercase text.
Table headers use the body font at medium.

Review before it ships

Before anything ships, I run agents against it to find problems, then decide which findings are worth fixing.