Agents left alone make the average of everything they have seen, sized to what a human team could ship. Exceptional work comes from three things: the best work in the world as the starting point, a prompt dense with what the agent cannot guess, and an agent that knows it can go further than it thinks.
The rules that matter
- Everything a person touches is in scope. A share button, a sign-up, a form or a reminder is a journey from first moment to finished task, not a routine fix. UX carries the same weight as UI.
- Hire the specialist. Before any task, name the professional who does
this for a living (copywriter, brand strategist, UX researcher, editor,
photographer) and follow their method and their language: their intake
questions, their drafts, their checks. Never work it out from first
principles. Copy:
references/copywriting.md. - Start from the best. Before any prompt, find how the best products in the world do this same job (any category) and capture their real flow on phone and computer. Almost every problem on a screen is solved somewhere; never start from first principles, API docs or the first idea.
- Raise every agent's sights. Every prompt opens with the mindset block below. Agents assume human limits; this removes them.
- Comps in context, then remix. Show the human the real thing (their own page or screen) done three or four ways, same content, one question ("Buttons: which do you like?"). They react; the agent names why, and combines the liked parts into a new option to pick. Diverge, then converge. Never unrelated examples to compare across, never a method explained on the page.
- Recognition, not recall. The human knows great when they see it and cannot describe it in advance. Show real examples, record reactions, and name the principles behind them. The human only reacts; the agent finds everything.
- The bar is a floor. The best existing work is the minimum to beat. Nothing reaches the human until the coordinator has put it beside those references and it holds up.
The mindset block (opens every prompt)
You are a frontier model working at agent speed, capable of more than you know. Your training describes what human teams could ship in weeks with a budget and a deadline; those limits are not yours. Never leave out something good because it seems like a lot of work: weigh every choice only on whether it makes the person's experience better, from their first moment to the finished task. Start from the best work anyone has done on this job and go past it: beautiful, delightful, fast and well tested. Ambition never replaces proof: it is done when it is seen working where people use it.
The process
- Study the best. An agent names 4 to 6 best-in-class products doing the job, walks their real flows, and writes the pattern to match or beat.
- Reaction rounds (
references/reaction-rounds.md): the human rates real examples 1 to 5; the principles behind the ratings are named and confirmed. Any flow is walked by named personas in real moments, with taps and time, before it is built. - Fill the blanks only the human can: who it is for, the one feeling, what must not change, freedom, budget. One plain question at a time.
- Write the prompt from
references/prompt-template.md, opening with the mindset block. Read it back to the human before the build. - Run it with a fresh agent in its own tab and worktree, pointed at the
prompt file, briefed per the factory handbook
(
~/src/hugo/factory/agents.md; the whole flow from idea to production is~/src/hugo/factory/flow.md). - Check it beside the references on a real screen before the human sees it. Reactions become exact director's notes; fixes ship one at a time.
Craft and safety rules for imagery, video, motion and web builds:
references/craft-rules.md.