CAPSLOCK is an independent AI innovation and design lab. That sentence is easy to say, so here is what it means on an ordinary day.

design is how we ask

Design isn't the coat of paint at the end. It's how we find out what a question really is: who it's for, what they would do with an answer, what "better" looks like before anything gets built. The loop a designer runs on an interface, we also run on a prompt, a workflow or a team habit. Stress-testing with synthetic users is that loop with a faster clock.

diagram: four materials, an interface, a prompt, a workflow and a team habit, all feed the same loop of three questions: who is it for, what would they do with an answer, and what does better look like, then back to the first. the loop yields the real question, and only then does something get built. stress-testing with synthetic users is the same loop with a faster clock
fig. the designer’s loop runs on any material: an interface, a prompt, a workflow, a team habit

we build to understand

Some questions only answer themselves once something is running. So we prototype early and small, put it inside real work, and watch what it does to the people using it. Nudge started as "how do we remind people about tasks" and only became a question about momentum once it was in use.

we write it down, including what didn't work

Every fieldnote ends with what carries, and most have a section called "the catch". A negative result is still a result. We sort the notes into thoughts, experiments and proofs so nobody mistakes a hunch for evidence. We don't invent numbers. When we don't have them yet, we say so.

flow: a result that worked and one that didn’t both become a fieldnote that ends with what carries and usually has a section called the catch. the note is sorted as a thought, an experiment or a proof, from hunch to evidence. then a check: do we have the number? if yes it goes in the note; if not yet, we say so. making one up is never an option
fig. what didn’t work gets a note too, and every note is sorted so a hunch never passes for evidence

people keep the judgment

AI drafts, proposes and checks. A person decides, especially where a step can't be undone or someone else's name is on the output. We split the doing from the checking (you can't let an agent mark its own homework), and we write down what failure looks like before the test, not after.

two-lane flow, a person above and ai below. first a person writes what failure looks like, before the test. then ai drafts and proposes. then ai checks the work separately, against that bar, so nothing marks its own homework. last, a person decides, especially when a step can’t be undone or someone else’s name is on the output
fig. a person sets the failure bar first and decides last; ai does and checks in between, never both at once

the loop

Question → prototype → operate → write it down.

  1. Question. Start from something we really don't know, phrased so a result could prove us wrong.
  2. Prototype. The smallest thing that can be used for real. Built fast with AI, judged slowly by people.
  3. Operate. Run it inside real work, long enough for the novelty to wear off.
  4. Write it down. A fieldnote: what we tried, what happened, where it's thin. That note is usually where the next question comes from.

It isn't a strict sequence. A proof can start a new thought, and a thought can stay untested for months. It's just the shape most of our work ends up taking.

cycle diagram of four steps. question: something we really don’t know, phrased so a result could prove us wrong. prototype: the smallest thing usable for real, built fast with ai and judged slowly by people. operate: inside real work until the novelty wears off. write it down: a fieldnote of what we tried, what happened and where it’s thin, which loops back to the next question. dashed side paths: a thought can stay untested for months, and a proof can start a new thought
fig. question, prototype, operate, write it down; the note is usually where the next question comes from

where from

CAPSLOCK is a Salvadoran company, built from El Salvador. The work itself could happen anywhere. What the place gives us is the setup: a tech law and export incentives that suit a lab like this, overlapping hours with the teams we work beside, and a reason to grow something at home instead of importing the future back at a markup. Why we build from El Salvador has the longer version.

how we use ai

AI is in almost everything we make, this site included. How we use it, where the line stays human, and how we handle data under El Salvador's laws has its own page: ai use.