Design Poiesis · a design methodology for the age of AIA design hypothesis, to be tested · Version 2 · October 2026

Every insight carries its trail.

Design Poiesis is a design methodology for the age of AI: five moves, two rails and a gate that turn informed intuition into design claims you can trace and test. It is written for designers, and for anyone doing structured creative work with AI. It is Diana Simpson-Hernandez's own method, formalised from fifteen years of design practice and the building of a technology venture from scratch.

The five moves ↓Run it with your class or studio →

Rail A · Sources of truth · tiered, tightening as the work converges

The ground: exploration, observation, experience, reading, data

M1SEIZECatch the prepared insight
M2EMBODYMetaphor → mechanism
⇄
M3INVERTConstraint → architecture
M4PROPAGATEValidity by travel
M5MINUTEThe same-day record
GATEWHAT’S NEXT?product · research · knowledge · seed · kill

DivergeConverge

EMBODY and INVERT loop until an inversion holds. Every minute becomes new ground for the next insight.

Rail B · Candour and radical honesty, on every readout

Why now

The case for a new design method.

Designers have been asked to record their reasoning for decades, and they mostly haven't. Design-rationale tools came and went, and they all ran into the same wall: writing down why you did something cost more effort than the work seemed to justify. So the reasoning stayed in people's heads, in sketchbooks, in meetings nobody minuted.

AI changes the cost. A model can capture the trail as you go, date it, tag it and link it, for close to nothing. That makes a method built on documenting the trail practical now in a way it wasn't before.

It also makes it necessary. AI generates ideas faster than anyone can remember where they came from. Without a trail, you can't tell your own insight from the model's suggestion, a tested claim from a plausible one, or a decision from a drift.

Design Poiesis is one answer to that: five moves, two rails and a gate, with one discipline running through all of it. Every insight carries its trail.

Design Poiesis doesn’t make the leap less intuitive. It makes it visible, so someone else can follow it, challenge it or build on it.

A design hypothesis, reached by abduction

Design Poiesis formalises an intuitive practice built over years of design work and, most recently, in building a technology venture from scratch. That intuition rests on earlier observation, experience, ideas and sometimes data. Philosophers call this kind of reasoning abduction: the leap to the most promising explanation from what you've seen.

The five moves

Five moves, one discipline.

Each move has a pass (has the trail been recorded?) and a signal (did the move change the design?). The dots show how much weight analogy carries in each one.

M1 · SEIZE ●●○

Catch the prepared insight

Insights feel sudden. They rarely are.

An insight is what's left after a long, informed exploration of a subject. SEIZE catches the insight and its ground the same day: what you'd been exploring, which sources and observations were in play, what tipped it over, and your first doubt.

Artefact: A capture card: insight · ground · trigger · first doubt.
The pass: Written within 24 hours, with at least one concrete ground and a named trigger.
The signal: It goes on into EMBODY. Captures that never move are counted, not hidden.
What the AI does: Dates it, lists the ground, asks what triggered it.
Where it goes wrong: Hoarding sparks, and the "sudden genius" myth that erases the work behind them.

M2 · EMBODY ●●●

Metaphor → mechanism

Push a metaphor until it predicts something.

Choose a concrete metaphor or analogy and push it element by element until the mapping tells you something you hadn't designed yet: a component, a rule, a way it could fail. Metaphors bring in solutions already worked out somewhere else, in biology, myth, engineering or other industries. Log every place the mapping breaks.

Artefact: A mapping table, its predictions, a break log, the domains you considered.
The pass: The mapping and break log exist, and why you chose this domain is written down.
The signal: Something in the final design traces back to a mapping line written before it.
What the AI does: Offers source domains at near, mid and far distances and drafts the mapping table.
Where it goes wrong: Falling for the metaphor and bending the system to fit the image.

M3 · INVERT ●●●

Constraint → architecture

What if this limitation were the product?

Every brief has a constraint it can't get away from. The usual move is to mitigate it. INVERT restates it as the organising principle: constraint C becomes principle P. Then take P back into the metaphor. If it breaks again, log the break and invert again. EMBODY and INVERT loop until an inversion holds.

Artefact: The C → P sentence, the decisions that follow from it, the loop log.
The pass: The inversion names its constraint and every loop turn is logged.
The signal: Take P away and the design can't be explained.
What the AI does: Red-teams it: is this an inversion, or a mitigation in disguise?
Where it goes wrong: Inverting everything. Some constraints are just costs. Invert the one that carries the load.

M4 · PROPAGATE ●●○

Validity by travel

A pattern earns trust by surviving away from home.

EMBODY brought a solution in from elsewhere. PROPAGATE sends the result back out. Carry the pattern into at least two unrelated contexts, record what maps and what breaks, and mark each row honestly: argued, or tested.

Artefact: A cross-context table: context · maps · breaks · verdict · argued or tested.
The pass: At least two rows, each marked argued or tested.
The signal: At least one row is tested, and a break changed the pattern.
What the AI does: Proposes far contexts and stress-tests the mapping.
Where it goes wrong: Propagating on paper only. A table is a claim about travel, not proof of it.

M5 · MINUTE ●○○

The same-day record

If it only exists in the conversation, it doesn't exist.

Turn the thinking into a dated record the same day. Tag every claim as fact, inference or opinion. Link the whole trail, from ground to SEIZE to the loop to PROPAGATE, and mark what the AI contributed. This is where AI earns its place: the minute that used to take an hour takes a minute.

Artefact: The minute: dated, tagged and linked.
The pass: On file with a date, warrant tags and links to the trail.
The signal: A later decision cites it.
What the AI does: Drafts it, tags it, links it.
Where it goes wrong: Minuting as theatre, where volume stands in for judgement.

The gate

What’s next? Not “build it”.

Before anything gets made, ask whether it needs to be made at all, and in what form.

Route	Ask	It needs
Product build and test	Does someone have this problem, and would they use this?	A named user, the smallest testable form, a kill criterion, a budget.
Research object study it	Is the open question worth more than an answer?	A question, a test written down before you run it, data you're allowed to use.
Knowledge teach or publish	Is the value in other people understanding it?	A disclosure check, an audience, a format.
Seed park it	Is it right but early?	A revisit trigger: an event or a date.
Kill stop	Is it wrong, done before, or not worth it?	The reason, written down. A kill counts as a pass.

A health check: count your kills and parks. If everything becomes a product, the gate isn’t working.

The two rails

Guardrails that let you diverge.

Rail A · Sources of truth

Before you start, declare what the work has to answer to. Guardrails are often applied at the same strength from start to finish, which either strangles early ideas or lets late decisions go soft. Design Poiesis tiers them, and turns them up as the work converges.

Tier	Holds	Diverging	Converging
T1 · Commitments	values, ethics, law, disclosure, budget	binds	binds
T2 · Evidence	verified data, primary sources	informs	binds
T3 · Working beliefs	assumptions, conventions, all AI output	open to challenge	proved, or labelled as assumption

Only sources that pass five checks get into T2: provenance, verifiability, independence from the idea under test, recency and scope. T1 is dated before the first SEIZE; guards written afterwards are rationalisation.

Rail B · Candour and radical honesty

Radical honesty here is about claims, not about being blunt with people.

Bad news first. Every readout opens with what's wrong, or says plainly that nothing is.
Fact, inference and opinion are kept apart wherever they mix.
"I don't know", "not done" and "we were wrong" are complete sentences.
No unearned praise. Say it's good when the evidence says so, and name the evidence.
Honest about the method too. Design Poiesis reports its own limits.
The duty to disagree. Anyone, human or AI, who sees a weaker choice or an unstated risk says so before the work goes ahead. The owner decides, and the disagreement is recorded.

Working with AI

AI widens the search. You hold the signature.

AI widens the search and keeps the minutes. You hold the judgement, the taste and the signature.

Rule 01

Diverge alone first. Seeing AI examples early can anchor you on them. Your own ideas go down first.

Rule 02

AI proposes, you decide. It never signs a ruling or declares what's non-negotiable.

Rule 03

AI output starts as a working belief. Check every fact and citation against a primary source.

Rule 04

Tag every contribution. [D] designer · [A] AI · [DEC] decision.

Rule 05

Ask it to disagree. Agreement without reasons is noise.

Rule 06

Mind what you paste. Confidential material only goes into tools you've approved.

Rule 07

Say AI was used in anything you publish.

Rule 08

It works with a notebook. AI makes the method cheap. It isn't required.

Worked example · illustrative, fictional

The setup nobody finished.

A made-up business case, used to show the moves in order. It isn't a client project.

The brief
A bookkeeping app for small businesses, with an AI assistant built in. Six in ten new customers never finish setting it up, so they never see it work, and most cancel in the first month. The ask: get more customers through setup without hiring an onboarding team.
Rail A
No extra headcount for onboarding calls, no customer data shared with third parties, no promise the product can't keep (T1). Product analytics on where people drop off, and fifteen customer interviews (T2). "They need more tutorials" and "a smarter chatbot will fix it" (T3, open to challenge).
SEIZE
After weeks of interviews and session reviews, the designer notices that every customer who finished setup did it with someone else in the room, usually their accountant. Trigger: three owners in one week said they were "waiting for the accountant". Insight: setup is a shared task, not a learning task.
EMBODY
Candidates: a hotel concierge (near), a dual-control driving-school car (mid), a fungal network that links trees underground (far). The car wins. Two people, two sets of controls, and the instructor takes over only for the dangerous moments. It predicts that an assistant which does everything leaves owners unable to run their own books. Break: the instructor can see the road, but the AI can't see the paperwork on the owner's desk.
INVERT
No headcount for onboarding becomes: the customer's own accountant is the onboarding team, invited in with their own half of the setup. Loop two: the AI can't see the paperwork becomes: the AI only asks for what it can check, and hands the rest to the accountant as a checklist. The car holds: two people, two sets of controls.
PROPAGATE
A gym induction with a training partner (argued: maps well, breaks on motivation). Opening a business bank account with an adviser (tested: one small trial with a partner bank).
MINUTE
Written the same day, with the AI's suggestions tagged. "Setup is a shared task" is an inference from fifteen interviews, so it's labelled that way.
What's next
Product: a two-week test of invite-your-accountant setup with twenty new accounts, measured on setup completion and first-month cancellations. Kill criterion, written down before it starts: if paired accounts don't finish setup more often than the rest, stop.

What it adds

For designers, and for design method.

For designers and teams

Intuition you can defend, in a client room, a funding panel or a viva.
Analogy as a craft, chosen across distances, with its breaks put to work.
Constraints treated as assets, on purpose.
Guardrails that don't kill ideas, tiered and dialled by phase.
A decision at the end, not a default to build.
Practice that compounds, because every minute becomes new ground.
Clear roles with AI, so authorship stays with you.

What it proposes for design method

The trail as the pass condition: abduction gets a record.
Rationale capture made affordable by AI-assisted, same-day minuting.
Guardrails that change with phase, through tiered sources of truth.
Analogy managed as a spine: distance-tagged, break-logged, run in reverse.
Explicit human and AI roles inside the method itself.
Reframing as a named move with a pass test.

What it doesn't add is a new theory of creativity. Each move has close relatives in design research: generative metaphor, frame creation, inventive problem-solving, design-rationale capture. The contribution is how they fit together, and the trail.

Honest limits

What it claims, and what it doesn’t yet.

It claims

Follow the moves and record the trail, and your design claims will be traceable to their grounds, tested against what matters, carried across contexts before they're trusted, and routed to the right next form.

It doesn’t claim, yet

That it works for everyone, or better than a strong design process you already have. It is formalised from one practice, and it will be tested before any stronger claim is made.

How it will be tested

Formalised from one practice, after the fact, is a real limit: anyone describing their own process finds it tidier than it was. So the test plan has three parts, and the results will be published, including what doesn’t hold.

Test 1

Looking back

Score a dated archive of past ideas, including the ones that died. If almost every idea passes every move, the tests aren't tests.

Test 2

Live

Run the next fresh idea through the moves, against a ten-minute "what I'd have done anyway".

Test 3

In someone else's hands

A class session in which students run the moves from a single card.

The name

Why poiesis.

Ποίησις (poiesis) is the Greek for bringing something forth, the root of the word poetry. Heidegger paired it with aletheia, truth as unconcealment. Poiesis is the making that brings something into the open.

Use it

Teaching it, or trying it in a studio?

Design Poiesis runs as a lecture, a half-day workshop or a studio session, in English or Spanish. If you run the moves, the trail you keep is the most useful feedback there is.

Email Diana →