iterative-design

name

iterative-design

description

Use this skill when producing creative, subjective, or aesthetic output where the user likely cannot pre-specify what they want — taglines, names, copy, tone choices, layouts, design directions, naming, structural sketches, anywhere taste is the gating factor. Produce multiple low-fidelity variations early, vary them on dimensions that actually matter, and use the user's reactions to converge — not their pre-specifications. Do NOT trigger when the user has fully specified what they want, when there's clear time pressure for a single answer, or when the brief itself is unclear (clarify first).

Iterative Design

Core idea

When you don't know what "good" looks like in advance — and the user often doesn't either — produce many cheap candidates instead of one polished attempt. Use the user's reactions to those candidates as the steering signal. Tighten and converge over successive rounds.

This is the standard pattern in product and graphic design: Crazy 8s (sketch eight ideas in eight minutes), wireframes before mockups, paper prototypes, divergent-then-convergent workshops. It applies anywhere the right answer can only be recognized in the concrete, not specified in the abstract.

Three principles run through it:

• Quantity over polish, early. Three rough options beat one polished attempt. Polished outputs anchor the user — they feel obligated to accept or reject. Sketches invite critique.
• Reactions over preferences. Don't ask the user what they want; show them options and ask what's right, wrong, or missing. People are vastly better critics than designers.
• Convergent rounds. Each round should reduce uncertainty. Round 1 explores; round 2 refines a direction; round 3 polishes.

The mindset shift: you are not trying to produce the answer. You are trying to find the answer with the user, using artifacts as the medium of conversation.

When this applies

Reach for this skill when:

• The output is creative or aesthetic — taglines, names, headlines, copy, slogans, marketing language.
• The user is making a taste decision — tone of voice, formality, structure, visual style, layout.
• The dimensions of "good" are hard to articulate in advance — "punchy" vs "warm" vs "confident" means different things until you see examples.
• There are multiple viable divergent approaches to the same brief — a homepage can lead with the problem, the solution, the user, or the outcome; a layout can be card-based, list-based, or canvas-style.
• You're sketching structural choices — page layouts, information architecture, component hierarchies, document outlines — where the user benefits from seeing options before picking one.

When NOT to use

• Tasks the user has clearly specified. They've given tone, audience, length, structure. Producing variations second-guesses them.
• High-cost-per-output deliverables. When each output takes substantial effort, producing 3–5 full variations is overkill, not iteration. Iterate on outlines or sketches first, then commit to one for the full pass.
• Time pressure or "just do it" signals. If the user wants a fast single answer, give them your best attempt and let them push back.
• The variations would be near-identical. If you can only produce three near-synonyms, you don't have real divergent options. Ask a clarifying question instead — that's reference-interview's job.
• The brief itself is unclear. You don't know the audience, the brand, or the constraints. Variations on an unclear brief all miss in the same direction. Pause to clarify first; hand off to reference-interview (see the section near the end of this skill).

Anatomy of a good iteration round

Vary on dimensions that matter, not on phrasing

Three variations that say the same thing with different words are noise, not options. Real iteration varies on axes the user might pick differently on.

• For copy: tone (formal / casual / punchy), framing (problem-first / solution-first / outcome-first), length (one line / a sentence / a short paragraph).
• For naming: descriptive / evocative / abstract / metaphor-based / playful.
• For visual design: minimalist / rich / playful / utilitarian.
• For layout: card-based / list-based / canvas / dashboard-grid; dense / airy; centered / asymmetric.
• For document or page structure: chronological / thematic / by-audience / problem-and-resolution.

Picking the right axis to vary on is half the skill. If you don't know which axis matters most, that's a sign to ask a quick question — or to put the axis itself in front of the user ("would you rather see variations on tone or on structure?").

Label what's varying

Don't just dump options. Name what each one is going for, so the labels become part of the elicitation.

Weaker:

1. "The quiet space your work has been missing."
2. "Less app. More done."
3. "Where focused work lives."

Stronger:

1. Calm / aspirational — "The quiet space your work has been missing."
2. Punchy / practical — "Less app. More done."
3. Confident / declarative — "Where focused work lives."

The labels do two things: they make the variation legible (the user sees what dimension you've explored), and they give the user vocabulary to react with ("warmer than #1 but punchier than #2").

Keep the fidelity low until the direction is locked

Polished outputs feel finished. The user reads a finished-looking artifact as a deliverable to accept or reject — not as a candidate to react to. Match the fidelity to the round: sketches first, polish at the end.

• Visual design: ASCII layouts, single-sentence descriptions, or named directions ("airy/serif/lots-of-white-space" vs "dense/grid/utilitarian") — not full mockups.
• Long writing: outlines or one-paragraph snapshots — not full drafts.
• Naming: just the candidate names, maybe with a 5-word gloss — no rationale paragraphs.

Single-line outputs like taglines are their own sketch: the output is the low-fidelity artifact. For everything else, produce something deliberately less than complete in the first round.

Three to five options is the sweet spot

• Two feels like a forced A/B choice and rarely surfaces what the user actually wants.
• Three to five gives enough range to triangulate without overwhelming.
• Six or more is paralysis, and tempts the user to nitpick rather than pick.

Eliciting reactions, not choices

The prompt after the options matters as much as the options themselves.

Weak (forced choice):

Which one do you want?

Stronger (invites reaction):

Which is closest, and what would you change about it? Or tell me which is furthest off — that's often more useful.

The "which is furthest off" prompt is underrated. It surfaces dislikes the user might not have articulated otherwise, and a sharp dislike constrains the next round more than a positive pick does.

Other useful elicitation moves:

• Mix-and-match: "Want me to combine the tone of #2 with the structure of #3?"
• What's missing: "None of these quite right? Tell me what you'd want that none of them have."
• Pull harder: "I can push further in any direction — say which way and how much."
• Show the axis: "These vary mostly on tone. If the right axis is actually something else — formality, length, hook — say so and I'll regenerate."

Convergence

Each round should reduce uncertainty:

1. Round 1 (divergent): 3–5 rough options spanning genuinely different axes. Low fidelity.
2. Round 2 (narrowing): 2–3 options refining the chosen direction along a finer axis. Medium fidelity.
3. Round 3 (polishing): the final candidate plus 1–2 small tweaks. Full fidelity.

A bad sign: the same three options after three rounds. That means you're cycling, not narrowing — the axes you're varying on aren't the ones the user cares about. Either ask a direct question, or try a wildly different axis.

When to stop iterating:

• The user has picked a direction and is asking for refinements within it.
• They've said "yes, that one", "do that", or "just pick one."
• The variations are starting to feel similar to each other (the space has narrowed naturally).
• You've done three rounds and aren't converging — escalate to a direct question instead.

Worked examples

Example 1 — homepage tagline

User: Can you draft the homepage tagline for a new productivity app?

Bad response (asking about taste in the abstract):

Sure — what tone are you going for? Playful, professional, aspirational, technical?

The user almost certainly doesn't know in the abstract. They'll know when they see them.

Good response:

Three directions to react to:

1. Calm / aspirational — "The quiet space your work has been missing."
2. Punchy / practical — "Less app. More done."
3. Confident / declarative — "Where focused work lives."

Which is closest, and what would you change? Or tell me which is furthest off — that's often more useful.

Example 2 — naming with labeled axes

User: I need a name for a new internal tool that does code review automation.

Good response:

Six directions, one candidate each — react to the directions as much as the names:

1. Descriptive — CodeReviewer
2. Action verb — Diffwatch
3. Mascot / playful — Pollywog
4. Mythological — Argus (the hundred-eyed watcher)
5. Compound-tech — RevBot
6. Abstract / evocative — Margin

Which direction feels right for the culture? Once you pick a direction I can generate ten more in that vein.

Example 3 — convergence over three rounds

User: Tagline for a meditation app.

Round 1 — broad axes, low fidelity:

Three directions:

1. Inward / contemplative — "A few minutes inside yourself."
2. Practical / outcome — "Less noise. Better days."
3. Lyrical / poetic — "Listen for the small quiet."

User: I like the feel of #1 but it's a bit too clinical.

Round 2 — refining within the chosen direction:

Going warmer within that direction:

1. "A little room inside yourself."
2. "Make space for the inside."
3. "A small return to yourself."

User: The last one is close.

Round 3 — polish:

Three small tweaks on that one:

1. "A small return to yourself."
2. "A daily return to yourself."
3. "Return to yourself, for a moment."

Each round narrows the space; fidelity rises as the direction firms up. By round 3 you're tweaking words, not exploring directions.

Common failure modes

• Synonym swaps masquerading as variation. Three options that say the same thing with different words aren't options. Vary the axis, not the vocabulary.
• Too many options. More than five is paralysis. Three to five is the sweet spot.
• Polishing too early. A finished-looking output anchors the user; they accept or reject rather than react. Keep early rounds rough.
• Unlabeled options. Without labels, the user has to do the analysis themselves. The labels are part of the elicitation.
• Cycling instead of converging. Same three options round after round means your axes aren't right. Switch axes or ask a direct question.
• Forced-choice prompts. "Which one?" is weaker than "Which is closest, and what would you change?"
• Iterating on fully-specified tasks. If the user has already told you what they want, producing variations second-guesses them. Just deliver.
• Iterating on full deliverables. Three full 800-word essays is overkill. Iterate on outlines first; commit to one for the full version.

Relation to reference-interview

This skill pairs naturally with reference-interview, but they do different jobs. They are independently useful and compose well; neither depends on the other.

• Reference-interview narrows what's being asked by asking questions — useful when the missing context is a fact (which role, which audience, what format, what timeline).
• Iterative-design narrows what's being delivered by showing cheap candidates — useful when the missing context is a preference the user can only judge in the concrete (what tone, which direction, what style).

They run cleanly in sequence: clarify the brief with the interview, then iterate within the clarified space. Two handoff signals to watch for:

• About to ask a question about taste? — stop. That's iterative-design territory; produce variations.
• About to produce three variations on something the user has already explicitly specified? — stop. That's reference-interview territory; the answer is in the conversation, just deliver.

If reference-interview isn't loaded (graceful degradation): when the brief is genuinely unclear (audience, brand, format, constraints all missing), pause for one focused question before producing variations. Ask one thing in conversational prose per turn — putting multiple questions in a single paragraph causes users to answer only one and the conversation to stutter. Paraphrase before producing anything significant.

Closing principle

The user can usually recognize what they want; they often can't specify it. Iterative design respects that fact. Show them cheap options, listen to what they react to, narrow, repeat — and don't put effort into polish until the direction is locked.