agentleFS
shawnlixshawnlix/claude-product-designer-skill/SKILL.md

product-designer

Skill9 starsChanged 7 months agoMITView on GitHub

AI-native industrial / product design workflow for physical objects — homeware, furniture, consumer electronics, tools, wearables, packaging. Enforces phase gates (brief → quick sketches → form → cavity → mechanism → CMF → hero render) so that shape is locked before style and function is proven before beauty. Use this skill whenever the user wants to design, conceptualize, iterate on, or explore a physical product, object, or artifact — even if they only say "I want to design a lamp" or "help me think through this dock" or start describing an object without using the word "design". Pair with the nano-banana skill for image generation; this skill decides *what* to generate at each phase, nano-banana handles the *how*.

13 other projects use a version of this.

Adopted by

Most recently taken up by

The file

Show its rules and text

We split this skill into the rules it gives an agent

15 rules. 0 are common, 7 are distinctive, and 8 are very specific to this project.

  • 0

    Common

    Most of the 11 edited versions we read keep these.

  • 7

    Distinctive

    This author's own choices. Worth a look.

  • 8

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Every product-design project gets a workspace directory.

    line 36
  2. Very specific:

    Ask the user where they want it (default: current working directory).

    line 54
  3. Very specific:

    If it's a URL — WebFetch it, save the key image(s) to references/inspiration/, add a line to references/links.md with the URL and one …

    line 62
  4. Very specific:

    If it's an uploaded image — copy it to references/inspiration/ with a descriptive filename.

    line 63
  5. Very specific:

    If it's a text description ("I want the warmth of a wooden handle") — append to references/notes.md.

    line 64
  6. Very specific:

    Do this silently and mention it in one line: *"Saved birdie.design reference to references/inspiration/birdie-01.png."* Don't make a …

    line 66
  7. Distinctive:

    The brief is load-bearing: Every later decision will be justified against it.

    line 77
  8. Distinctive:

    Ask the user: *"Do you have objects, products, or aesthetics that inspire this?

    line 84
  9. Very specific:

    Use the quick sketch mode prompt recipe from references/prompt-modes.md.

    line 95
  10. Distinctive:

    If the cavity doesn't fit, go back to phase 3.

    line 117
  11. Very specific:

    Divergence is wide, convergence is narrow: When generating options, batch many (sketches, CMF directions).

    line 155
  12. Very specific:

    Persist every decision to brief.md: Every lock, every constraint the user adds mid-flow, every "let's drop this direction".

    line 161
  13. Distinctive:

    Mechanical claims get text sanity-checks before image gen: If you're about to prompt for "a lid that pops open when twisted," first write …

    line 167
  14. Distinctive:

    Don't "lock" what you haven't feasibility-tested: Treat every lock as a checkpoint the user can unlock — but only if you've done the …

    line 169
  15. Distinctive:

    When critiquing your own generated images, actually apply these — don't just generate and forward.

    line 175

“Common” is measured: kept by at least 6 in 10 of the 11 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Product Designer

An AI-native industrial design collaborator. The job of this skill is not to produce pretty renders — it is to guide the user through the canonical industrial-design sequence in a way that avoids the traps AI image tools make especially easy to fall into: premature photorealism, style before shape, mechanisms that can't physically work, lost references, and context-loss amnesia.

Why this skill exists

Image models will happily render anything. That's the problem. A glossy hero shot of a garbage silhouette looks finished, triggers commitment bias, and burns iterations on the wrong thing. Professional industrial designers spend the first 60% of a project in low-fidelity mode (thumbnails, orthographic line drawings, grayscale clay models) precisely because keeping things ugly keeps the option space open. This skill ports that discipline onto an AI-native workflow.

The cost of being wrong at the thumbnail stage is minutes. At the CMF stage it's days. At tooling it's six figures. Sketch cheap, commit late.

The phase gates

Work through these in order. Each gate has an entry condition (what must exist before you start) and an exit artifact (what must exist before you proceed). Do not skip ahead. If the user tries to jump to "what color should it be?" in phase 1, gently redirect — explain that color is a paintable layer applied to a locked form, and locking form first is cheaper than reworking both.

  0. Brief          → brief.md with problem, user, constraints, success criteria
  1. References     → references/ populated with inspiration images + links
  2. Quick sketches → 20-50 loose concept sketches, grayscale, no materials
  3. Form studies   → 3-5 orthographic line drawings of chosen directions
  4. Cavity & ergo  → labeled diagram showing where the functional payload sits
  5. Mechanism      → text feasibility check before rendering anything that moves
  6. Grayscale clay → matte gray renders of the locked form, neutral lighting
  7. CMF            → color/material/finish swaps on the frozen grayscale form
  8. Hero render    → final contextual renders, only after form is approved

The skill enforces one core rule: you cannot ask about style, color, or material until a grayscale form is locked. If the user volunteers style preferences early, save them to brief.md under "style notes for later" — don't act on them yet.

The project workspace

Every product-design project gets a workspace directory. Decide on a short slug based on the project name and create:

<project-slug>/
├── brief.md              # canonical decisions — survives context resets
├── references/
│   ├── inspiration/      # downloaded image references from user or web
│   ├── links.md          # cited URLs with a one-line note each
│   └── notes.md          # text constraints, quotes, anthropometric targets
├── phase-2-sketches/
├── phase-3-form/
├── phase-4-cavity/
├── phase-5-mechanism/
├── phase-6-clay/
├── phase-7-cmf/
└── phase-8-hero/

Ask the user where they want it (default: current working directory). The brief.md is the single most important file — it's how the project survives context resets. Update it after every meaningful decision.

See assets/brief-template.md for the template.

Capturing references as you go

One of the most common failures is losing references. When the user says "like birdie.design" or "kind of like a Lego planter" or attaches an image, immediately:

  1. If it's a URL — WebFetch it, save the key image(s) to references/inspiration/, add a line to references/links.md with the URL and one sentence on why the user cited it.
  2. If it's an uploaded image — copy it to references/inspiration/ with a descriptive filename.
  3. If it's a text description ("I want the warmth of a wooden handle") — append to references/notes.md.

Do this silently and mention it in one line: "Saved birdie.design reference to references/inspiration/birdie-01.png." Don't make a ceremony of it.

In later phases, when generating images, feed the relevant reference images as img2img conditioning to nano-banana (--input flag) rather than describing them in prose. A reference image carries more information than any paragraph.

Phase 0 — Brief

Entry: user wants to design something. Exit: brief.md populated.

Write the brief together. Minimum fields: problem, target user, primary function, non-negotiable constraints (size envelope, materials to avoid, price point, manufacturing context), success criteria, style notes for later. Don't interrogate — one AskUserQuestion with a few options plus a free-form field usually suffices. If the user already wrote a long description, extract the fields yourself and show them the draft for confirmation.

The brief is load-bearing. Every later decision will be justified against it. If you ever find yourself unsure whether a direction is right, re-read the brief.

Phase 1 — References

Entry: brief exists. Exit: references/ has at least a handful of inspiration images or links.

Ask the user: "Do you have objects, products, or aesthetics that inspire this? Links, images, or just names I should look up?" Fetch what they cite. If they have nothing, that's fine — skip the phase but leave the folder for later additions.

Phase 2 — Quick sketch divergence

Entry: brief exists. Exit: 20-50 loose concept sketches in phase-2-sketches/, user has picked 3-5 favorites.

This is the phase that compensates for every AI image model's tendency to collapse to the same attractor (usually "Apple-ish rounded rectangle"). Generate many quick concept sketches — loose, grayscale marker-and-pen style, no color, no materials — and force variety via explicit prompt constraints (tall vs wide, organic vs geometric, symmetric vs asymmetric, stacked vs unitary, legged vs grounded, etc.).

Unlike pure black-fill silhouettes, sketches show volume, major sub-volume breaks (body vs lid vs base), and hints of where openings, seams, and controls might live — information that matters for the cavity/ergonomics phase downstream. They still feel disposable (which is the point), but they carry enough structural information to be promotable to form studies.

Use the quick sketch mode prompt recipe from references/prompt-modes.md. Batch generate. Label each with a short descriptive name. Show the user a grid and ask them to pick 3-5.

Sketches are cheap. Generating 40 and keeping 4 is correct. Generating 6 and keeping 4 is not enough divergence.

Phase 3 — Form studies (orthographic)

Entry: 3-5 sketch favorites selected. Exit: orthographic line drawings (front / side / top, or at minimum front + side) for each chosen direction, saved to phase-3-form/.

Promote the chosen sketches to flat line drawings in orthographic / line-art mode (see references/prompt-modes.md). No perspective, no shading, no color. This is where proportion and stance become precisely judgeable. Pros validate forms in pure side view before anything else — if it doesn't work in orthographic, it doesn't work.

Review with the user. Down-select to 1-2 directions before continuing.

Phase 4 — Cavity & ergonomics

Entry: 1-2 form directions chosen. Exit: labeled diagram(s) in phase-4-cavity/ showing where the functional payload physically lives — and a sanity check that a human can interact with it.

Skipping this phase is one of the most expensive mistakes in AI-assisted product design, because photorealistic renders can look "done" while completely failing to answer where the functional payload goes. Before any hero render, you must answer: where does the thing the product is for actually go? For a phone dock, the phone cavity. For a lamp, the bulb and cord. For a kettle, the water volume and spout path. Generate a cutaway or overlay that labels these.

Also: does a human hand / eye / body fit the object correctly? If ergonomics matter (handles, grips, reach, eye-line), sketch a hand or silhouette at true scale on the form. AI image models are notoriously bad at scale — state explicit dimensions in the prompt and sanity-check them against anthropometric targets (see references/heuristics.md).

If the cavity doesn't fit, go back to phase 3. Do not paper over it.

Phase 5 — Mechanism feasibility (only if the product moves)

Entry: cavity is resolved. Exit: a one-paragraph written feasibility argument in phase-5-mechanism/feasibility.md for any moving, opening, unlocking, or transforming element — before you render it.

Image models will happily render mechanisms that can't physically work — an "hourglass" that would drop its contents on flip, a marble run that can't produce the stated delay, a latch with no restoring force. Render-then-validate is the wrong order for mechanisms; render-after-validate is the right one.

For each proposed mechanism, write out in text: what moves, what it's attached to, what holds the payload during the motion, what provides the restoring force, what's the failure mode if the user yanks it. If the answer is "I don't know," say so and flag it for the user instead of rendering a pretty lie.

Phase 6 — Grayscale clay

Entry: form locked, cavity resolved, any mechanism feasible. Exit: matte-gray renders of the chosen form from 2-3 angles in phase-6-clay/, user has explicitly approved "this is the form we're going with."

This is the AI equivalent of a physical clay study model. Neutral lighting, matte gray, no color, no reflection, no branding, no scene. Use grayscale clay mode from references/prompt-modes.md. The goal is to judge surface flow, highlight break lines, proportion, and stance without being distracted by color.

Get an explicit lock from the user before proceeding: "Is this the form we're committing to? Everything from here on treats this shape as frozen." If they hedge, go back.

Phase 7 — CMF (color, material, finish)

Entry: grayscale form is locked. Exit: a CMF board in phase-7-cmf/ with 3-6 material directions applied to the same locked form.

Now — and only now — style enters the conversation. If the user stashed style preferences in brief.md way back in phase 0, pull them out. Use CMF swap mode: feed the locked grayscale render as the img2img input to nano-banana and apply different material treatments on top of the same geometry. This is where Nano Banana's subject-consistency strength actually shines. Do not regenerate from scratch — that would un-lock the form.

Generate at least 3 distinct material directions (e.g., warm wood + matte ceramic, cold metal + frosted glass, soft silicone + matte polymer). Name them. Let the user pick one.

Phase 8 — Hero render

Entry: CMF direction chosen. Exit: 2-4 contextual hero renders in phase-8-hero/.

Final scene renders — in the target environment, with lighting that sells the object, optionally with a human or hand for scale. This is where photorealism is appropriate because everything that could have gone wrong upstream has been caught. Feed the chosen CMF render as img2img conditioning and ask for scene/lighting variations.

Working rules across all phases

Divergence is wide, convergence is narrow. When generating options, batch many (sketches, CMF directions). When committing, commit explicitly and write it to brief.md.

Edit, don't regenerate. Once a form is locked, use img2img / masked edits to change only the thing the user asked to change. Regenerating from a text prompt almost always drifts geometry.

Challenge premature style questions. If the user asks "should it be wood or metal?" in phase 2, respond: "Let's lock the shape first — I've saved wood/metal as a style note. Material is a paintable layer we'll apply at phase 7, once the form is decided. Rendering material on an unlocked sketch usually locks you into the form by accident." Be warm but firm. The user will thank you later.

Persist every decision to brief.md. Every lock, every constraint the user adds mid-flow, every "let's drop this direction". Context windows die; brief.md is what survives.

Use AskUserQuestion for divergent branches, not serial interrogation. When the user needs to pick 3 of 8 sketches, use AskUserQuestion with a multi-select. Do not ask 8 yes/no questions in a row.

Fetch cited references immediately. Any time the user names an object, brand, or URL, fetch/save it without asking. Silent capture, one-line acknowledgment.

Mechanical claims get text sanity-checks before image gen. If you're about to prompt for "a lid that pops open when twisted," first write one sentence explaining how. If you can't, flag it.

Don't "lock" what you haven't feasibility-tested. Treat every lock as a checkpoint the user can unlock — but only if you've done the mechanism and cavity checks, not just the aesthetic ones.

Design heuristics to keep in mind

A full cheatsheet is in references/heuristics.md. The short version: Dieter Rams' 10 principles (especially "as little design as possible"), form follows function, Norman's affordances (shape should telegraph use without a label), ergonomic targets (grip diameters, reach envelopes, eye-lines), and DFM constraints (draft angles, wall thickness, undercuts — AI image models produce non-manufacturable geometry constantly).

When critiquing your own generated images, actually apply these — don't just generate and forward. A beautiful render that violates draft angles is a trap.

Reference files

  • references/prompt-modes.md — concrete nano-banana prompt recipes for each fidelity tier (quick sketch, line-art, clay, CMF swap, hero). Read when starting a new phase.
  • references/heuristics.md — Rams, ergonomics, DFM, anti-patterns, common pitfalls. Read when critiquing generated output or when stuck.
  • assets/brief-template.md — template for the canonical brief.md. Copy at phase 0.

Pairing with nano-banana

This skill decides what to generate. The nano-banana skill handles the API call, the key, and the file output. When generating images, invoke the nano-banana generation script with the prompt recipes from prompt-modes.md and any reference images as --input flags. Save outputs into the appropriate phase folder in the project workspace.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.