The tiny tool category; spec-first; the build loop; inspection; the scope fence.
All tracks · especially UX, motion, graphic design, 3D — 4 hours, drop-in friendly
Coding for Artists: Websites, Tools, and Tiny Apps with Codex/Claude Code
Turn a recurring annoyance in your practice into a working tiny tool — spec-first, built by AI, inspected by you. Zero coding experience required.
01This Week at a Glance
You will make: a tiny tool spec — user, problem, inputs, outputs, UI sketch, build prompt. Building it is the bonus round; many of you will.
A working app built live from a spec the room writes — including the inspection and the planned failure.
Your friction → your spec → (your tool).
60 seconds each, with honest ledgers.
Not to become software engineers — code does four things for artists: prototype, automate, publish, make interactive work. The barrier collapsed; the bottleneck moved from "can you code?" to "can you specify?" — and specification is design thinking. The spec is the skill; the build is increasingly free.
02Key Concepts
1 · The tiny tool category (where artists win)
One user (you), one problem, one file if possible, no accounts, no server, no other people's data. Examples students actually use: a moodboard prompt generator, an export checklist app, a riso separation previewer, a commission price calculator, a crit-notes formatter. Single-file HTML runs in any browser, can't touch anything else, and is fully inspectable — the most reliable, lowest-risk zone for AI coding.
2 · Spec-first: the interview method
Don't start with "build me an app." Make the model interview you — one question at a time, max 6. Your answers become the spec: user, problem with honest frequency, inputs (≤6), outputs, UI in words with exact button labels, what it does NOT do (the feature-creep fence), and acceptance checks (3–5 testable sentences). Template: tiny-tool-spec-template.md.
"Wrote the spec, directed the build, verified the result." Current, honest, hireable — and exactly how you'll work with human developers later.
3 · The build loop: client-revision thinking
One prompt builds the file. Open it, use it. Then: change requests describe outcomes ("the Copy button should visibly confirm"), and bug reports describe steps and symptoms. Same spirit as Blender week's error loop — you report, it repairs.
4 · Inspection habits (the trust layer)
After any build, three questions: what are the three main parts, where does my data go when I type it, does anything leave my computer? For a true tiny tool the answer is "nothing leaves your machine" — and you can verify it: skim the file for fetch or http. Ten seconds.
Privacy became checkable the day software became readable conversation.
5 · The scope fence
The moment a tool needs accounts, payments, or other people's data — it's not a tiny tool. App-builder platforms (Lovable/Bolt/v0-class) will happily build past the fence; the inspection habits stay mandatory and the risks multiply.
6 · The deeper route: creative coding as craft
Tiny tools are utility; creative coding (p5.js and friends) is art-making with code. AI is a patient tutor for it — and if today lights a fire, The Coding Train is where it goes next.
03This Week's Words
04Demo: The Moodboard Prompt Generator, Built Live
The room writes the spec by answering the model's interview; the spec becomes a working app; the app survives a change request, an inspection, and a planned failure.
Instructor backup / finished example: moodboard-generator.html. Use it if the live tool run fails, or to reveal the finished structure after critique.
- The interview: the model asks; the room answers; the instructor arbitrates. Six questions → a complete spec.
- The build: spec + build prompt → saved as
.html→ opened → used live with shouted test inputs. - The change request: the room picks one feature, pasted verbatim in client voice.
- The inspection: the three questions, then the proof — one scroll of the file: no
fetch, nohttp, nothing leaves the machine. - The planned failure: the change request breaks something; a client-voice bug report fixes it. "Describe the bug, not the code."
05Step-by-Step Workflow: Friction to Tool
- Catch the friction: what do you rebuild, retype, recalculate weekly? Write the honest frequency down.
- Interview: the model asks you 6 questions, one at a time.
- Spec it (template): user, problem + frequency, inputs ≤6, outputs, UI in words, NOT-list, acceptance checks.
- Build prompt: spec + "single self-contained HTML file, no frameworks, no build step, no network calls, mobile-friendly."
- Save as
.html, open in browser, run your acceptance checks. - One change request in client voice.
- Inspect — the three questions.
- Honest ledger: spec time / build time / savings per week. If saved < cost, killing the tool is also success.
06Prompt Recipes
07Lab: Tiny Tool Pitch (70 min)
Cold start friendly. Friction-starved? Standing options: export checklist for your main platform · commission price calculator · crit-notes formatter.
- Friction hunt (10 min): three recurring annoyances; pick the one with the most honest frequency. Write the number down. The boring candidate beats the impressive one.
- Interview + spec (25 min): run the interview; fill the template completely — NOT-list and acceptance checks are where specs earn trust. ✏️ the decisions only you could make.
- Build (20 min, bonus round): build prompt → save
.html→ open → run your acceptance checks. No build? Spec + build prompt is the full deliverable. - Inspect + one change request (15 min): the three questions; one client-voice revision.
Template filled by hand + hand-drawn UI sketch; the "build prompt" becomes a brief you could hand any developer. Arguably the purer version of the skill.
60 seconds: the friction (with frequency) → the tool → one acceptance check → the honest ledger. And every pitch answers: "Where does the data go?"
08Integrity Check
- Could you explain to a classmate, without hand-waving, what your tool does with their input?
- Did the tool pass your checks — or did you adjust your standards to what it built?
- Does anything in your spec smell like accounts/payments/others' data? Redraw the fence now.
- Shipping it anywhere public: "Built with AI assistance from my spec; inspected and tested by me."
09Troubleshooting
| Problem | Fix |
|---|---|
| "I don't have friction worth a tool" | You do. What did you retype, re-look-up, or recalculate this week? The boring-est candidate is the best one |
| The interview asks dull questions | "Ask harder questions — edge cases, misuse, what I'd never want it to do" |
| The tool ignores part of the spec | Quote the spec line in a change request: "The spec says X; the tool does Y. Fix to spec." |
| It looks ugly | Taste is a change request: "more whitespace, bigger type, calmer colours." You're the art director here too |
| File won't open / blank page | Check it saved as .html (not .txt — TextEdit trap), then: bug report with exact steps |
| "This feels too easy to be real" | Run your acceptance checks. If it passes, it's real. Spend the unease on a better spec |
10Resources
- LibraryLinkedp5.jsThe artist's coding canon — a friendly creative-coding library with a free in-browser editor. Where "tools" becomes "art medium."
- ChannelLinkedThe Coding TrainDaniel Shiffman — the most joyful creative-coding teacher alive. If today lit a fire, it goes here next.
- DocsLinkedMDN Web DocsWhat HTML/CSS/JS actually are, from the source — for when you want to read what you're running.
- DocsVerify before teachingClaude Code documentationThe agentic-coding tool from the demo; docs and capabilities move fast.
- DocsVerify before teachingOpenAI Codex documentationThe other coding agent; same caveat.
- PlatformsVerify before teachingApp-builder platforms (Lovable / Bolt / v0-class)Past-the-fence territory: powerful, fast, and where inspection habits matter most.
Optional extension
- Build the second tool — for a classmate, from their friction. Specs for other people are the real test.
- Style your tool with your Week 7 brand system — "apply these colours and type choices as CSS variables."
- Hard mode: open the file and change one thing by hand — a label, a colour, a default. You're now editing code on purpose.
Next week: Agentic Workflows → — when the AI doesn't just write the file, but organizes the whole folder. With checkpoints.