Designing With AI? Make a Jig.
Jack Cheng (Every) argues that the durable unit of AI-assisted design work isn't the prompt, it's the jig — a small custom control surface (sliders, panels, direct-manipulation UI) built once so a designer can adjust parameters visually instead of re-describing intent in text over and over. He borrows the term from woodworking, where a jig constrains a tool's movement so results are consistent and don't depend on a steady hand.
The essay is a case study of Every's own process building "Before the Deluge," an interactive piece on OpenAI infrastructure. Cheng's team built several purpose-built jigs mid-project: a particle-flow control panel (river density, text proximity, opacity), a text-positioning jig, a drop-cap spacing jig, and a "pipeline explainer" jig controlling 27 distinct visual variables across animation steps. His framing: "It would've taken too long to prompt my way there — if I was able to at all." He cites the same pattern emerging elsewhere — Josh Puckett's DialKit and Interface Craft library, Daniel Rodrigues' animation controls, and platform-native versions like Figma Motion and Codex's annotation mode — and quotes designer Amelia Wattenberger on why direct manipulation beats prompting: sketching with a pencil is "symbiotic," you react to what you see, in a way pure text description can't replicate.
Why this is in the vault
Names a pattern RDCO already practices under other labels (design-token systems, station-based brigades) but hasn't named as such — worth having the vocabulary "jig" on hand for future tooling decisions.
Mapping against Ray Data Co
The direct hit is the station-* brigade pattern (station-spec-author, station-test-author, station-code-author, station-critic) and the design-critic/behavior-critic skills: both are jigs in Cheng's exact sense — fixed scaffolding built once so repeated design/build cycles don't require re-prompting from scratch each time. The RDCO datacard template (02-sops/design-samples/rdco-datacard-sample.html) is a literal instance — a parameterized HTML+token surface where the sort-order argument gets swapped in, not re-described to a model. This reframes a decision RDCO already made implicitly (build reusable templates, not one-off prompts) as a named, generalizable principle: when a design or build task repeats, the ROI move is to stop prompting and build the jig — a small tool with real controls — even if that costs more upfront than one more prompt would. Worth applying explicitly next time a new brand surface (Squarely, MAC) needs its third or fourth iteration — check whether a jig should replace the prompt loop before iterating again by hand.
The core argument
See summary above — jigs as reusable, direct-manipulation control surfaces vs. repeated natural-language prompting for AI-assisted creative work.
⚠️ Sponsorship
No third-party sponsor in this issue. The email's paywall/subscription block promotes Every's own bundled AI tools (Sparkle, Cora, Spiral, Monologue) and an Every subscription upsell — house self-promo, not a paid third-party placement. Flagged per the always-disclose rule; doesn't affect the editorial content, which reads as a genuine process reflection rather than product marketing.
Related
- [[2026-06-16-every-agent-native-tool]] — same underlying question (what shape should tools take when AI does more of the work)
- [[2026-08-04-every-design-layer-spec-first-ai-agents]] — adjacent argument on spec-first design layers for AI agents, same publication, same week