I'm (finally) showing how I design with AI

2026-06-15 - 79 min - source - Read full transcript

Kyle (host)Rid

Key insights

Git worktrees, not branches, are what let multiple agents work the same codebase without colliding.

Conductor gives every 'mental lane' of work its own full clone of the repo. Agents on different worktrees never see or interfere with each other, whereas agents working different branches of the same checkout will bump into each other's changes. Rid treats each worktree as a parallel track (design exploration, playground build, research) rather than mapping worktrees to git branches.

parallel-agent-orchestration

Model choice is split by task type: Opus for creative/planning work, GPT 5.5 for bulk execution.

Rid says he almost always uses Opus as a thought partner and for plan-mode work, especially anything inside Paper, because he finds it stronger creatively. GPT 5.5 is his daily driver for execution because he finds it faster and better at nitty-gritty technical details when reviewing plans.

ai-assisted-design-workflow

Bouncing a plan between two different models with a numeric confidence score catches gaps neither model would surface alone.

Claude writes the plan as a markdown file; Rid opens a separate Conductor tab, switches it to Codex/GPT 5.5, and asks it to find edge cases, simplifications, and give a 1-100 confidence score that the plan will actually work. Plans typically start around 60 and Rid iterates the loop (Claude revises, Codex re-scores) until it reaches the high 80s or 90s before writing any code.

prompting-technique

Explicitly telling an agent your own limitations and naming a real human reviewer changes its behavior, and the framing sticks in memory.

Rid's global instructions tell every agent he isn't an engineer and name the specific colleague (Ver) who reviews his PRs, framing the agent's job as not embarrassing him in front of that person. He says this visibly nudges the agent toward the more conservative of two implementation routes, even in plan mode, and that Claude's session memory has picked up on the framing over time.

prompting-technique

Design tokens belong in a file referenced on demand, not inlined into the root instructions file that loads on every session.

Rid keeps a design-system.md separate from his CLAUDE.md/AGENTS.md, with an instruction to consult it only when the task touches UI, and to propose updates back to it when a change affects multiple places. He warns that inlining a heavy design file into the auto-loaded root file burns context on sessions that have nothing to do with UI.

ai-assisted-design-workflow

Paper is used for volume and speed, not pixel accuracy, because detail refinement now happens directly in code.

Rid explicitly doesn't care whether Paper output matches his real design tokens or components pixel-for-pixel; he treats it as medium fidelity purely for generating and comparing many directions fast. Once he picks a direction, refinement happens through tools like Agentation acting directly on the running code, which he now trusts more than a design file that constantly drifts from production.

ai-assisted-design-workflow

A self-built /playground/<feature> route with an internal version history turns disposable prototypes into a durable, reusable idea bank.

Rid scaffolds a versioned playground in code that runs in parallel with his Paper explorations, has each new duplicate spin up a fresh component so past versions are preserved, and defaults to the most recent version on refresh. He says he sometimes returns to a playground built a month or more earlier and reuses ('self-plagiarizes') an interaction he'd forgotten about for an unrelated feature.

prototype-feedback-loops

Sharing a live, working prototype closes the feedback loop faster than a single deployment-preview link because non-technical reviewers can pin comments to specific DOM elements.

Rid argues that jumping straight into a single in-progress preview link creates pressure to show only one direction and skips a useful 'go wide' checkpoint. Inflight instead lets him publish several versions from a playground, auto-generates a poll/context video, and captures spatially pinned feedback that flows straight back into his next agent prompt as the following task.

prototype-feedback-loops

Voice dictation produces richer prompts than typing because people narrate more context out loud than they'd bother typing.

Rid dictates almost every prompt via Whisper Flow (naming Monologue and Aqua as alternatives). He and Kyle both note that people talk differently than they type, giving agents more surrounding context for free, and that agents perform better the more context they're given.

prompting-technique

The specific tool stack will be irrelevant within months; the durable skill is learning to communicate intent to an agent.

Kyle notes that almost none of what Rid demoed in a January episode together still applies. Rid frames the current period as roughly a year or two of constant tool and model churn, and argues neither he nor his peers have any tool loyalty - they compare notes constantly and switch to whatever is best that week. The transferable skill is the underlying workflow philosophy (brainstorm, plan, work, review, compound), not any single product.

tool-agnosticism

Native iOS development is currently underserved by AI-assisted design-to-code workflows compared to web.

Asked whether his flow works for building iOS apps, Rid says it doesn't translate well: as soon as work moves into Xcode specifically, tooling like Agentation stops applying and the surrounding ecosystem is generally behind where web tooling has gotten, making native mobile a harder environment for this kind of agent-driven visual workflow right now.

tool-agnosticism

A Chrome extension that snapshots live DOM elements into the design tool removes the need for an MCP integration or from-scratch redesign.

Paper's Chrome extension lets Rid select any element on a deployed page and pull its exact markup into Paper because Paper's canvas is literally HTML/CSS/DOM. He uses this to grab something already in production, iterate on variations of it in Paper, and then paste the result back into code, treating it as his new default starting point instead of designing from a blank canvas.

prototype-feedback-loops

Media referenced

Dive Club episode with Rid, January 2026 - podcast - Kyle references an earlier Dive Club episode with Rid to note how completely the tool stack has already changed since then.

Companies

Anthropic - Maker of Claude and Opus, the models Rid uses as his default thought-partner and plan-writer throughout the session.
OpenAI - Maker of Codex and GPT 5.5, which Rid is testing as his daily coding driver and uses as the adversarial reviewer of Claude's plans.
Conductor - Rid's primary harness; spins up a git worktree per task with its own agent tabs, run scripts, and PR buttons. Its maker Charlie drops into the live chat.
Cursor - Discussed as Conductor's closest competitor; Rid notes its newer version has converged on worktrees and agent tabs similar to Conductor.
Paper - Rid's primary AI design tool - a Figma-like canvas that is literally the DOM, used for medium-fidelity visual exploration before code.
Quiver AI - SVG-generation model now built into Paper; Rid uses it to generate icon and graphic concepts he can't draw by hand.
Vercel - Deployment target Rid uses to push shareable Playground prototype links.
Inflight - Rid's own product, demoed throughout the episode: it embeds in preview links, lets reviewers pin spatial questions to DOM elements, and turns collected feedback into the next agent prompt.
Figma - Rid's former design tool; he says his Figma files fell permanently out of sync with production once he moved detail work into code, and he no longer maintains them.

Techniques and frameworks

Parallel git worktrees per task - Conductor gives each 'mental lane' of work its own full repo clone, so agents working on unrelated things never collide the way they would on shared branches.
Confidence-score plan review loop - Claude (usually Opus) drafts a plan as markdown, Codex/GPT 5.5 critiques it and returns a 1-100 confidence score, and Rid bounces the plan back and forth until the score reaches the high 80s or 90s before letting any code get written.
Persona and stakes framing in prompts - Rid tells his agents up front that he isn't an engineer and names the specific human reviewer he doesn't want to embarrass; the framing persists in the agent's session memory and visibly biases it toward safer, more conservative choices.
Design-system.md kept out of the auto-loaded root file - Rid keeps design tokens in a separate design-system.md referenced from CLAUDE.md/AGENTS.md rather than inlined, so non-UI sessions don't burn context loading rules that don't apply to them.
Playground routes with in-code version history - A self-built pattern (/playground/<feature>) that scaffolds a versioned set of live components in code, run in parallel with Paper explorations, that Rid sometimes returns to and reuses months later.
Dial Kit parameter sliders - A free tool (by Josh Pucket) that auto-generates tweakable sliders for any CSS, animation, or shader value on the page, replacing repeated back-and-forth prompting to dial in a number.
Paper's Chrome extension / DOM snapshot - Lets Rid select any live element on a deployed site and pull its exact markup straight into Paper as an editable starting point, skipping design-from-scratch entirely.
Compound engineering / Superpowers workflow plugins - Claude Code plugins that encode a brainstorm-plan-work-review-compound loop as reusable skills; Rid uses compound engineering, a colleague uses Superpowers, and both consider the underlying philosophy interchangeable.
Post-PR simplification prompt - A near-every-PR prompt asking the agent to look for consolidation opportunities 'so my engineer doesn't hate me' without risking visual regressions, and to separate real fixes from 'review theater.'
Storyboard-style animation breakdown (Interface Craft) - A paid resource by Josh Pucket that describes animations as individual labeled frames/moments, which Rid says helps agents understand and reproduce motion far more accurately than a generic prompt.
Skill Distillery - Rid's own open-source framework for building layered, multi-file Claude Code skills rather than single-file ones, released free on his GitHub alongside a design-motion-principles skill.

Summary

Rid, a designer building the feedback-collection product Inflight, spends most of this Dive Club episode doing a live, unscripted demo of his actual AI-assisted design workflow, while co-host Kyle interjects context and audience questions. Rid works inside Conductor, a harness that gives each parallel line of work its own git worktree (a full repo clone), so he can run several agents on unrelated tasks simultaneously without them interfering with each other's code. Within a session he alternates between Paper, a Figma-like design tool whose canvas is literally the DOM, and a running local build, dictating prompts by voice through Whisper Flow rather than typing them.

The episode is structured as three concurrent real tasks: reworking a grid-card component so more information is visible without hovering, fixing an "optical spacing" nit on the marketing site using a live parameter-tweaking tool called Dial Kit, and prototyping a new spatial-pinning interaction for Inflight's feedback questions using Claude in plan mode. Rid narrates his reasoning throughout: he deliberately treats Paper output as medium-fidelity, generating ten-plus visual concepts fast rather than caring about pixel-perfect adherence to his real design tokens, because refinement now happens directly in code with a tool called Agentation. He builds a versioned "/playground" route in parallel with his design exploration so he has a shareable, interactive artifact to send people, rather than a single deployment-preview link that would force a premature choice between directions.

Several process patterns recur. Rid keeps a design-system.md file separate from his auto-loaded CLAUDE.md so UI-specific instructions don't burn context on unrelated sessions. Before letting an agent write real code on anything nontrivial, he has Claude draft a plan, then opens a separate Codex/GPT 5.5 tab to critique it and assign a 1-100 confidence score, iterating the back-and-forth until the score climbs from a typical starting point around 60 into the high 80s or 90s. He also tells every agent, in his global instructions, that he isn't an engineer and names the actual colleague who reviews his pull requests, which he says visibly nudges the agent toward safer, better-reasoned choices and appears to have stuck in the model's session memory.

Guests from the tools themselves join briefly: Charlie, who builds Conductor, and (per Kyle's closing remarks) someone from Paper, both watching Rid use their products live. Rid uses the moment to plug free resources - a "skill distillery" framework he built for layered Claude Code skills, a design-motion-principles skill, and Josh Pucket's paid Dial Kit and Interface Craft resources, the latter using a storyboard method to help agents reproduce animations accurately. Toward the end, Rid closes the loop on the Inflight demo itself: he publishes several playground versions, Inflight auto-generates a feedback poll and context video, collects pinned comments from reviewers, and Rid pastes that feedback straight back into Conductor as the next prompt, describing the tool as functionally his to-do list.

The recurring meta-point, made by both hosts, is that the specific tools shown will likely be irrelevant within months - Kyle notes that almost nothing from a January episode with Rid still applies - so the transferable skill is learning to give agents rich context and clear intent, not mastering any particular product. Rid flags one clear gap: this workflow does not translate well to native iOS development, where tooling still lags behind the web.

Notable Quotes

"None of us have tool loyalty right now. I mean, you guys know that... whoever has the best thing right now, we're all talking to each other." - Rid

"The how do I work with AI effectively, that has a little bit more staying power... but the tool stack does not make the product or the design. We're all just, you know, like if you get a Craftsman power drill versus a Makita, they're all drills. Learn how to use a drill and how wood works." - Rid

"Tell AI that you're not an engineer too... I don't want to make big decisions. Please help me accomplish my intent without making big decisions basically." - Rid

"I ask Claude which items are review theater a lot versus which ones are actually worth doing, and it always eliminates some from the list." - Rid

"It's not about getting it pixel perfect... I've really started doing more of the refinement work in code." - Rid