Craft · Core Competency

AI in the Process, and AI in the Product

Two different problems share this page. In the process, Claude Code, Codex and Cursor are in the daily loop, and most of their value is rejecting output fast: generation is excellent at the tedious and known, and confidently wrong wherever the domain decides. In the product, a system that is only ever probably right needs an interface that says what it saw, how sure it is, and what a person does when it is wrong — I designed that for face recognition at metro gates and ATMs, and it is the same problem an agent has when it proposes an action on somebody else’s machines. What I have not done is ship that agent. The closest thing is a design, and it is on this page.

In the Process
Claude Code, Codex, CursorCode-first components, Storybook, Figma downstream
The Team Rule
AI writes what we can verify quicklyIt does not decide business rules or untested edge cases
In the Product
Confidence, threshold, retry, fallbackBiometric authentication at gates, ATMs, self-checkout
Not Yet Shipped
The human side of an agentDesigned once, for an IT-operations platform, 2026
3 Process
AI Coding Assistants in the Daily Loop
0 Rule
Untested AI Output Allowed into Production
~1 s Product
To Turn a Match Score into a Decision at a Gate
8 Design
Stages in the Agent Action Lifecycle I Modelled, Signal to Audit
01 · Evidence

Where this showed up in real products

IBWTAI in the Process
Where AI Writes the Code, and Where It Is Not Allowed To
Claude Code, Codex and Cursor scaffold components, state permutations and responsive variants in code first; the result is published to Storybook with its states and documented in Figma afterwards. The boundary is the part that matters: generation is confined to what a reviewer can check in minutes, and kept away from business rules and from edge cases nobody has mapped — it will produce a plausible order-entry screen that would lose someone money.
Read the case →
Generative PipelinesVisual Asset Ops
Model Conditioning & Seed Control for Unified Asset Sets
Style references, seed control and model conditioning so a set of production illustrations and visual assets reads as one family across formats and can be extended later in the same key. No set is shown on this site.
02 · Method

How I actually do it

STEP 01
AI Scaffolding in Code
Use AI assistants to generate boilerplate component scaffolding, TypeScript interfaces, responsive CSS variants, and accessibility attributes in seconds — the tedious and known parts, which is where generation is reliable.
STEP 02
Interactive Verification in Storybook
Never trust a static preview. Render generated code in Storybook and poke at the real edge cases, keyboard focus, dark-mode tokens and extreme viewport widths. Most of the value of the tools is how fast a wrong variant can be thrown away.
STEP 03
Domain Rules Guardrails
AI does not guess financial calculations, compliance constraints or business rules. High-risk logic is mapped with stakeholders beforehand, and a generated screen that looks plausible and is wrong is treated as worse than no screen.
STEP 04
Where the Agent May Act, Where a Person Must Check
The same rule inside the product. A model’s output is a proposal with a confidence attached, and the interface decides where the machine acts alone and where a person must check. The threshold half I shipped, in biometrics. The agent half — friction that grows with the blast radius, one machine one click, many machines across clients a stop that names them; rollback declared before the run, not discovered after — I have designed once, for an IT-operations platform, and not yet shipped. Both follow the fintech rule: you do not confirm a payment without seeing the fee and the total.
03 · Stack

Tools, in service of the above

AI Coding Assistants
Claude CodeCodexCursorClaude Design
Patterns — One Shipped, One Designed
Confidence thresholds & retries (shipped)Proposal → scope → approval gate (designed)Dry run & rollback (designed)Audit trail (designed)
Generative Production
Style referencesSeed controlModel conditioning
Frontend Runtimes
ReactVue 3Storybook
Design Tokens & APIs
Figma VariablesTokens StudioFigma Plugin API
The full version

What the complete version of this page adds

  • The code-first pipeline end to end: prompt, component, Storybook entry, review
  • Where generation was allowed to decide and where it was explicitly not — including the Figma bridge, whose writes are typed commands and whose batches are one undo step
  • The threshold asymmetry from the biometric platform: what a false reject and a false accept cost at a gate and at a cash machine, by client
  • The 2026 take-home for an IT-operations platform as completed: domain model, both screens, the failure states

It lives at alex.zhovnir.com/craft/ai-workflow, behind a password — client agreements, not theatre. Send me a note and I’ll open it: alex@zhovnir.com.

Related: IBWT