From chat tools to a production pipeline
February: chat tools, no repo access. Eight months later: 77% of the product's AI-tooling commits, and handoff straight from prototypes.
- My role
- Design Director
- Team
- 4 designers since Q3 (6 before), with an org-wide AI specialist
- Duration
- February 2026 to present
- Scope
- Pre-beta poker product, all design squads
Design's share of AI-tooling commits rose from 42% in March to 88% in September
Context
The asymmetry
Engineering had
- Coding agents as a daily environment
- Repo access and pull requests
- Its own tooling
Design had
- ChatGPT and image generators
- No repo access, no pull request ever opened
- No way to build a tool
- No bridge from the repo to Figma
February 2026 survey of both teams.
Atomic AI adoption, in four phases
One rule: give designers the leverage engineering had, then let them build tools for their own pain.
- March Access VS Code, Git and Claude Code for every designer. Weekly sessions on GitHub, branches and pull requests. First commit 20 March. Four designers committing within four weeks.
- April to June Atoms One pain point, one slash command: find-figma, figma-annotate, design-qa, and bind-tokens. 9,658 token bindings, zero errors, in a designer's first pull request.
- July to August Pipeline Commands chained into an 8-stage pipeline, intake to handoff, taught through an Education Hub. Then two Figma plugins and an intake bot. 36 tools, three languages.
- September to October Production Ship to engineering from the prototype; Figma polish in parallel. A bridge imports prototypes into Figma as real library instances. Home, PKO and Leaderboards shipped from prototypes.
The foundation under the pipeline
AI output is only as good as the system it's bound to. I set three owners, took over a stalled contractor proof of concept in July, and agreed the token pipeline with engineering.
Library and tokens
A token architecture engineering can consume.
Four-collection token set, reviewed twice with engineering, proven on a full library copy.
- About 27,000 bindings relinked
- 17 component changes in Sep, 8 clean-ups in Oct
- Russian review, 173-term poker glossary: 8 critical mistranslations found
Design tech support
Prototype to Figma without losing the components.
A bridge imports prototypes as real library instances. Seven manual steps became two approval gates.
- 114 commits to the prototyping kit since 17 Sep
- Copy-to-Figma fixed 8 Oct
- Product lead, 14 Oct: "no longer a blocker"
Programme and pipeline
Owners, baselines and the rules the agents follow.
Design-to-development charter, coverage baseline, component-token rule, and a UX rules file for AI prototypes.
- 15 of 45 code components matched to Figma; target 90%
- Six languages approved; only reviewed strings ship
- Weekly sync and a live leadership page since July
| Dimension | July 2026 | 17 October 2026 |
|---|---|---|
| Token architecture | Ad hoc; engineering couldn't add tokens for months | Proven on a full library copy; live migration is the Q4 item |
| Design-system intake | Chat threads, untracked | Figma comment to Jira ticket: 43 requests since 27 Aug, 18 shipped, 8 in QA |
| Figma to code | Demoed in April, never integrated | A pull request on every Figma token change, in a test repo; frontend pending |
| Prototype to Figma | Seven manual steps; flat frames | Library instances intact; two approval gates |
| Component coverage | 15 of 45 matched to Figma; 4 of 82 interactive components had a focus state | Target 90%; 25 component changes; focus states added |
| Localisation | Not started | Simplified Chinese QA, 13+ issues fixed; Russian review done |
Roles, never names. Three people carried this.
Key decisions and what each one cost
Skills first, the agent last
I built a UX orchestrator agent in March, without asking designers what they needed. I shelved it and inverted the order.
Traded off: our most demo-able thing, for small tools designers asked for.
Every tool earns its place
Every tool ends in adopt, adopt-with-limits or drop, backed by a written pilot evaluation, scorecards and synthetic-persona runs.
Traded off: speed. The Claude Design pilot: 316 prompts, 13 to 16 hours, one ticket; 8% of prompts decided the design. It became the exploration tool, not the handoff.
Prototype-first delivery
From 15 September, prototypes go to engineering; Figma polish runs in parallel.
Traded off: Figma is no longer the source at handoff.
Turning a critical review into owned fixes
In October the product lead listed what was still broken. Each flaw got a mechanism, an owner and a status.
- Fixed
Prototypes looked like the product but weren't made of it
bind-tokens, an adherence plugin that scores and rebinds, and the Figma bridge. Off the friction list.
- Mechanism live
Design brought in late; UI shipped without design
July: 11 of 79 stories had a design link. Now "Design required" creates a design ticket, and Done fires a Design QA ticket.
- Baseline set
No baselines to measure against
Time and iteration tracking since 21 Sep, a weekly KPI snapshot, a 25-ticket Jira study.
- Designed, not shipped
Product flaws found by research
A silent 7 to 20 minute import, hidden filters and a dead-end "Skip for now" went into the Home and import redesigns.
- Q4
Design tokens reach the frontend by hand
Token architecture proven on a library copy; Figma-to-pull-request pipeline proven in August. Live migration is the Q4 item.
- Open
Communication with engineering after handoff
Owners named. Engineering still reports no change; a buddy developer per designer is proposed.
Roles, never names. Status at 17 October 2026.
Before and after, and how each was measured
| Metric | Before | After | How measured |
|---|---|---|---|
| AI-tooling commits from design | No repo access (February) | 77%, 361 of 471 (March to October) | Git history |
| Pull requests opened by designers | None ever | 65% of the repo's pull requests, 74 of 113 | GitHub pull-request list |
| Designers delivering from AI prototypes | 1 (Q2) | 4 of 4 (October) | Per quarter |
| Design tickets per designer, per quarter | 14 (Q4 2025) | 33 (Q2 2026) | Jira; tickets held, not closed or sized |
Where a ticket's time went
About 80% of a ticket's life was waiting. Features without design took the same 52 days and failed QA more.
Jira study of 25 tickets
What research found
First moderated study: 0 of 5 participants reached their analysis unassisted. Synthetic personas: 0 of 4 understood the progress block. Both fed the Home and import redesigns.
5 moderated sessions, August · persona runs, September
To a peer
Tooling does nothing on its own; adoption is the work
Adoption was uneven. One designer wrote 28 of 37 commands; one opened 43 pull requests, 37 merged. Weekly sessions made the first commit possible, a few hands made the volume, and now every designer ships from prototypes. The lesson: ask designers what hurts before building for them.
Not proven yet
- Hours saved: not measured. One scorecard filled; target is 80% of runs.
- Tickets are counts, not effort: sizing at pickup is 50 to 67%.
- Design tickets at Ready for Dev: 10 in Q3, 8 in early October. Promising, small.
- Credit limits and prototype hallucinations still interrupt work.
Q4 asks: token auto-sync with an engineering owner, a buddy developer per designer, the AI design-engineering hire.