Web & Digital

Procol's design team rebuilt a procurement product in code with Claude, then built the design system last

Procol's head of design says six designers rebuilt the product in code from one Figma frame, then an audit-and-fix loop turned the mess into a system.

Done withClaudeFigmaFigma MCP

Head of design's own account of his company's product (two-post series)


WHAT THEY DID

Yugandhar Bhamare, who heads design at Procol, says the team rebuilt the look and feel of its procurement product by giving Claude one complex frame through the Figma MCP and rebuilding the product in code. Six designers pushed their own modules straight to main, with an iteration switcher and a persona switcher on every page to compare versions on real data.

STEP BY STEP

  1. Gave Claude one complex product frame through the Figma MCP and rebuilt the product in code with a new look and feel.
  2. Six designers pushed their own modules straight to main; every page got an iteration switcher and a persona switcher.
  3. A harness audits every file (token use, type roles, components, duplicates and 20 rules) on one live coverage page.
  4. The head of design decides the rules and logs each call in a governance ledger; decisions are codified into a living UI library inside the app.
  5. Auditor, fixer and verifier agents in Claude fix drift using only settled rules; a guard fails new violations and refuses production deploys unless every check passes.
  6. Second phase: Figma components mirror code names and props; Figma is used for documentation and experiments, and Claude syncs changes both ways.

HOW IT WENT

He reports that three weeks in, baselined violations went from 221 to 0, shadows on the system from 80.6% to 100%, and canonical components from 29 to 69. A week later he reported 105 canonical components and 93 Figma masters bound to variables named as in code.

Reported results: Three weeks in: baselined violations 221 to 0, shadows on the system 80.6% to 100%, colour and radius at 100%, one style per text role, canonical components 29 to 69. By 13 Nov: colour, type, radius and shadow at 100% with 0 new violations, 105 canonical components, 93 Figma masters bound to variables named as in code.

WORTH KNOWING

An unusual order of work, design system last, with a named loop (Explore, Select, Audit, Decide, Codify, Fix, Guard) and a clear split: agents fix only settled rules, and anything needing judgement goes back to the head of design.

TRY IT

  • When exploring a redesign, add an iteration switcher to each page so the team compares versions live instead of in static frames.
  • Log each design-system decision (which duplicate wins, one style per text role) in a ledger before letting agents fix drift.
  • When binding colours to variables, check alpha too; the team's first pass turned 300 hairline borders solid dark grey.
Read the original ↗

Rates the entry, not the tool.

Source: Yugandhar Bhamare (@yug_design) · XOpen as a page

  • Work done: not stated in the source (Not stated as dates. The 6 Nov post says 'Three weeks in'; the 13 Nov post compares with 'a week ago'.)
  • Source published: Nov 6, 2026
  • On 4TC: added Nov 17, 2026 · checked Nov 17, 2026
About this source

Type: Documented method · Firsthand report

  1. We rebuilt the look and feel of our entire procurement product, and we built the design system last.
    Yugandhar Bhamare (@yug_design) · X · Published Nov 6, 2026 (Posted 2026-11-07T02:38:06Z, which is 6:38 PM PT on 6 Nov 2026.) · Checked Nov 17, 2026
    Primary: the head of design's own post about his company's product (bio: 'Heading design @procol_io'). X paid_partnership flag false; no vendor sponsorship found.
    Access: X post; read in full via the X API.
  2. Last week I wrote about building our design system last ... This is the second half
    Yugandhar Bhamare (@yug_design) · X · Published Nov 13, 2026 (Posted 2026-11-13T18:24:32Z (10:24 AM PT).) · Checked Nov 17, 2026
    Primary: second post in the same series by the same author.
    Access: X post; read in full via the X API.

Source check: Researched from the public page; not yet re-checked by an editor. Nobody at 4TC has tested this.

Claims and where they come from

  • The team rebuilt the look and feel of its procurement product and built the design system last. (reported) [1]
  • They gave Claude one complex frame from the product through the Figma MCP and rebuilt the product in code with a new look and feel. (reported) [1]
  • Six designers pushed their own modules straight to main; every page got an iteration switcher and a persona switcher. (reported) [1]
  • The harness runs as a loop: Explore, Select, Audit, Decide, Codify, Fix, Guard; it audits token use, type roles, components, duplicates and 20 rules. (reported) [1]
  • Auditor, fixer and verifier agents in Claude fix drift using only settled rules; production deploys are refused unless every check passes. (reported) [1]
  • Three weeks in, baselined violations went from 221 to 0, shadows from 80.6% to 100%, and canonical components from 29 to 69. (reported) [1]
  • The first pass of binding colours to variables turned 300 hairline borders solid dark grey, and the next audit caught it. (reported) [2]
  • As of 13 Nov he reported 105 canonical components, up from 69 a week earlier, and 93 Figma masters bound to variables named as in code. (reported) [2]

4TC interpretation

Our reading, not something the source says.

He says designing a Figma library first locks in components before you know what a reimagined product needs. The cost of exploring freely was the same card built five ways and raw hex colours, which the audit loop then cleaned up. In the second post, Figma becomes documentation and a place for experiments, while code stays the tiebreaker.

Designer impact

Alleged [1]

Free exploration with an agent produced the same card built five ways, raw hex colours, one-off shadows and hand-written buttons and inputs, which he says made QA by eye hectic.

  • Pair exploration with an automated audit before converging.
  • Let agents fix only rules a person has already decided; route judgement calls back to a designer.

Established = documented by primary sources or the company’s own statements. Disputed = contested with evidence on both sides. Alleged = reported, not established.

Votes rate whether this entry was useful. They are not an endorsement of the tool, and don’t mean voters agree with the company’s practices.

Report an issue or correction

Share your own experience

4TC does not endorse a tool unless an explicit 4TC endorsement is shown.

Get the best of these weekly

The weekly design brief, starting soon: the most useful workflows, tool updates and developments, with sources. Sign up now and the first issue will come to you.