Automate File Hygiene with Generative Plugins

Workflow area: Design Ops · Tier: Foundation · Hands-on time: ~30 minutes

Design ops debt accumulates in five-minute increments: the pre-handoff rename, the page setup for every new project, the spacing cleanup after a messy exploration. None of it is design. All of it used to be either manual or "learn to code a plugin," and the second option cost more than the debt. Prompt-built plugins change the math — you describe the operation, the AI generates the tooling, and you validate that it does what you meant on a copy before you trust it. The skill is specifying operations precisely, which is a design skill wearing overalls.

When to use — and when not to

• Use it for ops you've done three timesthe third repetition is the automation signal; the description you'd give a junior designer is your prompt.
• Use it for structural hygienerenaming to conventions, normalizing spacing to your scale, scaffolding standard pages; deterministic operations with checkable outcomes.
• Use the meeting-notes route for recurring design ritualscritique, design review, triage: meetings whose outputs should land in a log without a human courier.
• Skip automation for judgment opsdeciding which frames to keep, what a component should be named conceptually: automating judgment produces confident wrong answers at scale.
• Skip it on files you can't afford to manglealways run a new automation on a duplicate first; bulk operations have bulk failure modes and undo has limits.

The recipe

1. 1.

   Pick one op and write its spec in plain language. Precision here is the whole game. "Clean up my layers" produces chaos; this produces a tool:

   Copy

   Operation: rename all layers in the selection to our convention.
   Rules:
   - Frames: PascalCase describing intent ("ApprovalCard"), never
     appearance ("Group 5", "Blue box")
   - Text layers: prefix "txt/" + content role ("txt/amount-label")
   - Icons: "ic/" + name ("ic/chevron-down")
   - Leave component instances' names untouched
   - Output a summary of every rename so I can scan for mistakes
   

2. 2.

   Have the AI build it. Give that spec to Figma's AI plugin-generation surface (or paste it into Claude/ChatGPT and ask for a Figma plugin you can run via the development console, following the tool's own import steps). You are the specifier and the acceptor — the implementation layer is the machine's problem. If the result misbehaves, refine the spec, not the implementation: every rule you add is a defect you stop.

3. 3.

   Validate on a sacrificial copy. Duplicate a messy real page and run the automation there first. Check the summary output line by line — good validation looks like:

   Copy

   RENAMED: "Group 12" → "ReceiptThumbnail" ✓
   RENAMED: "Rectangle 4" → "DividerLine" ✓
   FLAG: "Frame 9" → "ApprovalCard" — WRONG, it's the rejection
   card; rule needs content-awareness → add rule: read the text
   layer inside before naming the frame
   SKIPPED: instance "Button/Primary" ✓ (rule respected)
   

   Iterate spec → regenerate → revalidate until a full page passes clean. Then it graduates to real files.

4. 4.

   Scaffold the other two classic ops the same way. Spacing normalization ("snap all gaps in selection to the 4px scale, list every change") and page scaffolds ("create our standard project structure: Cover, Explorations, WIP, Ready-for-Dev, Archive, Decisions — with the cover template filled from the file name"). Each follows the same loop: spec, generate, validate on a copy, adopt.

5. 5.

   Route meeting notes into the decision log automatically. Let Granola (or your notetaker) capture design reviews, then a Zapier route sends each meeting's notes to the tool holding your pb-09 decision log (Notion, Slack channel, etc.) — with a summarization step in the middle:

   Copy

   From this design-review transcript, extract ONLY:
   1. DECISIONS made (what was decided, in one line each)
   2. ACTION ITEMS with owners
   3. OPEN QUESTIONS explicitly deferred
   Do not summarize discussion. If no decisions were made, output
   "no decisions". Never attribute a decision to a person unless the
   transcript states it explicitly.
   [TRANSCRIPT]
   

   The route delivers a draft; a human still confirms decisions before they're canonical — pb-09's correction step doesn't get automated away.

Make it yours

• Solo version, no Zapier: skip the pipeline; a weekly 15-minute ritual of pasting the week's meeting notes through the extraction prompt achieves 80% of the value with zero infrastructure.
• Team plugin library: once a prompt-built tool passes validation, publish it privately to the team workspace with its spec attached — the spec is the documentation, and teammates extend tools by editing the spec, not the output.
• Claude for the spec-writing itself: before generating anything, paste your fuzzy op description and ask "what edge cases does this spec not cover?" — five minutes of adversarial spec review saves three regeneration cycles.

Quality check

• Every automation ran on a copy before a real fileno exceptions, including "trivial" regenerations of a tool that worked yesterday.
• The rename output is semantic, not cosmeticnames describe intent; pb-34's agent-readable files and Dev Mode handoff both inherit whatever this tool writes.
• The summary output exists and gets scannedan automation without a change log is a bulk edit you can't audit; the flag-and-fix loop is the safety system.
• Decisions in the log were confirmed by a humanthe pipeline drafts; people ratify. An unratified auto-logged "decision" is a rumor with a timestamp.
• The automation didn't touch what it shouldn'tinstances, locked layers, other pages: verify the boundaries held, because scope creep in bulk tools is silent.