What this is
Two small tools, built to answer a question I kept running into: what does it take for someone who does not write code to use an AI tool on something that matters, and keep using it after the person who built it walks away.
The setting was Figma. Someone needs to update copy across an existing set of layouts. The AI can do it. The risk is not that the copy comes out badly. The risk is that a text edit quietly changes a frame's auto-layout, and nobody notices until a deck goes out looking wrong.
For these two I also did the handoff.
The problem
A working prototype is not a deliverable. It is a demonstration that the thing is possible.
The failure mode is predictable. The tool works when its author drives it. It is handed to an operator. The first time it produces something unexpected, the operator cannot tell whether they did something wrong, whether the file was already broken, or whether the tool is having a bad day, and there is no safe way to find out. So they stop using it, and they go back to doing it by hand, and the project failed.
The design work is deciding what the operator must not be able to do, and then removing it from the tool rather than from the instructions.
What the constraints actually are
The content guardian is text-only, and that is enforced rather than requested. It may propose and apply changes to existing text content, and it must preserve layout, typography, color, spacing, auto-layout, components, variants, tokens, and geometry. It is not a tool for generating new design. It is a tool for protecting existing design while the words change.
Changes travel a fixed path with no way around itC1: export a normalized snapshot from the plugin, generate a proposal, check it against guardrails, approve it with a named approver and a note, run a dry run, then apply.
The read path and the write path are deliberately different systems. Reads go through the API. Writes happen only through the plugin. There is no code path where an automated process can write to a design file on its own.
The brand manifest is a single source of truth, and a test fails when generated files drift from it. Configuration control here is a check that goes red.
The second tool is not code at all. It is a written harnessC3 that lets a non-technical operator drive an AI coding agent against a design file: standing instructions, safety rules, change modes, prompts, a checklist, and setup notes. Its main rule is inspect first, plan second, write only after confirmation. For anything touching layout, the agent creates a variation beside the original rather than overwriting it, so the recovery step is doing nothing rather than undoing something.
The decisions I would defend
Take the dangerous options away rather than documenting them. A safety rule in a document is a rule an operator can miss. A capability the tool does not have is a rule that cannot be missed.
Make the safe move the lazy move. Creating a variation beside the original is less work than overwriting and recovering. Operators do the easy thing, so the easy thing has to be the correct one.
The handoff is the product. The checklist, the setup guide, the quick start, and the standing rules are the part that determines whether the tool survives contact with someone who did not build it.
Evidence
- 27 test files coveringC2 the approval pipeline, the guardrail builder, the patch generator and validator, and manifest drift
- 14 documentation files covering scope rules, the handoff checklist, setup, and the operator workflow
- The packaged harness as delivered, with its safety rules, change modes, and quick start
Status
Working and handed off. The guardian pipeline runs end to end, and the harness was packaged and delivered as a kit.
Known gaps. These are small tools at personal and single-team scale, not products. The handoff went to one operator, and one operator is an anecdote rather than evidence that the design generalizes. Some support was still needed after delivery, which is why the checklist exists. Nothing here has been tested with an operator I did not have direct access to.