Skip to content

Writing

Learning to design with agents through real client work.

Each Monday note looks back at StickerGiant, Loopy, and the local agent setups I use to turn ideas into prototypes, tests, and better decisions.

The method came back to the client

Loopy stopped feeling like a side project when it changed how I approached a live StickerGiant problem.

  • StickerGiant
  • PDP visualizer

I started testing the whole loop

A clean demo tells me almost nothing about whether a client can trust the product.

  • Loopy
  • Full-loop QA

A design system needs a story and a source trail

Tokens are not enough if nobody can see what they mean or tell whether they came from a real source.

  • Loopy
  • Design-system evidence

Security became a design problem

A useful agent can also become a very efficient way to mishandle client work.

  • Loopy
  • Trust boundaries

I stopped making users translate their design system

The product should do the archaeology before it asks a designer to fill in the gaps.

  • Loopy
  • Source-first onboarding

The loop moved into the conversation

I wanted Loopy to begin where product decisions already happen, not in another empty dashboard.

  • Loopy
  • Buzz + Hermes prototype

Pencil became part of the build loop

The design file stopped being a picture of the idea and became the source for the next implementation decision.

  • Loopy
  • Pencil with agents

The agents' work became the interface

A flow diagram was explaining Loopy while hiding the only thing worth watching.

  • Loopy
  • Live iteration experience

The demo had to survive Monday morning

Loopy was becoming a product, which meant the boring failures were now design material.

  • Loopy
  • Hosted reliability

I called it Loopy

The old name described infrastructure. The new one described how I wanted teams to work.

  • Loopy
  • Product identity

The scorecard finally had somewhere to go

Evaluation became useful once it changed the next decision. A score alone just decorated the last build.

  • Loopy
  • Eval-led routing

Loopy started as a hand-built harness

I built the product only after the manual loop started helping with real client work.

  • Loopy
  • Origin

I started drawing with agents

An editable design file gave the conversation a place to land.

  • Practice
  • Pencil experiments

I refused to pick concepts from prose

Three descriptions are not three design directions if I cannot actually see or use them.

  • StickerGiant
  • Live concept review

I put the brief before the prompt

The agents needed the client's world before they needed another instruction.

  • StickerGiant
  • Human feedback loop

I gave the agents different jobs

One chat should not brainstorm, build, critique, test, and approve its own work.

  • StickerGiant
  • Agent roles

The design system became an argument with receipts

Extracting Figma values was easy. Preserving where they came from mattered more.

  • StickerGiant
  • Design rules

I gave the agent a room of its own

The environment around the model started mattering as much as the prompt.

  • Practice
  • OpenClaw local environment

Twenty-one folders taught me one thing

The StickerGiant directory was full of experiments, but the folder count was not the value.

  • StickerGiant
  • Practice review

A scaffold is not a pipeline

AI can build a convincing shell long before the hard handoff works.

  • StickerGiant
  • Prototype honesty

I started turning decisions into systems

The reusable part of a design decision is the rule and the evidence behind it.

  • StickerGiant
  • Badges, tokens, deploys

Product education belongs in the interaction

Customers should not need a support article to understand the choice in front of them.

  • StickerGiant
  • Product discovery

The review stopped being a vibe

I wanted critique that engineering could use, including the moments when Figma was wrong.

  • StickerGiant
  • Design review

I connected AI art to print production

A generated image was not a sticker until the manufacturing rules came with it.

  • StickerGiant
  • Print-ready generator

The agent learned to say what it could not see

A confident design review is dangerous when the source API did not expose the thing being reviewed.

  • StickerGiant
  • Figma audit

I designed the states the ticket left out

The happy path was one screen. The real customer experience was several different products hiding inside it.

  • StickerGiant
  • Customer states

Brand rules became something agents could query

A PDF tells people what the brand was. A shared source can help people and agents keep using it.

  • StickerGiant
  • Brand playbook

The first useful outputs were review surfaces

AI was most helpful when it made existing client evidence easier to see and discuss.

  • StickerGiant
  • Early prototypes