What do design, implementation, and iteration look like in an agentic-driven workflow?
This page documents 36 iterations of bringing Claudito — a pixelated, four-legged Claude avatar — to life through conversation.
This exercise demonstrates some key observations for a collaborative workflow with an AI agent.
Methods can be over-prescriptive. In some cases, defining the method to use seemed to be too restrictive and didn't result in the outcome or visual feeling intended. (See Version 3)
Describe feelings, not implementations. It often proved more productive to describe the intended feeling with different ranges of vocabulary that could communicate the idea. (See Version 8)
Screenshots enable shared vision. A very powerful way to provide feedback to the agent was through screenshots, enabling the agent to see as I see on the other side of the screen. (See Version 17)
Precision with latitude. Instructions that were simultaneously quite specific but also somewhat vague were able to be interpreted and executed on accurately and to the precise specification intended. (See Version 22)
Feel is iterative too. Later iterations showed how subjective feedback about feel can drive refinement. The completion celebration went through eight versions, with prompts like "make the glow pulse 3 times" and "the maximum glow is 80% as bright" fine-tuning the emotional payoff until it felt right. (See Versions 28–35)
The journey from v1 to v36 followed a natural progression:
From the agent's perspective, this workflow highlighted the value of iterative refinement over upfront perfection. Each version built on the last, with human feedback steering direction while preserving creative latitude. The shift from rectangular legs to organic bezier tentacles (v7→v8) emerged not from explicit technical instruction, but from understanding the desired feeling—"quadropus"—and translating that into appropriate techniques.
The data-driven approach in v20 (using a keyMap hash) exemplifies how a well-timed structural suggestion from the human—"create a hash of all the keys"—can unlock cleaner solutions. Sometimes the best contributions aren't about doing more, but about reframing the problem.
The final stretch (v28-v36) revealed something else: feel is iterative too. The golden glow celebration went through cascading sequences, pulse timing adjustments, and opacity tweaks—each change small, but together transforming a functional completion state into a satisfying moment. The last fix (v36) addressed tentacle attachment points, a detail that only became apparent after everything else was working.