Strike a Pose

I recently built the same small browser-based game two different ways. The game itself is simple: capture a reference pose, then try to recreate it closely enough to produce a match score. The “gamelet” uses a webcam, detects body landmarks with MediaPipe Pose, and draws the camera feed to a canvas. Current detected landmarks are drawn in red, while a saved target/reference pose is drawn in blue. The user can snap a reference pose, optionally with a 3-second delay, then try to recreate it. The app calculates an average landmark distance between the current pose and the reference pose and shows a green or red “Match score” depending on whether the score is below.

The first version was built with AI code assistance. I used Codex as a helper: asking for snippets, debugging help, implementation ideas, and incremental improvements. That version took about four hours.

Assistance version

Agentic version

The second version was built with an agentic coding workflow, using Deft (Directive and vBrief) as the harness around Codex. That version took about 35 minutes. Deft is a pretty cool brainchild of Jonathan Taylor, one of the early (like in the first 10) OpenClaw devs.

With assistive AI coding, I was still driving at the level of individual implementation decisions. I asked for help, reviewed the output, pasted code, tested it, adjusted the prompt, and repeated the cycle. The AI acted like a capable pair programmer, but I remained responsible for breaking the work into small pieces and keeping the whole thing coherent.

With Deft, the workflow changed.

Deft is not just another prompt. It is a layered framework for AI-assisted development. Its own description is direct: “One-shot, anti-slop.” It provides standards files, deterministic task tooling, project rules, conventions, context management, and verification structure so an AI agent can follow repeatable development patterns instead of relying on whatever instructions happen to be in the current chat.

Deft helps turn AI coding into a more structured delegation process.

In the first version, the work felt incremental. The code grew through a conversational loop: add this, fix that, now make the score work, now clean up the UI, now handle saved poses. That is a useful workflow, but it is still mostly manual orchestration. The developer has to carry the architecture in their head and continuously steer the assistant back toward the goal.

In the Deft version, more of that orchestration moved into the harness. The goal, conventions, and development process were more explicit up front. The agent had a stronger frame for what to build and how to proceed. That reduced the amount of back-and-forth and made the result feel less like stitched-together assistance and more like a bounded task being executed.

There were code differences too.

The assistive version reflects the way it was made: it has more signs of iterative accumulation. Features appear as they were requested: pose snapshot, timed capture, saved pose targets, JSON loading, scoring, and visual feedback. That is not bad. In fact, it is often how small tools get built. But the structure tends to mirror the conversation that produced it.

The Deft version felt more like it came from a single implementation pass. The code path was more direct because the agent was operating from a clearer task definition. Instead of repeatedly negotiating the next small change, the harness gave the agent a stronger operating context. The result was not simply faster code generation; it was less workflow friction.