I'd rather let the sessions speak.
AI is a real revolution, and revolutions scramble language. Titles get recombined, and overnight self-proclaimed experts are everywhere. I'm introverted enough as it is, so rather than add to that chorus, I'd rather let my work speak for me.
What follows is how I actually work with the model, as measured by apply new, a tool built by Play New, from forty-five days of my logged Claude Code sessions. I didn't grade myself; the logs did. The unedited output is linked at the bottom.
Source: Claude Code, full capture. Log-consistency screen 100/100 (a screen, not a proof). Window and counts are lower bounds, since logs rotate and old sessions are pruned.
A freelance designer and semiotician who also ships production code: consumer and creative apps across web and native iOS, prototypes ported into design systems, presentation decks crafted by hand. A daily driver, mostly co-thinking. Detailed briefs to launch greenfield work, then tight iterative refinement. Returns to a handful of products repeatedly: depth over breadth.
I switch to directing mode to launch work: detailed implementation plans, thorough codebase audits, and full product briefs set the frame. Once building, I co-think, steering through short, colloquial reactions, corrections and continue-prompts rather than fresh specifications. Roughly three quarters of the exchanges.
I decompose greenfield work up front through detailed plans, then track execution with structured task lists and frequent clarifying exchanges before committing to a direction. Verification is the defining habit: I inspect model output against the source, catching silently changed strings, omitted fields and prototype-to-screen divergence rather than accepting work at face value. Risk is calibrated by context: I let the model push freely on personal projects, and stay tight and low-churn on client deliverables (zero reverts).
A mature stack, used as standard tooling rather than novelty:
- Builds: a custom skill authored, hooks edited, five project-memory files maintained.
- Tooling: a self-built multi-tool MCP server driven through 303 calls, alongside a public one (125 calls).
- Orchestration: sub-agent delegation, custom commands, and 186 structured task-tracking events.
- Deliberation: 57 clarifying questions raised before committing to a direction.
- Web & native app development — consumer and creative apps across web and iOS, plus a design-to-code sync toolchain. 19 sessions, 111 commits.
- Design-systems & design-to-code — porting interactive prototypes into a maintained design system; localization of structured content. 11 sessions.
- Presentation & talk design — an HTML conference deck on AI-assisted qualitative analysis, with build tooling. 9 sessions.
- Developer tooling — auditing and repairing a local agentic setup. 2 sessions.
Over the window the practice shifted from exploration toward confident execution. Prompts grew more specified and delegation roughly doubled, while the ratio of research to actual changes fell: less circling, more building. Verification rose in step, so faster execution did not cost scrutiny.
research : mutation 0.75 → 0.60 ↓
verification rate 22% → 26% ↑
These are screens, not proofs. The honest move is to read the raw data and judge for yourself.
The unedited output of the tool, exactly as generated. The page above is just the readable rendering.