Building with AI, not just designing for it.
Skills, loops, and working practices from shipping AI products — and from running my own. Everything here is in production use.
Agentic engineering is design work: shaping the loop — plan, act, evaluate, revise — until the system earns trust. These are the loops and the toolchain behind everything above.
A git archaeology of the ways of working
The operating system wasn’t designed up front — it accreted, one dated artifact at a time. Git history reconstructs the whole evolution, with quality signals computed from the commit stream.
Service-blueprint skill
Turns discovery-call transcripts into an interactive HTML service blueprint — the process steps, the systems of record behind each one, the common happy path, and where reality deviates from it.
github.com/datapanda/product-discovery-skills ↗Diagram-quality skills
Three Claude Code skills that pressure-test process diagrams — scoring completeness, decomposition, ownership, and data provenance, then separating what the artifact shows from what a human must verify against reality.
github.com/datapanda/product-discovery-skills ↗---
name: assess-diagram-quality
description: Assess a process diagram's
overall quality against a rubric, then
separate what the artifact shows from
what a human must verify against reality.
allowed-tools: Read
model: opus
---
## Scoring rule
Rate each dimension 1–4:
1 Absent · 2 Partial · 3 Solid · 4 Strong
- No score without evidence. Cite the
box numbers or edges for every rating.
- When evidence is mixed, score down.