Skip to content

Matt Pocock — Full Walkthrough: Workflow for AI Coding (AI Engineer 2026)

Source: https://youtu.be/-QFHIoCo-Ko — 1h36m25s workshop at AI Engineer 2026 conference. Ingest was user-curated: only agentic-coding workflow content distilled; live-coding filler, audience surveys, jokes, and breaks omitted.

Raw: raw/transcripts/matt-pocock-ai-coding-workflow-2026

Entity

  • matt-pocock — TypeScript-educator-turned-AI-educator; Total TypeScript / AI Hero; shipped Sandcastle harness.

Concepts minted (one per user theme)

LLM constraints - smart-vs-dumb-zones — context occupancy determines model quality; keep system prompt tiny - memento-principle — reset over compact; deterministic empty state beats noisy summary

Alignment phase - grill-me-skill — AI interviews human until design concept is shared - prd-as-destination — PRD is a reference marker, not a compiler input; don't over-polish; expect doc rot

Task planning / execution - tracer-bullets — vertical slices cross every layer so each slice emits feedback - afk-implementation — night-shift parallel execution via Sandcastle once the Kanban is curated - tdd-feedback-loop — feedback-loop quality is the ceiling on agent output

Codebase architecture - push-vs-pull-standards — push rules to reviewers (always on), let implementers pull skills on demand

Notable tensions with existing wiki

  • Memento ↔ compaction. Pocock argues /clear beats /compact; context-compaction (via Zoneraich's reverse-engineering of claude-code-master-loop) describes H2A auto-compaction at ~92% as load-bearing. Both can be true: H2A keeps autonomous long-horizon runs alive, whereas Pocock's manual workflow externalises plan state into PRD/issue files so clearing is cheap. Marked in memento-principle frontmatter.
  • Spec-driven vs code-driven development. Pocock explicitly rejects "specs-to-code" — pairs as a counterpoint to any wiki page that treats PRDs as the primary artifact (see prd-as-destination).

Framing

A rare "known software craftsman, applied to agents" source. Most of the wiki's 2026 ingests come from lab/infra/tool-builder angles; Pocock comes from the teaching fundamentals tradition and keeps arguing that agents don't replace Fowler / Hunt & Thomas / Ousterhout, they reward them. Useful anchor when evaluating new agent-workflow claims: does it respect the smart zone? Does it produce deep modules? Does it run a real feedback loop?