Matt Pocock live-demos Wayfinder — the evolution of his Grill Me skill, for the pre-spec stage of software work. Wayfinder charts a “map” from the current codebase to a destination (a locked set of decisions / a spec), breaks a foggy requirement into parallel tickets, and spawns agents to work them. The demo builds a TikTok creator feature for his Course Video Manager app — 8 of 9 map tasks in ~75 minutes, without writing implementation code.

What Wayfinder is

  • “A sort of Grill Me on steroids” — an orchestrator over grilling sessions instead of a single one
  • The map’s job is to find the way and resolve open decisions, NOT to build the thing — output is a spec handed to an implementation agent
  • Works on Opus 4.8 medium (“perfectly fine for grilling, I don’t need Fable for this bit”)

The map + ticket types

  • Creates a real GitHub issue with sub-issues and blocking relationships (not GitHub-dependent — Linear/ToDoist/Jira all work)
  • Decision tickets (grilling sessions) — resolved by conversation
  • Research tickets — Wayfinder auto-spawns sub-agents to tackle these now, findings committed to branches
  • Prototype tickets — build real A/B UI variants in a work tree to raise fidelity
  • Progressive disclosure: ticket → thick docs → summary, walked down only when needed

Prototyping (the highlight)

  • For “what should it look like” questions, Wayfinder builds actual UI so you look at screens, not text
  • First pass: “New TikTok” button variants — he picks B (grid of cards): “stuff that I would not have put time into”
  • Integration beats generation: he rejects the AI layouts and asks the agent to explore the REAL editor code — reuse the existing reducer, componentize shared logic into the new view
  • Prototyping burns context fast (“used to things going to like 200k”)

Decisions surfaced by grilling

  • Enum over boolean: format: standard | short — “what happens if we add a third type? Three booleans = six possible states; an enum gives you literally three”
  • No pitch relationship for TikToks: “a TikTok video should just literally be: okay, I’ve got an idea, let’s go” — low friction by design
  • No backfill: assume all existing videos are landscape, no migration
  • Render pipeline at arm’s length: extract the Remotion project to a standalone repo, CVM invokes via shell-out, defer the pull-in
  • The TikTok blocker: TikTok’s API treats apps as un-audited until they build the full mandated content-sharing-guidelines UX — “damn, this is brutal.” Workaround: Buffer via Dropbox + Zapier

The model/harness/environment framing

  • “Swap out the model and it’ll be great” is a small part of the picture — you can also make the harness better or the environment better, and that’s free. “I would say 50/50”

The full flow

  • Wayfinder → to-spec → to-tickets → implement → code review → human review
  • Complete the map → to-spec turns it into a spec → close the map → implementation sub-tasks, each in a separate context window with an AFK agent → code review on the whole diff → human review
  • Decision tickets resolve by deciding; implementation tickets reify those decisions in code
  • Reference: a prior Wayfinder spec produced a ~5,000-line PR via a GitHub Action working each ticket in a separate session

“So much of misalignment from AI is AI not asking these questions, not understanding what your values are, what you prioritize over other stuff.”

“You’re the lead. They’re the junior. You need to be on top of them. You need to be driving it.”

Watch on YouTube