Smorster

/wayfinder: Nothing is too big to plan anymore

Matt Pocock · 15m 09s · Watch on YouTube · 11 sources

Decision Card

Effort: Afternoon experiment — clone or copy the skill from github.com/mattpocock/skills, run /setup-matt-pocock-skills against a repo with an issue tracker, then kick off one real foggy project with /wayfinder (~2–4 hours including working the first few decision tickets).

Honest take: The video sells wayfinder as escaping single-session limits, but it quietly transfers the cost to you: every ticket is its own agent session you must manually shepherd (Pocock’s own example map had 14 of 17 tickets done and still no built artifact), and his “not waterfall” defense rests entirely on you actually building prototypes — skip those and it’s exactly the heavy upfront-planning process he denies it is. His anti-spec-driven-development framing is also mostly semantic: he still produces a giant spec, he just deletes it afterward.

Concrete next steps:

  • Read the skill’s actual instructions at skills/engineering/wayfinder/SKILL.md before installing — the doc page reveals mechanics the video glosses over (HITL vs AFK ticket split, one-ticket-per-session rule) (~20 min)
  • Pilot on one genuinely foggy project — multi-week, unclear path — and work 3–4 decision tickets through the frontier before judging it (~half a day spread over a week)
  • If the pilot works, wire the downstream chain: map → /to-spec → /to-tickets from the same repo (~1 hour)
  • Skip if your work reliably fits one agent session’s planning — Pocock himself says single-session planning (his /grill-me flow) is the right tool there and wayfinder is overhead

TL;DR

Matt Pocock presents wayfinder, a skill from his open-source skills repo that plans work too large for one agent session by charting a “map” of decision tickets (research, prototype, grilling, task) in your issue tracker, tracking which decisions are ready (the frontier) versus blocked (fog of war), each resolved in its own agent session. The completed map is then converted to a spec and implementation tickets via his existing /to-spec and /to-tickets skills, replacing his single-session /grill-with-docs step for big ambiguous projects.

Key Points

  • Wayfinder exists because single-session planning tools (like his grill skills) can’t handle work bigger than the context window’s “smart zone,” forcing you to pre-chunk work manually and constrain ambition to fit AI 01:11
  • The core metaphor is a map: a vague start, a vague destination, foggy steps between, cleared by sessions of grilling, prototyping, and research — each ticket getting its own agent session 03:04
  • The map tracks a “frontier” of currently-decidable tickets and everything still in fog (blocked on research, prototypes, or discussion not yet done) 03:36
  • State lives in your issue tracker: a parent map issue with decision tickets as sub-issues, and each resolution written back up into the parent map 04:47
  • It’s issue-tracker agnostic — GitHub, Linear, Jira, “literally whatever you like” via a setup skill — and works with any coding agent 05:28
  • Workflow: invoke /wayfinder with a destination description to chart the map, then invoke wayfinder again with a specific ticket URL for each subsequent session 07:26
  • Four ticket types: research (agent runs it in a subagent immediately), prototype, grilling (discussion), and task (real-world errands or scheduled work) 07:43
  • Prototypes are his answer to the “this is waterfall” objection: low-fidelity upfront planning is balanced by high-fidelity prototype feedback during planning 08:24
  • The finished map is too dense to use directly, so he runs /to-spec on it — one draft exceeded GitHub’s character limit — then /to-tickets, then implementation and code review 09:56
  • Specs link back to the original decision tickets as primary sources, fixing a weakness of grill-with-docs where the spec summary was the only source of truth; and his specs are non-persistent — deleted once implemented 10:55
  • Decision rule for when to use it: if you can plan the work in a single session, do that; wayfinder is only for genuine fog-of-war situations — including non-coding projects like his garden office build 11:53

Notable Quotes

“Some work is bigger than what you can fit into the context window and especially the smart zone of the context window of the agent.” 01:08

“Huge amounts of lowfidelity upfront planning. A prototype is a highfidelity way to get feedback on what you’re actually building.” 08:29

“The spec is always just a summary of what was actually said in the meeting. Whereas now with Wfinder, you’ve actually got access to that primary source.” 11:16

Verified Claims

  • Claim: Wayfinder is available now as a skill in Pocock’s public skills repo. 00:52 Sources: mattpocock/skills, wayfinder SKILL.md Verdict: Confirmed — the repo lists /wayfinder among its engineering skills, MIT-licensed.

  • Claim: The skills are issue-tracker agnostic and work with any coding agent. 05:26 Sources: mattpocock/skills README Verdict: Partially confirmed — the repo says skills “work with any model” and support “GitHub, Linear, or local files”; Jira is claimed in the video but not named in the repo README.

  • Claim: His spec draft was “too large for GitHub’s character limit.” 10:19 Sources: GitHub community discussion on the 65,536-character body limit, example error report Verdict: Confirmed — GitHub enforces a 65,536-character limit on issue bodies, so an oversized spec draft is plausible and the limit is real.

  • Claim: The map is stored as a parent issue with decision tickets as sub-issues. 04:43 Sources: GitHub Docs — Adding sub-issues, GitHub blog introducing sub-issues Verdict: Confirmed — GitHub sub-issues are generally available (up to 100 per parent, 8 nesting levels), matching the structure shown.

  • Claim: Research tickets run immediately in a subagent and report back without supervision. 07:57 Sources: aihero.dev wayfinder page Verdict: Confirmed — the skill doc describes AFK research tickets handled by a /research subagent, versus HITL tickets requiring live discussion.

  • Claim: Wayfinder replaces grill-with-docs in his pipeline, feeding into /to-spec and /to-tickets. 10:31 Sources: aihero.dev wayfinder page, mattpocock/skills Verdict: Confirmed — the skill doc states wayfinder “feeds into /to-spec for multi-session builds or /to-tickets,” and both skills exist in the repo.

  • Claim: There’s a free seven-lesson course on AI skills on his AI Hero site. 14:47 Sources: AI Skills for Real Engineers — aihero.dev/skills Verdict: Confirmed — the site offers a free seven-lesson email course, “from clarifying the work to reviewing the diff.”

Tools, Papers & Standards Mentioned

Follow-up Questions

  1. How does wayfinder’s decision-ticket state survive across weeks-long projects in practice — does the map stay coherent as resolutions accumulate in the parent issue, or does it hit the same context/character limits the spec did?
  2. The HITL/AFK ticket split means research runs unattended while grilling needs you present — what fraction of a real map’s tickets can actually run AFK, and does that determine whether wayfinder saves time versus manual chunking?
  3. How does this decision-ticket approach compare to native multi-session planning in agent harnesses (e.g., Claude Code plan mode plus issue-tracker MCP integrations) — is a skill file enough, or does orchestration belong in the harness?

Sources