Matt Pocock, author of the "AI Skills for Real Engineers" project — which has more than 220,000 stars on GitHub — and creator of a YouTube channel with about 347,000 subscribers, released a new skill called /wayfinder. The skill is designed to help humans and agents organize work when the final state of a project is not fully known at the start.
What /wayfinder does
/wayfinder aims to reduce the burden of early-stage planning, which Pocock describes as the "fog of war." It helps when you have a project to begin but can't decide every detail up front. The tool splits planning into multiple threads — for example prototyping, research, and tasks that require human intervention — and orchestrates those threads so an AFK (Away From Keyboard) agent can run larger, longer jobs without the user needing to micromanage session handoffs.
Pocock explains that he used to spend a lot of time scheduling work for AFK agents and converting specs into tickets, and that managing session context (token budget, depth of session) made planning onerous. He built /wayfinder to act as an orchestrator layer that runs planning sessions, prototypes and research threads, and gathers results back into more detailed specs that can then be handed to AFK agents.
The map–ticket–session model
A core design choice of /wayfinder is to use precise, consistent terminology so an agent can clearly map entities and information flow. Pocock identifies three primary concepts:
- map: a document representing the rest of the project and the decisions already made (the "map");
- ticket: the specific piece of work that will be executed in a session;
- session: the actual working session in which an agent performs the ticket's work.
Using these distinct terms prevents ambiguity — if everything is called a "ticket" or labeled inconsistently, agents can become confused and produce unexpected behavior. Defining a small set of leading words lets the agent understand the role of each entity and the information flow between them.
Use cases and examples
Pocock reports using /wayfinder for engineering tasks, non-engineering work and course planning. The system defines ticket types such as grilling (detailed alignment or planning sessions), prototype (create prototypes), research (do research), and task (work a human must perform). This typology makes the approach adaptable to many project types.
He tested /wayfinder on a personal website rearchitecture that involved reorganizing more than 20 years of content. In this context, early decisions open up further exploration on the map, and the fog-of-war metaphor helps guide agents into discovering what’s unknown.
Terminology and a shared language
Pocock stresses the importance of a ubiquitous language between human and agent. Recently he has focused on terminology and is assembling an AI coding dictionary — a graph of terms that describes what an agent, a harness, or a model is — and has updated his courses to rely on that dictionary. All of his skills use the same leading words and assumptions so they interoperate more predictably.
He argues that agents are good at suggesting alternative wordings and at domain modeling, and that creating a consistent vocabulary reduces misunderstandings and improves handoffs.
When to use /wayfinder vs. "grill me"
Pocock recommends using the "grill me" skill when the whole plan can be created in a single session and you simply need alignment before starting — typically for small features or tasks where the path is clear. Use /wayfinder when you sense the fog of war — when you don’t know the full path forward and need to explore, prototype, and research across multiple threads until the way becomes clear.
Summary
/wayfinder provides a structured method for handling uncertainty in early project stages by defining clear entities (map, ticket, session), supporting multi-threaded planning (prototyping, research, tasks), and enforcing consistent terminology between human and agent. Inspired by Pocock’s work with AFK agents and his emphasis on a shared vocabulary, the skill is applied across engineering and non-engineering contexts, including extensive personal projects and course design.
Resources and demonstrations of Pocock’s methods, including the "AI Skills for Real Engineers" project and the /wayfinder documentation, are available via GitHub and his YouTube channel.
