Emerging Ideas for AI-Assisted Software EngineeringSeven field guides from Alpha Compose
Alpha ComposeResourcesAll resources

Matt Pocock's current AI engineering workflow

From fog
to frontier.

The interesting shift is not a better prompt. It is a sequence of human decisions, durable artifacts, small executable slices, and independent feedback loops.

ideagrillprototypespecticketsreviewdestination

01 / Current spine

A five-step chain, with judgment before autonomy

In Matt's July 2026 walkthrough, the everyday route is now grill-with-docs, to-spec, to-tickets, implement, and code-review. Wayfinder sits upstream when the work is too large or too foggy to specify in one session.

01/grill-with-docs

Create shared understanding

Interrogate the idea, sharpen the project's language, and capture durable context before writing a specification.

02/to-spec

Describe the destination

Synthesize the decisions already made. The specification records shared understanding instead of inventing it.

03/to-tickets

Cut tracer-bullet slices

Turn the spec into vertical tickets with explicit blocking edges, ready for sequential or parallel execution.

04/implement + tdd

Build through stable seams

Implement one behavior at a time, with tests aimed at public interfaces and regular type and test feedback.

05/code-review

Review on two independent axes

Fresh agents check standards and spec fidelity separately, then human QA judges the product in use.

Matt Pocock showing the five-stage skills workflow08:23 ↗

The current main flow. The sequence keeps alignment, specification, slicing, implementation, and review as distinct jobs.

02 / What changed

The workflow moved closer to reality before the spec

Earlier emphasis

Conversation - PRD - issues - agent loop

The April workshop established the durable skeleton: shared understanding, vertical slices, TDD, fresh review, and human QA.

Current emphasis

Research and prototypes can precede the spec

When words are too low fidelity, the agent builds an artifact to react to. The validated answer then becomes primary material for the spec.

Large efforts

Wayfinder resolves decisions before planning delivery

It does not pre-slice an unknowable project. It advances only the visible frontier and lets later questions emerge from resolved ones.

Matt Pocock showing a gallery of prototypes used before specification
03:50 ↗

Raise fidelity only where conversation fails

Cheap prototypes turn taste and behavior into observable evidence.

03 / Wayfinder

Do not pretend the whole route is visible

A Wayfinder map is deliberately incomplete. It records the destination, indexes resolved decisions, exposes what can be decided now, and leaves still-vague territory as fog.

startresearchgrillingprototypetaskspecdestination
01

Destination

The decision, spec, or change the map is finding a route toward.

02

Decision tickets

One precise question per child issue, sized to one agent session.

03

Frontier

Open, unblocked, unclaimed tickets that are genuinely ready now.

04

Fog

In-scope questions that cannot yet be stated precisely enough to ticket.

05

Out of scope

Work consciously beyond this destination, closed so it never leaks onto the frontier.

What is most interesting

The map stores uncertainty honestly. It advances the boundary of the known instead of laundering fog into a confident checklist.

04 / John Rengwu

The planning memory becomes a planetary instrument

John Rengwu released wayfinder-maps after posting a planetary view of the process. The open-source tool reads a local Markdown adapter, reports frontier status, lints the map, and renders it as a star field.

Planetary demo ↗

Not just a skin

Legibility is part of correctness

  • Status is derived from answers, rulings, and claims instead of stored twice.
  • Lint is fsck for dangling edges, cycles, duplicate tickets, stale claims, and contradictory closures.
  • Spatial memory persists because the deterministic layout keeps tickets where the operator learned them.
  • The frontier glows so readiness is visible without opening every issue.
Inspect the repository ↗

05 / Two layers

Matt designs the decision process. Rengwu designs its observability.

What owns the truth?

A shared issue-tracker map indexes decisions and child tickets.

Structured local Markdown owns the facts; the viewer is a read-only projection.

Keep one canonical graph. Treat every interface as a view, not a second database.

How is readiness seen?

Native blocking links expose the frontier in the tracker's UI.

Gold stars, glow, position, and a status command make readiness spatial and immediate.

Machine-readable dependency edges should also be glanceable to a human supervisor.

How does the system resist drift?

The map stays low resolution and points to the ticket that owns each decision.

Status is derived from file content; lint behaves like fsck for the planning memory.

Avoid duplicated status, then verify the remaining contract continuously.

Where does the human belong?

Scope, grilling, prototypes, and QA remain explicitly human-in-the-loop.

The planetary view compresses fleet state so a human can choose where to look next.

Automate navigation and evidence. Preserve judgment at high-leverage decisions.

06 / Recommended combination

A navigable decision graph feeding a governed execution fleet

The strongest synthesis adds Kun Chen's runtime after Matt and Rengwu's planning layer. Wayfinder decides what the work means; the star map makes readiness legible; FirstMate can dispatch only the implementation tickets that have genuinely cleared the frontier. Mycelium adds a distinct layer when those tickets need to become application code: an executable control-flow graph with schema-bound cells that agents can implement in bounded contexts.

  1. 1
    Chart

    Use Wayfinder for a foggy destination. Resolve research, prototype, and grilling decisions.

  2. 2
    Observe

    Render the map as a deterministic star field and lint it after every structural change.

  3. 3
    Compile

    When the route is clear, produce the spec and tracer-bullet implementation tickets.

  4. 4
    Dispatch

    Let FirstMate lease worktrees only for ready slices, preserving claims and dependency order.

  5. 5
    Verify

    Run TDD, two-axis fresh review, and human QA before delivery changes the shared truth.

Do not merge the layers into one mutable dashboard.

Keep the decision map, execution state, and visual projection as separate contracts. Integration should pass references and events, not duplicate truth.

Continue with the Kun Chen comparison Continue with the Mycelium contract layer

07 / Critical verdict

The workflow's real product is disciplined attention

What transfers well

Interview before specification. Prototype where words are weak. Keep tickets vertical. Test through stable seams. Review spec and standards independently.

What needs restraint

A heavy map is wasteful when one sharp session will do. Skills can conflict. Prototypes can seduce. Visual polish can disguise an incoherent graph.

The durable principle

Use automation to move evidence and state. Reserve human attention for scope, taste, architecture, and acceptance.