Pocock, never start with /wayfinder, one-shot the tiny diffs
His rule for picking between a planning session, a full map and just letting the agent write the code comes down to expected diff size.
Matt Pocock posted a decision rule for the moment before you hand work to an agent. Three options, as he frames them: run a /grill-with-docs session, run /wayfinder, or just one shot the code. The deciding input is how big you expect the diff to be.
Never start with /wayfinder
Two rules come out of it. First, if the diff is tiny, do not grill. Pocock's reasoning is that /grill-with-docs exists to save you from the case where the agent built the wrong thing, and when the thing being built is tiny, that cost "drops to nothing". One-shotting is fine for small changes.
Second, never start with /wayfinder. The failure mode he describes is specific. You start grilling, the solution turns out to be far simpler than you thought, and you are now holding a map and a set of tickets for a problem that did not need them. The artifacts outlive their usefulness before you have written a line.
The ordering
So for a medium to large expected diff, always start with a /grill-with-docs session. If planning starts to get out of hand during that session, escalate in place with "/wayfinder let's turn this into a map".
For working engineers, the command names are Pocock's and the shape is the point. Cheap planning first, heavier planning artifacts only once the problem has demonstrated it needs them, and no planning at all when the change is small enough that being wrong costs a re-prompt. The common failure is picking the planning depth up front, based on how important the task feels rather than how large the change turns out to be, and then being locked into a plan you have to maintain.
It is also worth noticing what the rule does not do. It does not try to classify tasks by type or risk. The only signal is expected diff size, and the escalation path is one sentence typed mid-session, which means the cost of guessing low is small and the cost of guessing high is a map you throw away.
This is Pocock's own working rule, not a benchmark result. Treat it as one practitioner's heuristic.