Splitting Skills

Split one agent document into two only when the cut earns the load it spends: by sequence, when later steps tempt the agent to rush the current one, or by invocation, when a distinct leading word or another skill needs to reach part of it independently.

Evidence: moderate

Every split spends one of the two loads in context load versus cognitive load: a new model-invoked skill adds an always-loaded description, and a new user-invoked one adds something for the human to remember. Matt Pocock's writing-for-agents skill therefore names only two cuts that earn it.

Splitting by sequence applies when a run of steps lets the agent see work ahead of it and rush the step in front, which is premature completion. Keeping the later steps out of view drives more legwork on the current one, but only if the split crosses a real context boundary.

Splitting by invocation applies when part of a skill has its own leading word that should trigger it independently, one you actually use in prompts, or when another skill must reach it. That is how grilling became its own model-invoked primitive behind several thin orchestrators.

Neither cut is "the file is long". Length alone is sprawl, cured by disclosing reference behind pointers rather than by creating more skills.

Boundary: merging is also a decision. Combining two sequences exposes the first to the second's steps; keep them apart when the first has a fuzzy completion criterion.

---

Sources

• https://github.com/mattpocock/skills/tree/main/skills/productivity/writing-for-agents (read 2 Oct 2026, v1.2.3)