← narwal.one/Second Brain
SecondBrain
Ask the Brain
Index/Conceptupdated Sat Sep 26 2026 08:00:00 GMT+0800 (Philippine Standard Time)

Leading Words

promptingsteeringskillsagentsreasoning-traces

Leading Words

Matt Pocock's name (he nods to Leitwort from literary theory) for the core steering technique in agent skills and prompts: specific terms that pack dense meaning into a small space, chosen so they trigger the model's prior. Put the leading word in the skill/prompt text and the agent repeats it back to itself — in its thinking tokens and output — re-emphasizing the behavior you wanted.

The canonical example

Problem: agents code layer by layer (all database → all schemas → all endpoints → all frontend) instead of getting something small working end-to-end. Rather than a paragraph of "don't code layer by layer, build something small first…", use the leading word "vertical slice" — a well-known dev term that carries the whole behavior — and repeat it consistently through the skill.

Why it's more than a prompt trick

  • It's verifiable. You know the steering worked when the leading word shows up in the reasoning traces ("okay, we'll do this as a thin vertical slice"). This gives skill-authoring a feedback signal short of full evals.
  • It's compression. A leading word is the token-efficient form of an instruction paragraph — the same economy that motivates minimal SKILL.md files (Skill Checklist (Pocock) pruning) and cheap context (Context Engineering).
  • "English is a pretty wide API." Pocock frames leading-word hunting as function discovery: many candidates exist, and agents are good at helping you find them. If the agent isn't doing what you want, make the leading words more consistent and more powerful before adding more instructions.

2026-09-26 — Where leading words come from (AI Skills with Matt Pocock (The Pragmatic Engineer))

  • Origin story: the specs-to-code failure sent Pocock to The Pragmatic Programmer (unread, still in plastic) → software entropy, don't outrun your headlights, programming by coincidence, tracer bullets. He started using those phrases in prompts and noticed the agent saying them back ("Okay, I'll turn this into a tracer bullet") in its reasoning traces.
  • Mining method: treat classic books as a leading-word quarry — 25-year-old texts are in the model's prior. Tracer bullets / vertical slices (Pragmatic Programmer) fix the agent's habit of building the whole DB layer, then the whole app layer, then the whole UI, and only integrating at the end; deep modules (Ousterhout); ubiquitous language (Evans).
  • Leitwort — the literary term he borrows: a phrase repeated a few times in the skill or prompt to change behaviour.
  • It's jargon, rehabilitated (Gergely Orosz's framing): jargon excludes newcomers but compresses meaning between professionals — exactly the property you want with an agent. Kent Beck and Ward Cunningham kept a thesaurus on the desk while naming patterns; the same search for the precise word is now an agent-steering skill.
  • Project-specific leading words are the bridge to Ubiquitous Language: once the prior-level terms work, coin domain terms (e.g. materialization cascade) with the agent.

Vault connections

Existing vault frameworks already rely on this mechanism without naming it — Sandeep's named frameworks (TRAP Framework, DRAG Framework, AIM Protocol) and coinages like Vibe Coding or Code Is Free work because they're dense, prior-triggering handles. Leading Words is the explicit theory of why naming things well steers both humans and models.

Cross-links