Leading Words

A steering technique — the source's "main thing I want you to get from this talk" — for the common failure where you specify something in a skill, think you were clear, and the agent just doesn't do it. A leading word is a short phrase that packs a lot of meaning into a small space (a term of art the model already has a strong prior for). You place it in the skill's text and repeat it consistently; the agent then re-emphasizes that word in its thinking tokens and its output, and because the word encodes the behavior you want, that re-emphasis pulls the behavior into line.

The worked example: agents tend to "code layer by layer" (all the database, then all the schemas, then all the endpoints, then the frontend) instead of getting a small end-to-end piece working and seeking feedback. Rather than the verbose "don't code layer by layer, build a small slice first," use the leading word "vertical slice" — a well-known development term. It triggers the agent's prior, compresses paragraphs into two words, and gives you a built-in verification signal: you can watch the reasoning traces and see the agent say "we'll do this as a thin vertical slice." If the agent still isn't complying, make your leading words more consistent, more powerful, or find better ones — "English is a wide API," and agents themselves are good at brainstorming candidates.

Leading words are the positive-steering counterpart to Negative Prompting (bound what not to do); both are compact, capable-model steering levers. They're also why a skill can be smaller — a leading word is a high-compression substitute for a paragraph of instruction (Minimal Skill Surface).

Claims

Leading words — short, meaning-dense phrases repeated through a skill — steer agents because the agent re-emphasizes the word in its thinking and output, and that re-emphasis changes behavior. principle — durable: it rests on how models propagate a salient, prior-laden token through their own generation, not on a specific model or harness.
Choose a phrase the model already has a strong prior for (e.g. "vertical slice"), so a couple of words unpack into the full intended behavior. best practice — context: steering toward a concept that has established terminology; a coined phrase with no prior won't carry the same load without more explanation.
Repeat the leading word consistently throughout the skill — inconsistency dilutes it. best practice — context: any skill relying on a leading word to steer; the technique is about consistent reuse, not a single mention.
You can verify a leading word landed by watching it surface in the agent's reasoning traces; its absence there is a signal to strengthen or replace it. best practice — context: iterating on steering when you have access to reasoning/thinking traces; without trace visibility you lose the built-in check.
When steering fails, treat it as a leading-word problem first: make them more consistent, more powerful, or find better ones — agents can help generate candidates. best practice — context: debugging "the agent won't do what I asked"; presumes the instruction itself is sound.
Invoke a well-cited canon's vocabulary as leading words: naming Martin Fowler's refactoring smells (mysterious name, feature envy, data clumps, primitive obsession, repeated switches…) in a code-review skill, one sentence each, unlocks capability already deep in the model's priors — it repeats the term back ("I found some message chains") and acts on it. best practice — context: a second, independent application of the technique by the concept's originating author; the canon must be old and well-cited enough that the terms are genuinely in-prior ("~10 lines, outrageously useful"). A coined smell name would not carry the same load. See Shared Review Surface for where the smells land — the diff.
Steer a fuzzy behavioral boundary with a leading-word distinction: telling a grilling agent to separate "facts" (found by exploring the code) from "decisions" (made by the user) stopped it grilling itself and running past the human. best practice — context: fixing an agent that conflated two roles; a pair of prior-laden nouns drew the boundary that a paragraph of instruction hadn't. Reported as especially fixing misbehavior on one model (the source names Fable).

Negative Prompting — the complementary lever: leading words steer toward a behavior; negative prompting bounds it. Both are compact steering for capable models.
Minimal Skill Surface — a leading word compresses a paragraph into a phrase, so it directly shrinks the skill surface.
Concise Prompting — same "pack meaning, cut length" instinct at the prompt scale.
Step Isolation — the other steering lever in the talk: when a leading word isn't enough, restructure so the agent does more leg work on the step.
Skill Authoring Checklist — pillar 3 (Steering).
Resolution-Typed Tasks — the ticket-type labels (HITL/AFK, research/prototype/grilling) are themselves compact prior-laden terms that steer how a ticket is resolved.
Distillate: Building Great Agent Skills: The Missing Manual
Distillate: Skills v1.1: Wayfinder, the SDLC flow, and naming the artifact right — two further applications by the same author: Fowler's refactoring smells as leading words in /code-review, and a facts-vs-decisions distinction that stopped a grilling agent grilling itself.
Context-Independent Review — where the Fowler smells actually land: they are the fallback standard a fresh-context reviewer checks against when the repo documents no conventions of its own.
Distillate: The whole flow, end-to-end: the smart zone is the unit of work — the smells shown in their fallback role: "if it doesn't detect any [standards docs], then it uses some classic ones kind of from Martin Fowler."

Leading Words

Claims

Related

Linked from