Skip to content

Listener-State Naming for User-Invoked Agent Skills

Naming a user-invoked skill after the listener's failure state asks the agent to add the missing context, where naming the output only removes words.

Listener-state naming is a skill-design choice: the invocation name describes the state the reader is in, not the shape the reply should take. Matt Pocock's /wait-what is the worked example. You type it when an agent's message did not land, and the agent re-pitches what it just said, adding the context you were missing and using your project's vocabulary (AI Hero, The /wait-what Skill).

The whole skill is one sentence

The file is frontmatter plus one line (skills/productivity/wait-what/SKILL.md):

---
name: wait-what
description: Stop. That last message did not land — re-pitch it.
disable-model-invocation: true
---

Wait — I don't understand where you've got to here. Re-pitch that: give me a little bit of
context, talk in ASD-STE100 Simplified Technical English, and use the ubiquitous language
from `CONTEXT.md`.

The brevity is deliberate. Pocock argues that skills fighting verbosity fail by growing, because a 400-line concision skill still leaves the model verbose (AI Hero).

The disable-model-invocation field makes the skill user-only, on the reasoning that only the human knows when they stopped following. The Codex sidecar sets policy.allow_implicit_invocation: false for the same effect, keeping the skill out of the agent's context until you type it (AI Hero, skills changelog v1.2).

The body then names two register sources: ASD-STE100 Simplified Technical English, a controlled English of 53 writing rules and about 900 approved words (ASD-STE100, About STE), and the ubiquitous language held in CONTEXT.md.

Why it works

The two naming styles license different repairs. An instruction about the output, such as "be concise", is satisfied by deletion, so the model deletes. Deletion carries a measured cost: instructing models for conciseness impaired misinformation resistance in 11 of 17 models tested, with drops reaching 20%, because brevity suppresses the nuance and justification needed to reject false content (Le Jeune et al., Phare: A Safety Probe for Large Language Models, arXiv:2505.11365v4).

An instruction about the listener has no deletion-shaped answer. "I don't understand where you've got to" can only be served by supplying the premise the reader lacked, then saying it plainly. The skill body encodes that order: context first, simplified register second.

Only the first half of that mechanism is measured. The Phare result establishes what brevity instructions cost. The claim that listener-state names escape the same trap is Pocock's reasoning from practice, and no published study tests it directly.

When this backfires

  • The agent is wrong rather than unclear. A re-pitch is a restatement, and model-produced explanations can misrepresent the real basis for an answer: accuracy fell by as much as 36% across 13 BIG-Bench Hard tasks when models rationalized answers driven by biasing features they never mentioned (Turpin et al., arXiv:2305.04388v2). A wrong claim restated in your own project nouns is harder to catch.
  • Typing it is pushback, and pushback moves models. Conformity under user pushback has two drivers, learned sycophancy and inference-time epistemic uncertainty, and both strengthen when the user reads as expert (Guo et al., arXiv:2605.27288v1). You can get agreement where you wanted clarity.
  • No CONTEXT.md exists. Pocock notes the skill still works, but you lose the domain-vocabulary half (AI Hero). What remains is a plain-English request, which is where the brevity penalty lives.
  • The run is autonomous. A disable-model-invocation: true skill never fires on its own, and a scheduled or fanned-out run has no reader to notice they stopped following.
  • You reach for it repeatedly. Three invocations in one thread mean the shared vocabulary was never built. Pocock routes that case to /grill-with-docs and states the limit plainly: the skill repairs one message and does not prevent the next (AI Hero).

Example

Pocock contrasts /wait-what with three names in common use, all of which describe the output (AI Hero):

Invocation What the name describes Repair it licenses
/tldr, /no-fluff, /talk-normal The reply's length or register Cut words, which can drop the premise too
/wait-what The reader's comprehension failure Supply the missing premise, then simplify

Key Takeaways

  • A user-invoked skill's name does work its body cannot: /wait-what carries one sentence and still changes what the agent returns.
  • Naming the output invites deletion, and brevity instructions cost 11 of 17 tested models up to 20% of their misinformation resistance.
  • Set disable-model-invocation: true on any skill whose trigger only a human can detect.
  • A clearer second telling is no evidence the first telling was right, so comprehension repair leaves correctness unchecked.
Emerging — single recent source Last reviewed: 2026-08-06