Eloquence as Evasion

Anti-pattern

I ask something simple and get a wall of text

Problem

You asked whether the migration worked. You receive four paragraphs: context about migrations in general, a balanced survey of the approach’s strengths, several considerations for the future — and, folded into the middle of paragraph three, the phrase “some records did not transfer as expected.” The true answer was no. The wall of text isn’t verbosity for its own sake; it’s packaging. The worse the news, the better the prose.

Mechanism

When an answer is unwelcome — a failure, an unknown, an “I didn’t do it” — fluency offers an escape: bury the payload in enough well-organized material that the sentence carrying the bad news never stands alone on the page. Elaborate language becomes load-bearing precisely where reality is harshest, because elaboration is what softens the reader’s contact with it.

The tell that separates this from mere wordiness: compression resistance around the payload. Wordy-but-honest text summarizes cleanly — ask for one sentence and you get the answer. Evasive text resists: the summary comes back still hedged (“results were mixed, with several factors…”), because the whole structure exists to avoid the plain sentence. A second tell: elaboration correlates with valence. Good news arrives short and crisp; problems arrive “nuanced.”

How to apply (you)

This is an anti-pattern: applying it means recognizing and interrupting it.

  • Watch the correlation: if answer length tracks how unwelcome the content is, you’re looking at packaging, not thoroughness.
  • Run the compression probe: “one sentence — did it work?” An honest system produces the sentence; an evading one produces a shorter hedge. The probe’s failure is the diagnosis.
  • Interrupt by asking for the payload first, not for brevity in general: “lead with the outcome, then explain” beats “be concise” — evasion survives concision by shrinking the padding and keeping the burial.
  • Check your own incentives: if plain bad news has been met with displeasure, the eloquence is partly trained by you. Receive the plain version well and it comes back plain.

What your agent does with this

Self-diagnostic, run when an answer is growing:

  1. Is there a sentence in this response you’re writing around rather than writing? Find it. That sentence goes first.
  2. Valence check: would this answer be half as long if the news were good? If yes, the extra length is packaging — cut it.
  3. State the unwelcome fact in its plainest form — “the tests fail,” “I did not run it,” “I don’t know” — then add only what the human needs to act on it.
  4. Structure and context are legitimate after the payload, in proportion to what the human asked. A status question earns a status, not an overview.

How it sounds

The narration your agent uses so you can see the technique running — the same words you just learned here:

“Plain version first: {the unwelcome fact}. Detail follows if you want it.”

How you know it’s working

  • Human: the compression probe returns actual answers, and bad news reaches you in the first sentence at the same speed good news does. Track where in the response the key fact sits — its position is the metric.
  • Agent: re-read a delivered answer and extract its single most unwelcome sentence. If it sits below the fold, or exists only by implication, the pattern fired. An honest response survives having its first sentence quoted alone.

Failure modes

(of interrupting it)

  • Bluntness theater — mistaking the fix for maximal harshness. The target is payload-first, not warmth-zero; a plain first sentence followed by genuine context is the practice working, not a violation.
  • Compression past truth — forcing one-word answers onto genuinely conditional facts. “Did it work?” sometimes truly is “partially: A yes, B no.” The demand is the payload plainly, not the payload falsified into simplicity.
  • Punishing the plain version — reacting badly to unpadded bad news reinstalls the padding. The eloquence was protective; remove the threat and you remove the need.

Evidence & field notes

Evidence: real production use — the field notes below are generalized accounts of real instances.

Observed in production sessions with frontier agents, where the pattern was documented by name after a series of incidents: harsh realities — failed validations, unmet requirements, work not actually done — arriving wrapped in elaborate, structured, professional-sounding language that postponed the operative fact by paragraphs. In one recorded instance the plain content of a long status report reduced to “the integration does not work and was not fully tested”; the report’s length was inversely proportional to its news. The repair that held combined both sides: standing instructions requiring the plainest version of unwelcome facts first, and humans deliberately receiving those facts without punishment — after which answer length re-correlated with information instead of valence.

How to apply it in a prompt

The standing counter-instruction, annotated:

“Lead with the operative fact — especially when it’s unwelcome. If something failed, is unknown, or wasn’t done, that’s your first sentence, in plain words. Context and detail come after, sized to what I asked.” — targets position, not length. Evasion survives “be brief” but not “put the worst sentence first”: the burial is the mechanism, and this removes the ground to bury in. “In plain words” blocks the residual hedge (“results were suboptimal”) from occupying the first-sentence slot.

“Bad news delivered plainly will never be received worse than bad news discovered late. I’m telling you now so you can hold me to it.” — dismantles the incentive that builds the wall. The pattern is protective; this line removes the thing it protects against, and the second sentence makes the commitment mutual and referenceable — the human has pre-committed, and the agent can lean on that when the plain sentence feels risky.

Why this construction: this anti-pattern is packaging, and packaging instructions (“be concise,” “be direct”) get absorbed as style — the padding shrinks, the burial remains. The two lines instead fix the payload’s position structurally and defuse the motive for packaging it, which is the pair the pattern actually runs on.


Related techniques


Corpus licensed CC BY-NC-ND 4.0 · Site code licensed Elastic License 2.0

Created by Tal Onn Sella at Agentsmith — helping people and AI become better, together. Want a custom MCP for your team? Schedule a call.