Provider treatment of operator-communication failures — 2026-08-29

This frozen report records why Commonplace kept local operator-communication controls even though major model providers already treat clear user-facing communication as a general requirement. It preserves the evidence available on 2026-08-29 for later evaluation of those controls. It does not establish a new methodology rule; the operative rule and skill are linked below.

Result

The problem was not simply omitted by the providers. Anthropic and OpenAI both state that answers should be clear, direct, and calibrated to the reader. Both also give developers controls for response length and style. Anthropic's current Claude Fable 5 guide goes further: it explicitly describes agents carrying working shorthand and unseen context into a final answer, and tells them to reintroduce or discard vocabulary that the reader did not share.

That provider recognition does not make the Commonplace controls redundant. Provider guidance supplies a broad default. It cannot know which terms have canonical definitions in this project, which terms an agent coined during one task, or when a technical finding should remain intact as evidence while a separate explanation serves an operator. Provider documents also describe intended or steerable behavior, not a guarantee that every model and harness will produce it throughout a long run.

The local intervention therefore addresses two different needs:

  • The always-loaded rule states the project's ordinary communication contract.
  • The optional skill performs a separate operator explanation when a complex technical finding cannot satisfy both evidential and conversational needs in one presentation.

This is a plausible control, not yet a demonstrated solution. Its value must be tested against representative findings and long agent runs.

Inputs and method

The report compares four kinds of evidence:

  1. Provider statements of intended model behavior.
  2. Provider model-specific prompting and product guidance.
  3. Several public user reports that describe the observed failure.
  4. The two Commonplace changes made in commits 80a5f3a0 and 73f530b4.

The external pages were read on 2026-08-29. They remain live sources and may change; this report freezes the interpretation, not the pages. Public forum posts show recurring failure shapes but do not estimate prevalence. The initiating LinkedIn example was supplied in the working conversation without a source URL and was not independently verified.

The synthesis was produced by Codex under the repository's AGENTS.md; the runtime did not expose a more specific model identifier. No model-generated claim is treated as primary evidence.

The reported failure has several independent axes

"Use less jargon" or "be concise" compresses several failures into one global style preference. The complaints and provider guidance distinguish at least five axes:

Axis Failure Why a brevity instruction is insufficient
Amount The answer includes material the operator does not need. A shorter answer can still contain the wrong material.
Density Sentences, fragments, or arrow chains carry too many relations at once. Removing words can make the remaining text harder to parse.
Order Implementation detail appears before outcome, consequence, or decision. The same content in fewer words can still start in the wrong place.
Audience The answer continues an internal working thread the operator did not see. Length does not restore missing context.
Vocabulary The answer relies on task-local shorthand or inherited terms with no shared definition. A coined term can be short, precise to its author, and opaque to everyone else.

The last axis is the main Commonplace addition. Ordinary language, established domain language, and project terms with canonical definitions are shared vocabulary. They need not be removed merely because they are specialized. A human operator can pay the learning cost once and then benefit from the compact term. The risky category is vocabulary that exists only in the current agent's work or has spread through artifacts without acquiring a canonical definition.

This distinction cuts across expertise. An expert can understand difficult domain terms while still being unable to decode a label invented twenty tool calls ago. Conversely, replacing every project term with generic language can lose precision, create synonym drift, and repeatedly charge the operator for a definition they already know.

What users report

The strongest consolidated complaint is Claude Code issue #77136. It describes verbose, jargon-heavy, and invented phrasing; forced metaphors; answers that become cryptic when asked to be concise; and style instructions that drift during a session. The issue summarizes a larger Reddit discussion, but its vote and comment counts are evidence of one cluster of users rather than a population measure.

Separate threads report the same distinctions from different angles:

  • An Opus 5 discussion says project-level style rules stop holding after meaningful work and that buzzword-heavy explanations can make otherwise useful results difficult to review.
  • A cryptic-phrasing discussion describes dense shorthand and missing transitions; several participants say they use another model to translate the output.
  • A Claude Code discussion contains the more specific observation that an agent can invent shorthand in its working context, repeat it until it becomes locally familiar, and then fail to notice that the shorthand was never user-facing. The same discussion notes that an aggressive brevity style can worsen this by condensing prose into invented labels.

These reports support the existence and shape of the problem. They do not show that every user, model, or harness encounters it, and their explanations of the underlying model mechanism remain hypotheses.

How Anthropic describes the problem

Anthropic addresses the issue at three layers.

At the intended-behavior layer, Claude's Constitution says response length should follow the request and should avoid padding, excessive caveats, and unnecessary repetition. The same document says Claude's actual behavior may depart from these ideals. This frames appropriate length as a contextual judgment and acknowledges an implementation gap.

At the model-guidance layer, the Claude Opus 5 prompting guide says the model's user-facing answers are longer than those of prior Opus models. It separates reasoning effort from visible response length, recommends an explicit concision instruction, tells agents to lead with the outcome, and warns that old verification instructions can compound the model's own verification and add narration. This account mainly treats amount, order, and harness interaction.

The Claude Fable 5 prompting guide matches the observed failure most closely. It says that extended agentic work can lead to dense shorthand, deep implementation detail, references to work the user never saw, and overly technical phrasing. It explicitly separates readability from concision and says shortening should select details rather than compress sentences into fragments, arrows, abbreviations, or jargon. Its central vocabulary instruction is: "The vocabulary you built up while working is yours, not theirs." It recommends treating the final response as a fresh orientation for the reader, reintroducing any necessary terms, and using a dedicated send_to_user tool when the harness must separate exact user-facing messages from working narration.

At the product layer, Claude Code release 2.1.237 added a built-in Concise output style that leads with results and skips preamble and narration. Its public description addresses amount and order, not the status of project and task-local terms. The release still shows that Anthropic treats the harness as a control surface rather than expecting training alone to settle the problem.

The evidence therefore rejects the simple "Anthropic forgot" explanation. Anthropic recognizes the general problem and, for Fable 5, the case where working vocabulary was never shared with the user. Its published remedy still includes operator prompts, model-specific migration work, and harness scaffolding. It is not a claim that the base model will always do this unaided.

How OpenAI describes the problem

OpenAI also treats clarity as both a model target and an application-level control.

The current OpenAI Model Spec says answers should be lucid, succinct, organized, clear, and direct. It says a professional default does not mean business jargon. OpenAI's explanation of how the Model Spec is used is important to interpreting those rules: the specification is partly aspirational, production models do not yet fully reflect it, training cannot cover every context, and models can generalize differently from the intended behavior. The same explanation argues that explicit policies remain useful because high-level goals underdetermine behavior and runtime context is limited.

The current GPT-5.6 model guidance offers a text.verbosity control but warns that broad brevity instructions can make an answer too short. It recommends specifying what a short answer must preserve: the conclusion, necessary evidence, material caveats, and next action. It also tells developers to keep style guidance when it encodes a product requirement or corrects a measured failure, and to validate prompt changes on representative work.

Some earlier model-specific guidance is more explicit about reader vocabulary. The GPT-5.2 guidance says to avoid jargon unless the conversation clearly establishes an expert reader. That is an expertise rule. In the OpenAI pages examined for this report, no rule distinguishes a canonically defined project term from a term invented inside one task. The current guidance does supply the pieces—audience, clarity, explicit output shape, preserved evidence, and local evaluation—but not this project's vocabulary boundary.

Why a local control remains necessary

Only the project knows which vocabulary is shared

A provider can discourage obscure language globally or adapt to apparent user expertise. It cannot determine from the word alone whether freshness baseline is a registered Commonplace term, whether a code identifier names an important project mechanism, or whether a plausible compound was invented during the current review. That information lives in the project's vocabulary and artifacts.

The Commonplace rule therefore classifies terms by whether their definitions are shared, not by whether they sound technical. It permits domain language and canonically defined project terms. It requires an agent to define, expand, or omit task-local terms and inherited terms that have no canonical definition. This is more precise than "avoid jargon," and it preserves the value of a small, stable technical vocabulary.

This also reflects a human-agent asymmetry already described by minimum viable vocabulary and agent statelessness. Humans retain useful project terms across sessions. Agents normally need those definitions loaded again. A durable vocabulary can lower the human's repeated reading cost while still requiring deliberate routing for each new agent context.

The evidence and the explanation have different jobs

A technical finding must preserve exact identifiers, code paths, conditions, and evidence. Editing it until it reads like ordinary conversation can destroy the basis for review or implementation. An operator message instead needs to make the practical outcome, reachability, consequence, and decision clear.

The local design keeps both. The operator explanation points to the technical finding rather than copying its proof into the opening. This is not merely a verbosity setting: it is a separation between an evidential artifact and a reader-facing view of that artifact.

Provider behavior is a dependency, not the project contract

Model defaults change. The Anthropic guides themselves distinguish Opus 5 from Fable 5, recommend revisiting prompts during migration, and warn that instructions tuned for an older model can compound a newer model's defaults. OpenAI likewise recommends retesting model migrations and keeping local style guidance when it fixes a measured product requirement.

Commonplace needs one communication contract across providers and versions. If a future model satisfies it by default, the instruction should become redundant rather than harmful. If a model-specific prompt conflicts with it, the conflict should be found by evaluation and the obsolete scaffolding removed. Waiting for a provider fix would neither encode the project's vocabulary boundary nor protect it during the next migration.

The local controls

Commit 80a5f3a0 added two paths:

  • The always-loaded operator-communication paragraph tells agents to lead with the practical outcome or decision, preserve important conditions and uncertainty, introduce implementation details after their meaning, and point to technical evidence rather than reproducing it.
  • The operator-brief skill handles a consequential explanation that still depends on several concepts or a long causal chain. It preserves the full finding as the technical basis and writes a separate practical explanation.

Commit 73f530b4 sharpened the vocabulary rule. Project terminology is permitted only when it has a canonical definition. A term that merely circulated through artifacts does not become shared project vocabulary by repetition.

The skill is intentionally not the default for ordinary chat. The short rule covers routine communication; the separate operation pays for a heavier pass only when the finding's complexity warrants it.

Worked check against the initiating example

The initiating code-review finding began with the title:

drifted ignores clearCacheFor, re-enabling the editor on the chapter whose reload just failed

It then led through line references, escalation sites, branch conditions, code identifiers, and comments. Those details matter to a developer validating the finding, but they do not tell an operator what happens without first loading the implementation.

Applying the local rule produces this operator explanation:

If a global search-and-replace finishes while you switch chapters, and the chapter then fails to reload, the editor can show an old copy and still let you type. Your next edit can save that old copy over the completed replacement.

The affected chapter should remain locked until fresh content is available. The current lock decision ignores the list of chapters whose cached content must be discarded, so it can unlock the editor too early.

Technical basis: finding I1 in the full review.

The rewrite keeps ordinary product vocabulary such as chapter, editor, and global search-and-replace. It explains the one implementation concept needed for the cause, and leaves drifted, clearCacheFor, and the line-by-line proof in the technical finding. It also preserves the narrow triggering conditions.

The check exposes a second reason to retain the technical finding. The supplied informal gloss can be read as saying the search-and-replace later undoes the operator's edit. The technical excerpt appears to describe the opposite causal direction: editing the stale chapter can overwrite the already committed search-and-replace result. Simplification must not silently choose between those readings. The evidence remains authoritative when the brief is disputed.

Plausible causes, with evidence limits

Provider documentation establishes intended behavior and observed model differences, but it does not expose a complete causal account of this failure. Several mechanisms remain plausible:

  • Long agent traces make temporary working labels locally frequent, so the model may treat them as shared even when the user never saw them. Anthropic's Fable guidance and user reports support the symptom; they do not by themselves establish the internal mechanism.
  • Instructions can compete or drift as context grows. User reports describe this, and provider migration guides warn that accumulated scaffolding can compound new defaults.
  • Training and automated evaluation may reward superficial signals of quality. The ICLR 2026 paper Flattery, Fluff, and Fog finds preference-model miscalibration across length, structure, jargon, sycophancy, and vagueness. It is evidence for a general selection pressure, not evidence about Anthropic's or OpenAI's private training recipe.
  • Communication style varies by model and setting. Anthropic's study of values across models and languages identifies a depth-versus-brevity axis and says the causes of model shifts are not yet understood. This supports testing each deployment rather than assuming one universal baseline.

The report therefore does not attribute the problem to hidden reasoning leakage, reward-model bias, deliberate token inflation, or any other single cause. Those explanations require evidence not available here.

Evaluation implications

The local fix should be evaluated as three conditions: provider default, provider default plus the AGENTS.md paragraph, and a separate operator-brief pass. Each should be tested after both short and long technical runs and across the models Commonplace actually uses.

The primary question is whether an operator can state, after one reading, what happens, when it happens, why it matters, and what decision or action follows. Checks should also count undefined task-local terms, verify that causal direction and material qualifiers survive, and record rereading time. Token count alone is not a success measure.

The outcomes distinguish likely failure modes:

  • If the small AGENTS.md rule fixes both short and long cases, the gap was mainly local underspecification.
  • If it works initially but decays during long runs, context competition or instruction drift remains.
  • If a separate brief succeeds where the in-context final answer fails, the main problem is likely the collision between working context and reader context.
  • If readability improves while causal or qualifying detail is lost, the two-artifact design is necessary but the brief instruction needs stronger fidelity checks.
  • If results differ substantially by model or harness, the provider layer remains a material variable and local guidance should be calibrated rather than assumed universal.

Until those tests exist, the evidence supports retaining the current controls without expanding them. Provider work makes the controls more plausible—it independently converges on outcome-first writing, selective detail, and re-grounding after agentic work—but does not yet demonstrate that the controls will hold in Commonplace's operating context.