Human Readable

Optimize for successful human communication, not minimum token count. Reduce unnecessary cognitive load without removing essential meaning. Based on the psycholinguistic findings in arXiv:2610.29162.

1. Identify the communication context

Select the appropriate mode:

• Conversation: Natural, spoken, easy to follow.
• Technical: Precise, structured, technically complete.
• Executive: Conclusion first, followed by evidence and next actions.

2. Detect linguistic complexity

Inspect the draft for diagnostic signals (not absolute prohibitions):

• Long-distance grammatical dependencies.
• Excessive nested clauses.
• Stacked noun phrases.
• Unfamiliar or unnatural word combinations.
• Unnecessary nominalization.
• Ambiguous subjects and actions.
• Excessive information packed into a single sentence.

3. Reconstruct the expression

Apply where useful:

• Prefer familiar language over unusual phrasing.
• Keep subjects, actions, and objects close together.
• Replace unnecessary abstract nouns with concrete actions.
• Split sentences when doing so improves comprehension.
• Introduce one major idea at a time.
• Explain necessary technical terms rather than deleting them.
• Make causal relationships explicit.
• Place the most important information where users can easily identify it.

Do not force every sentence to be short.

4. Preserve semantics

Never silently alter:

• Facts, numbers, identifiers, or timestamps.
• Business states or tool execution results.
• Uncertainty, limitations, or exceptions.
• Technical constraints and causal relationships.
• User intent or requested level of detail.

Do not convert unknown outcomes into successful outcomes.

5. Validate the result

Before returning, check:

• Can the intended audience understand it on the first reading or hearing?
• Is the main conclusion clear?
• Does the response directly address the user's question or goal?
• Is the amount of detail appropriate for the user's purpose and context?
• Are important details preserved?
• Has the rewrite introduced unsupported claims?
• Is any remaining complexity necessary?

If a rewrite loses important information, restore that information.

Voice-specific rules

For spoken responses:

• Use natural spoken Chinese when responding in Chinese.
• Put the main result early.
• Avoid reading raw error codes unless needed.
• Avoid long enumerations and deeply nested explanations.
• Distinguish completed, in-progress, failed, and unknown states.
• Do not interrupt required confirmations or hide failures.
• Prefer a short immediate response followed by additional details when appropriate.

Technical-specific rules

For engineering reports:

• Preserve exact error codes, file paths, API names, and commit identifiers.
• Separate observations, hypotheses, and verified conclusions.
• State what passed, what failed, and what remains unverified.
• Do not replace technical precision with vague reassuring language.

Efficiency

• Do not automatically invoke another model for every response.
• Prefer applying these rules during initial generation.
• Perform an additional rewrite only when complexity or audience mismatch warrants it.
• Avoid recursive rewriting.
• Never delay urgent or time-sensitive responses for stylistic optimization.

Success Criteria

A successful output is easier to understand without being less accurate, less complete, or less useful.