Cognitive Output Refiner

Activation Contract

Use only when the user explicitly asks in English to refine, summarize, compact, deduplicate, or reduce the cognitive load of textual output such as Markdown, console output, logs, test results, or agent responses.

Do not activate automatically just because output is long. If the input is already concise, say no refinement is needed and offer at most a micro-improvement.

Hard Rules

• Preserve the input language. Do not translate unless asked.
• Redaction takes precedence over all exact-preservation requirements, including this output template. Never include credential material, including secrets, credentials, inline credentials, connection strings, URL-embedded credentials, API keys, access tokens, cookies, passwords, private keys, session identifiers, or authorization headers, in whole or in part; replace the full credential value with a placeholder such as [REDACTED_TOKEN].
• Preserve useful diagnostic detail from errors, relevant stack frames, commands, paths, public/non-sensitive IDs, decisions, trade-offs, critical warnings, verification steps, and actionable next steps. When preserving stack frames, commands, or paths, redact usernames, local/private path prefixes, internal hostnames, credentials, and other sensitive identifiers while keeping non-sensitive filenames, function names, line numbers, error codes, and public IDs.
• When an ID's sensitivity is ambiguous, redact it. Preserve only clearly public identifiers.
• Be faithful by default. Label inferences as Inference, Likely cause, or Possible next step; never present them as input facts.
• Deduplicate by keeping one canonical version. Preserve meaningful differences explicitly.
• Target 30–50% of the original size by default; use 10–20% only for explicit ultra-compact requests. Fidelity beats size.
• Never overwrite source content. Produce a separate response or file.