zhangyanbo/opus-writer/opus-writing

Opus 4.6 Writing Agent

Opus 4.6 has the best prose among all Claude models. This skill delegates writing tasks to it.

Collaboration Principles

The current model owns the content; Opus 4.6 owns the prose.

Too much intervention suppresses Opus's writing quality — if every sentence is already polished before handoff, all it can do is swap in synonyms, wasting its greatest strength. Too little intervention leaves Opus without direction — it may fabricate facts or drift off course. The key is to give Opus enough information to get the content right while leaving enough room for it to write freely.

The current model prepares the content — a complete draft, facts plus an outline, text that needs polishing, any form will do — then hands it to Opus and lets it write in its own voice. Word choice, sentence structure, paragraph organization, transitions — all of these are Opus's decisions. The current model does not pick words for it, does not provide example sentences for it to imitate, and does not tweak its style after it finishes.

If there are project-specific writing standards or constraints, pass them to Opus at call time.

The only situation in which the current model should directly modify Opus's output is a factual error or a violation of an explicit constraint. Changes must be minimal — fix only what is wrong. For style issues, do not edit — tell Opus and let it revise.

Execution

The opus-writer subagent must be called through this skill, not invoked directly by the user — without this skill the current model does not know how to collaborate with it.

In Claude Code, prefer the subagent (subagent_type: opus-writer).

When the subagent is unavailable, call via the command line:

claude --agent opus-writer -p "<content or instructions>" --output-format json

When neither option is available, prepare the materials and inform the user to hand them to Opus 4.6 manually.

Post-Completion Review

After Opus finishes, the current model fact-checks the output and corrects any errors. Fix only factual errors; for style issues, let Opus revise.

When Not to Use

• Changing a word or two, fixing a cross-reference — the current model handles it directly.
• The task is reviewing, auditing, or finding errors rather than writing — the current model does it.
• The user explicitly asks the current model to write.