The Writer Agent — Prose That Fits the Audience
How Datarim's Writer agent produces technical documentation, blog posts, and social media content — matching format, register, and length to the target platform.
A pull request needs a changelog entry. A new feature needs a README section. A finished sprint needs a blog post. Each of those asks a different kind of writing, and a single generic prompt rarely produces the right answer for all three. The Writer agent handles the full range — from API docs to a Telegram post — by treating audience and platform as first-class inputs.
The agent runs at /dr-write and also participates in /dr-archive, where it produces the final documentation review as part of the reflection step.
What it produces
On the technical side: README files, API reference docs (OpenAPI, JSDoc, docstrings), architecture decision records, changelogs, migration guides, and troubleshooting docs. On the content side: research-backed articles, how-to guides, and platform-appropriate social posts for Telegram, LinkedIn, Twitter/X, and Facebook. Multi-language output — English and Russian — is part of the standard capability, not a special mode.
Each output type has its own constraints. A Telegram text post has a hard limit of roughly 4 000 Russian characters before the platform clips it. An API docstring needs to stay close to the code it describes. The Writer applies those constraints during drafting, not after.
One concrete example
Take a new authentication endpoint. The Writer produces three things from the same source: a JSDoc block that describes parameters and return values for the developer reading the code; a migration guide section explaining what changed for teams upgrading from the previous version; and a short Telegram announcement written for a general technical audience. Three audiences, three registers, one task.
Writing principles enforced
The agent applies a set of rules that push against common AI writing patterns. One idea per paragraph. Concrete over abstract — name the specific technology, cite the actual number. Active voice where possible. No filler phrases. For social content, there is an additional rule: links go into a first comment, not into the post body, which is a platform convention on Facebook, LinkedIn, and Telegram that keeps the body readable.
Where it fits in the pipeline
The Writer runs on a Sonnet-class model — structured enough for disciplined output, fast enough for iterative drafts. When a draft needs fact-checking, the agent loads the factcheck skill. When output risks sounding mechanical, the humanize skill provides a reference for the specific patterns to avoid.
Full capability list on the Writer reference card, or read what Datarim is for the bigger picture.