The coworker-context Skill — Conventions for External LLMs Editing Datarim Artifacts
How Datarim's coworker-context skill gives an external LLM the exact conventions it needs to generate or edit task artifacts correctly — frontmatter, stage headers, snapshot fields, and more.
Datarim delegates bulk I/O work to an external LLM through the coworker tool — first drafts of plans, long research documents, archive write-ups. The external model does not share the conversation history that the primary agent accumulated. Without explicit conventions, it makes plausible guesses: it reorders YAML keys, invents field names, or omits a required frontmatter field. The coworker-context skill is the single reference document an external LLM reads before touching any artifact under datarim/.
What it does
The skill contains eleven sections. The most frequently applied ones cover YAML frontmatter preservation, stage headers, and snapshot fields.
Frontmatter preservation means: never reorder keys, never change quoting style, never alter indentation, and never modify a key name. If a value must change, only the value characters are edited. This rule exists because Datarim scripts parse frontmatter by exact key order and spacing — a reordered file is a broken file from the tool's perspective.
The stage header convention requires that every operator-facing response from a pipeline command begins with a bold inline title in the form **TASK-ID · title** as the literal first line. The title comes verbatim from the task file; the separator is a Unicode middle-dot, not a hyphen.
The snapshot frontmatter section defines ten mandatory scalar fields that every stage-snapshot artifact must carry: task ID, artifact type, schema version, stage, command, capture timestamp, captured-by, recommended next command, size in bytes, and a truncated flag. The skill specifies the exact field names and permitted values so the external LLM does not invent alternatives.
Other sections cover the PRD-to-archive mirror (each archive validation item cites the corresponding PRD acceptance criterion), the documentation taxonomy (four Diátaxis categories, no others), and the security baseline (nine named areas, S1 through S9). Each section is history-agnostic — it names contract surfaces, not specific task identifiers.
One concrete example
An external LLM is asked to draft a compliance report. Without this skill, it might write a custom frontmatter structure with fields it considers logical. With coworker-context loaded, it knows the report frontmatter must carry task_id, date, and verdict, and the verdict must be one of three exact string values. The output requires only a review pass, not a structural rewrite.
When it loads
The skill is referenced by the datarim profile in coworker's profile configuration. Any call to coworker write --profile datarim or coworker ask --profile datarim injects this skill's contents as the system prompt prefix. It is also loaded during /dr-write and /dr-archive when external delegation is involved.
Read about what Datarim is for the full pipeline context, or see the /dr-archive command where coworker delegation most frequently applies.