Artifact Templates
Workflows don't just do work — they produce documents: a plan before the code, a result after it, a learning for the next time. Artifact templates define what those documents look like. A template is a named Markdown form; an agent fetches it by name, fills it in, and saves the finished artifact attached to the task it was working on.
The built-in three
| Template | Written when | What it captures |
|---|---|---|
| Task Plan | Before implementation | Objective, scope, ordered steps, risks, acceptance criteria, validation strategy |
| Task Result | When implementation completes | What was delivered, plan deviations, validation results, commit references |
| Task Learning | When implementation completes | What went well, issues and their fixes, which reusable patterns emerged |
The built-in workflows reference these by name — so changing what a "plan" must contain in your workspace is just editing the Task Plan template. Every future plan follows the new form; nothing to redeploy.
An artifact can be anything — add your own
Templates aren't limited to the built-in three. Any document your process produces can be a template: a design note, an incident post-mortem, a release checklist, a QA report, a decision record. Create one in the Templates section of the AI Workflows page — give it a name, a description (when an agent should write this artifact), and the Markdown body to fill in — then reference it from a workflow step ("fill in the Design Note template and save it").
A few properties that keep this low-maintenance:
- Agents save and read artifacts by task key + template name — storage, identifiers, and task linkage are handled by the platform. Saving again simply updates the same artifact.
- Renaming a template never breaks anything already saved.
- The template editor's Agent tab shows the exact form a connected agent receives.
Where artifacts end up
Artifacts are saved in the knowledge base, attached to their task — the plan an agent wrote is there to review before you approve implementation, the result to audit after it, the learning for the next time. Together with reusable learnings (see Context), they're how each task leaves the workspace a little smarter than it found it.