Skip to main content
Prompting the same background into every run gets old, and it drifts. Put it in the workspace instead: agents load these files themselves, so any run in that workspace - from the API, the product, the CLI, or MCP - starts with the same instructions. They are ordinary workspace files, so you create them the same way you create anything else - upload, write through the API, or let an agent write them.

​
Project instructions

CLAUDE.md (or AGENTS.md) at the workspace root is read on every run. Keep it short and durable: conventions, where things live, what never to touch.

​
Skills

A skill is a folder under .skills/ with a SKILL.md inside. The frontmatter is what the agent sees when deciding whether the skill applies, so write the description as “when to use this”, not as a title.
Write it to .skills/monthly-report/SKILL.md:
A workspace loads up to 50 skills. Only the name and description sit in the agent’s context; the body is loaded when the skill is actually used, so a long skill costs nothing until it applies.

​
Subagents

A subagent is a markdown file under .agents/. The main agent delegates a self-contained task to it, and the subagent works in its own context - useful for work that would otherwise flood the main conversation, like reading fifty files to answer one question.
A workspace loads up to 50 subagents.

​
Seeing what a workspace offers

This returns the skills and subagents a run in this workspace would pick up, which is the fastest way to confirm a new file parsed correctly. A skill or subagent whose frontmatter is missing name or description is skipped silently, so check here after adding one.

​
Turning them off for a single run

Skills and subagents apply automatically. To exclude one from a specific run without deleting it:
  • Agent runs - the run options these fields belong to
  • Unattended runs - where durable instructions pay off most
  • Files - writing these files through the API