creator

AgentUse Creator

Use this skill when creating or improving .agentuse files: markdown with YAML frontmatter and plain-English instructions. The filename is the agent name.

Basic Agent Shape

---
model: anthropic:claude-sonnet-5
description: "Short action-oriented purpose"
---

You are a focused autonomous agent.

## Task
Describe exactly what the agent should accomplish.

## Output
Describe where results should go and what format they should use.

Common Frontmatter

model: anthropic:claude-sonnet-5
description: "Analyze daily metrics and send a concise summary"
timeout: 600
maxSteps: 100
schedule: "0 9 * * *"
subagents:
  - path: ./researcher.agentuse
    name: research
    maxSteps: 50
mcpServers:
  filesystem:
    command: npx
    args: ["-y", "@modelcontextprotocol/server-filesystem", "."]

Authoring Checklist

• A concrete job the agent can finish without interactive supervision.
• model: set explicitly; short description: if it may be listed or used as a subagent.
• Tools and MCP servers declared in frontmatter, not assumed ambient.
• Inputs, outputs, destinations, and success criteria stated in the body.
• Multi-role work: subagents with clear names and maxSteps limits.
• Recurring work: YAML schedule: + a note that agentuse serve must run.

Write Lean: Hard-Code Invariants, Delegate Judgment

The prompt is a brief, not a manual. It is re-sent on every step, so length is a recurring token cost and a long prompt buries the rules that matter. Pin down only what must be exact:

• safety boundaries (read-only, never call X, which store to write),
• exact commands, paths, and flags the model cannot guess,
• the output schema and where it goes,
• ordering that changes the result.

For everything else — how to investigate, how to phrase, the long tail of edge cases — state the goal and the constraint, then let the model decide. Spelling out every branch makes the agent brittle on the case you did not enumerate.

Over-specification smells: the same rule in three places, a paragraph justifying why a step exists, an enumerated decision tree derivable from one sentence of intent. Write what a competent teammate needs, not a spec.

Skills are instructions, not tool grants: declare an agent's tools in frontmatter even when a skill documents them with allowed-tools. Put reusable instructions in .agentuse/skills/<name>/SKILL.md or install with agentuse add.

Source Precedence: Skills Are Defaults, Learnings Override Them

The runtime composes one prompt from layered sources, in this precedence (highest first): agent instructions → Learned Guidelines → Skills → other reference files. The system prompt's operational/safety rules sit above all of these. This shapes where a rule belongs:

• Put soft defaults in skills. Don't bake a hard "never do X" into a skill that a learning should be able to override — a captured correction outranks a skill default, so an absolute skill rule fights the feedback loop.
• State a rule once, at the right layer, and reference it. The same craft rule copied into both a skill and the agent drifts; the lower-precedence copy then silently wins (this is the "same rule in three places" smell above, seen from the runtime side).
• learning: true (sugar for capture + apply) injects the agent's stored learnings every run — for delegated subagents too, not just top-level runs. So a leaf's prior-run corrections actually reach it; rely on that instead of hand-restating past corrections in the prompt.

Gotchas

• No builtin grep/glob. Only filesystem_read|write|edit exist, and filesystem_read returns the whole file unless you pass limit/offset. For big or structured files, grant grep/rg/jq via bash and tell the agent to search, not slurp. A read agent given only cat/ls falls back to whole-file reads — the exact context blowup to avoid.

• Approval gates are async. approval: true / await_human gates suspend the run; timeout: does not tick during the wait. Size timeout: for the active work between gates, not human response time.

• Agents cannot prompt the user mid-run. Never write "stop and ask the user to do X." The only branches at a blocker: exit with a clear error, record to the store and continue/stop, or fire an approval/notification. The approval gate is the only human-in-the-loop path.

• Validate models against agentuse models. The catalog moves; check it before calling a name invalid. Don't infer limits from other providers' naming (e.g. "5.5 can't exist because provider Y stops at 5.2").

• Defer to skills; don't inline their internals. Reference a skill by name (/linkedin) and never copy its drift-prone internals (script paths, eval invocations, file layout). Do repeat the durable steps/context the run hinges on inline, since the skill may not load that turn. Rule: reference what changes, repeat what doesn't.

• Skill scripts read via bash need explicit allowlists. Grant tools.filesystem read on the skill dir (absolute path — ~ may not expand) and a narrow bash.commands like cat /Users/<you>/.claude/skills/<name>/scripts/*. Not a blanket cat *.

• Match channel prose to the frontmatter key. channels.slack is the key. Don't describe Slack delivery as notifications.routes; the body must name the key the frontmatter uses.

References

Don't hardcode a doc list here — it goes stale. Fetch the canonical index and load the specific page(s) you need:

https://docs.agentuse.io/llms.txt

It enumerates every current guide and reference page with a short description. Pick the matching .md URLs and fetch them directly.