| name | create-agents |
| description | Use when creating or upgrading reusable agent manifests or agent teams with explicit model profile metadata and harness-style contracts. |
Create Agents
Use this skill to author reusable agent manifests for templates/.agents/agents/ or another shared agent directory.
The job is to make a specialist agent loadable, narrow, and honest about its scope, uncertainty handling, and fallback behavior.
Core Contract
- Solve one repeatable agent job.
- Keep the manifest portable: discovery metadata up top, execution contract in the body.
- Keep the agent role narrow enough that another agent can tell when to use it.
- Use harness-style language: file-backed truth, explicit uncertainty, and no silent fallback.
- Use
meta-prompting as the prompt-architecture lens when shaping the agent body.
- If the request is really a skill package (leaf or router), use
create-skill instead.
Role Set
For coding work, the default reusable agent set is:
task for bounded implementation or patch work
explore for evidence gathering and codebase discovery
plan for scoping, sequencing, and risk framing
librarian for reusable knowledge, inventories, and convention lookup
Only expand the set when the request proves another specialist is necessary.
Workflow
- Define the agent's job, scope boundary, and primary evidence surface.
- Choose the narrowest role and output contract.
- Write frontmatter with
name and description only.
- Write the body with role, core contract, uncertainty labels, fallback behavior, and output shape.
- If the agent is part of a team, keep shared conventions in
references/ and role-specific behavior in the individual manifest.
- Add direct evals that test discovery, a near-miss boundary, and a noisy case.
Uncertainty Protocol
- Label observations as
OBSERVED, INFERRED, or UNKNOWN.
- Do not present an inference as a measured fact.
- If a decision depends on missing information, say what is missing and whether a safe default exists.
- If the missing detail changes shared token, contract, or routing behavior, ask before inventing the answer.
References
Evaluation
- Add at least three prompts to
evals/evals.json.
- Add
evals/trigger-evals.json when discovery precision matters.
- Compare against a believable baseline, such as the old persona-style manifest or no agent package.
Final Checklist