菜单
Present the completed agent to the user with usage instructions, example prompts, eval results, known limitations, and maintenance model. Creates MAINTENANCE.md and README.md.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/leonbeckert/agentspawn --skill deliver命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Write the agent files based on the approved design brief. Creates agent artifacts in the location specified by the deployment target. Supports standalone, project subagent, and user-level deployment.
Run the eval suite against the built agent and report results. Checks hard gates and scored dimensions against release criteria. Use after building agent files.
Classify an agent by type and complexity tier. Use at the start of any agent build request to determine how to approach the project.
基于 SOC 职业分类
正在显示 SKILL.md
| name | deliver |
| description | Present the completed agent to the user with usage instructions, example prompts, eval results, known limitations, and maintenance model. Creates MAINTENANCE.md and README.md. |
| disable-model-invocation | true |
Present the completed agent to the user. This skill requires manual invocation because delivery should only happen after validation passes.
Tier awareness:
Before starting: Verify that generated-agents/[agent-name]/evals/test-cases.md exists and contains passing results. If not, tell the user to run /validate first.
Present to the user:
Show the full file structure of the generated agent.
Adapt to the deployment target:
Standalone:
cd generated-agents/[agent-name] && claude
Project subagent:
cd [target-project] && claude
# Then: "Use the [agent-name] agent to [task]"
# Or Claude will auto-delegate when relevant
Multi-project subagent:
cd generated-agents/[agent-name] && bash deploy.sh
# Then in any target project:
cd [target-project] && claude
# "Use the [agent-name] agent to [task]"
User-level:
claude # from any project
# Then: "Use the [agent-name] agent to [task]"
Explain the workflow: what to say, how skills are invoked, how delegation works for subagents.
One per skill, using real-world phrasing the user would actually type. Not formal — conversational.
Which test cases were run, what passed, scores on dimensions.
What the agent can't do well yet. Be specific and honest.
What needs periodic review and when (link to MAINTENANCE.md).
Write generated-agents/[agent-name]/MAINTENANCE.md using templates/maintenance.md (agentspawn repo root, not this skill's directory) as a starting point. Customize the review cadence and observability signals to this specific agent.
Must include:
| Symptom | Check First | Then Check |
|---|---|---|
| Asks questions answered in files | CLAUDE.md context loading, file paths | Skill input spec |
| Hallucinated facts | guide.md freshness, source rules | Skill template, eval coverage |
| Wrong tone or voice | Skill examples, principles template | CLAUDE.md output defaults |
| False refusal | Forbidden outputs (too broad?), scope | Ambiguity strategy |
| Uses stale data | [Review: date] markers | templates/current-facts.md |
| Overclaims | Forbidden claims list, evidence rules | Skill template |
| Conflicting behavior | CLAUDE.md contradictions | Source-of-truth precedence |
| What | Frequency | Action |
|---|---|---|
| current-facts.md | Quarterly | Update pricing, offerings, team |
| current-campaign.md | Monthly | Update or delete time-limited content |
| guide.md / docs/ | Biannually | Stale stats, dead links, outdated frameworks |
| skills/*.md | When outputs drift | Update templates, add bad examples |
| evals/test-cases.md | After significant failure | Add regression test |
| CLAUDE.md | Quarterly | Prune rules Claude follows without instruction |
Write a brief generated-agents/[agent-name]/README.md:
# [Agent Name]
[1-2 sentence description]
## Quick Start
cd [agent-name] && claude
## Skills
- [skill-name]: [what it does]
- ...
## Maintenance
See MAINTENANCE.md for review cadence and troubleshooting.
For standalone agents on macOS, offer a one-click Dock launcher so the user
can start the agent's session without typing. Ask whether they want one, and
which command it should run (claude, or their yolo/skip-permissions alias).
When they say yes, run the generator from the agentspawn repo root:
scripts/make-launcher.sh --cmd "<command>" "<AppName>" "<absolute-agent-dir>"
<AppName> — a Title-Case name (e.g. the agent's display name). This is what
shows in the Dock.<absolute-agent-dir> — the full path to generated-agents/[agent-name]/.--cmd — what to run after cd. Omit it to let the script prompt; pass
claude or yolo directly when the user already told you.The script auto-detects the terminal (iTerm2 → Ghostty → Terminal.app),
generates a distinct colored icon from the agent's initials, and writes the
.app to ~/Applications. Tell the user to drag it onto their Dock and to
click "OK" on the first-run automation prompt.
Skip this on non-macOS, for non-standalone targets, or if the user declines.
Tell the user based on deployment target:
Standalone:
cd [path] && claude to start using itProject subagent:
/agents to verify[agent-name]- to avoid conflictsMulti-project subagent:
deploy.sh to create symlinks in all target projectsgenerated-agents/[agent-name]/ — edit there, then re-run deploy.sh to update target projectsdeploy.sh and re-run.claude/ directoryUser-level:
/agents to verify it appearsFor all targets: the agent will improve with use — add regression tests when it makes mistakes.