| 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 |
Deliver Phase
Present the completed agent to the user. This skill requires manual invocation because delivery should only happen after validation passes.
Tier awareness:
- Tier 1: Skip README.md. Generate only MAINTENANCE.md (abbreviated — just the observability signals table and mistake-handling workflow). Keep the delivery presentation conversational and brief.
- Tier 2+: Full delivery as described below.
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.
Delivery Presentation
Present to the user:
1. Project Tree
Show the full file structure of the generated agent.
2. How to Use It
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.
3. Example Prompts (3-5)
One per skill, using real-world phrasing the user would actually type. Not formal — conversational.
4. Eval Results Summary
Which test cases were run, what passed, scores on dimensions.
5. Known Limitations
What the agent can't do well yet. Be specific and honest.
6. Maintenance Expectations
What needs periodic review and when (link to MAINTENANCE.md).
Generate 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:
When the Agent Makes a Mistake
- Fix the immediate output
- Ask: recurring pattern or one-off?
- Recurring → add rule to CLAUDE.md or skill (only if decision boundary, not edge case)
- Edge case → add to evals/test-cases.md as regression test
- If adding a rule exceeds line budget → prune a less-important rule first
Observability Signals
| 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 |
Review Cadence
| 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 |
Pruning Rules
- If Claude consistently follows a rule without it being in CLAUDE.md → remove it
- If a rule was added for a one-off and hasn't triggered → remove it
- If two rules contradict → resolve, keep one, document why
- If CLAUDE.md exceeds budget → move content to skill or template
Generate README.md
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.
Offer a Dock Launcher (Standalone + macOS only)
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.
After Delivery
Tell the user based on deployment target:
Standalone:
- The agent directory is self-contained and can be moved anywhere
cd [path] && claude to start using it- On macOS, a Dock launcher can be generated (see "Offer a Dock Launcher" above)
- Review MAINTENANCE.md for ongoing care
Project subagent:
- The agent is now active in the project — restart Claude or use
/agents to verify
- Claude will auto-delegate relevant tasks, or you can request the agent explicitly
- Subagent skills are namespaced with
[agent-name]- to avoid conflicts
- Review MAINTENANCE.md for ongoing care
Multi-project subagent:
- Run
deploy.sh to create symlinks in all target projects
- The source of truth is
generated-agents/[agent-name]/ — edit there, then re-run deploy.sh to update target projects
- To add a new target project: add its path to
deploy.sh and re-run
- To remove from a project: delete the symlinks in that project's
.claude/ directory
- Review MAINTENANCE.md for ongoing care
User-level:
- The agent is now available in all projects
- Restart Claude or use
/agents to verify it appears
- Review MAINTENANCE.md for ongoing care
For all targets: the agent will improve with use — add regression tests when it makes mistakes.
