add-agent

Prompt file imported from anayatkhan1/kit (.cursor/commands/add-agent.md). Copyright stays with the author.

Scope, plan, and ship a new AgentCN agent — context first, SPEC in repo, one commit per todo.

Workflow (follow in order)

Read implementation-workflow.md for the full picture. Summary:

Required reading (before Phase 1)

1. agent-foundation-strategy.md — lanes, 3–5 tools, anti-patterns
2. agent-spec-template.md — SPEC structure
3. build-agent/SKILL.md — file-level checklist per phase

Phase 0 — Context intake

Parse what I provide after /add-agent. If missing, ask:

• Job — one sentence
• Lane — Web | Files | Channels
• Primary provider — API/SDK (e.g. Firecrawl, Telegram)
• Prototype — path (tmp/…), URL, or greenfield
• Must-have tools — what v1 must do
• Env vars — which keys I will use
• Out of scope — what to defer
• Commits — commit each step | commit at end | no commits (default: only when I ask)
• E2E template — wire in examples/agent-ui-template/? yes/no

Do not invent providers or tools I did not specify or that fail the foundation gate.

Phase 1 — Plan (no tool code)

1. Apply foundation strategy (gate, overlap with web-agent / extraction-agent).
2. Create ai/agents/<short>/SPEC.md using agent-spec-template.md.
3. Fill the Implementation plan table with concrete commit messages.
4. Reply with:
   • Link to SPEC path
   • Summary (job, lane, tools, providers)
   • Implementation plan table (todos 0..N)
   • Gate recommendation: proceed | narrow | extend existing agent

Do not write agent.ts, tools, or registry entries until I approve.

If I included implement or go ahead in the same message without approving a plan, still show the plan first unless I explicitly said to skip approval.

Optional todo 0 commit (when I ask):

docs(agent): add <short>-agent spec

Phase 2 — Build (after approval)

Work one SPEC todo at a time:

1. Implement only that todo's files
2. Run verify command from SPEC plan row
3. Report done; if I said commit each step, create one commit with the planned message
4. Move to next todo

Standard order: source → tests → registry → docs+demo → verify → (optional) agent-ui-template.

Match extraction-agent commit style:

• feat(agent): add <short>-agent source for …
• test(agent): add <short>-agent tests
• feat(registry): register <short>-agent
• docs(agent): add <short>-agent docs and demo
• chore(agent): verify <short>-agent build
• feat(examples): wire <short>-agent in agent-ui-template

Update SPEC Status and acceptance checkboxes as you complete work.

Phase 3 — Ship

• All acceptance criteria in SPEC checked
• Offer /pr-description with "implements ai/agents/<short>/SPEC.md"
• Remind: production needs web deploy for npx agentcn add <short>-agent

Do not

• Skip SPEC.md for new registry agents
• Batch source + registry + docs in one commit (unless I explicitly allow)
• Create one agent per vendor demo category
• Port full wizard UIs from tmp/ — tools only
• Ship 8+ v1 tools

Example invocation

/add-agent

Provider: Firecrawl (FIRECRAWL_API_KEY)
Prototype: tmp/firecrawl-migrator — use map + crawl API routes only
Job: map a website and batch-scrape pages to JSON in workspace
Tools: map_site, batch_scrape, save_results
Out of scope: migrator UI, per-industry agents
Commit each step
E2E template: yes