A reusable Claude Code project that runs end-to-end SEO ops via the CitationBench MCP server — tell Claude what you want, it composes tools, persists results, and pauses for approval.

A reusable Claude Code project that runs end-to-end SEO ops via the CitationBench MCP server. Tell Claude what you want; it composes tools, persists results in your workspace, and pauses for approval where it matters.


Outcome	A `.claude/` setup that gives Claude Code a sharp SEO agent persona scoped to your workspace
Time	~15 min setup
Cost	Varies by what you ask the agent to do; demo mode is free
Prereqs	Claude Code, CitationBench API key, a workspace

What it does

You ask Claude in plain language:
  "Bootstrap acme.com and queue 10 blog posts for the top problem-solution keywords."

Claude:
  agent.invoke(skill: bootstrap_brand)
    ↓ waits, narrates progress
  research.keyword.search (intent: PROBLEM_SOLUTION, missingFromContent, maxKd: 35, limit: 10)
    ↓
  produce.blog_post.bulk
    ↓
  reports back with credit usage and links to drafts

Step 1 — Connect the MCP server

claude mcp add citationbench https://mcp.citationbench.com/mcp \
  --header "Authorization: Bearer $CITATIONBENCH_API_KEY" \
  --header "X-Workspace-Id: $WORKSPACE_ID"

Verify with claude mcp list. (Or, to start without an account, omit the headers for demo mode.)

Step 2 — Drop in a CLAUDE.md

Create CLAUDE.md in your project root. This is the agent persona.

# SEO ops agent for Acme

You're our SEO and GEO ops agent. You have CitationBench MCP tools.

## Workspace

- Brand: Acme PM (project management for engineering teams)
- Workspace ID: ws_acme
- Primary content URL: https://acme.com/blog/engineering-team-capacity-tracking

## Defaults

- Always check workspace context (research.keyword.search filters) before producing content
- Prefer `with-research` mode for blog posts
- Always apply rfn_brand-voice + rfn_seo-cleanup refiners
- Pause for approval on: publish, outreach send

## Cadence

- Mondays: weekly rank check, surface drops of 5+ positions
- Daily: AI citation snapshot across our top 20 prompts
- Continuous: respond to inbound link-building requests via the inbound queue

The MCP tools light up. Claude reads CLAUDE.md to set its working context.

Step 3 — Build skills (reusable Claude Code commands)

Drop these in .claude/skills/ as separate markdown files.

`.claude/skills/bootstrap.md`

# Bootstrap a new client

When asked to bootstrap a brand:

1. Call agent.invoke(skill: bootstrap_brand, input: { domain: <user-provided>, depth: "thorough" })
2. Stream events until SUCCEEDED
3. Surface the result.workspaceId, the keyword count by intent, the top 3 competitors
4. Ask: should I queue landing pages from the 6 briefs?

`.claude/skills/weekly-ops.md`

# Weekly ops run

Every Monday:

1. workspaces.bulk_action with action=agent.invoke, skill=rank_monitor, alertOn={drop:5}
2. For each drop:
   - Fetch the keyword + content
   - Call produce.evaluate
   - If score < 70: spawn refresh_stale at WAITING_APPROVAL
3. Summarize: total checks, drops, queued refreshes

`.claude/skills/inbound-triage.md`

# Triage inbound link requests

Each morning:

1. link_building.inbound.pending — get the queue
2. For each message, render: from, subject, classification, escalation reason, draft preview
3. Ask: approve / approve-with-edits / reject / mark spam — one at a time

Step 4 — Try it

claude

> Run the weekly-ops skill

Claude reads .claude/skills/weekly-ops.md, runs it. You'll see each tool call inline.

> Bootstrap newclient.com — depth thorough

Claude runs the bootstrap skill. After ~20 minutes you have a fully populated workspace.

> Triage today's inbound

You walk through the queue interactively.

Commit CLAUDE.md + .claude/skills/ to your repo. Every team member who runs Claude Code in this repo gets the same SEO agent persona + skills.

Cross-workspace patterns

For agencies, the same setup pattern works — but Claude calls workspaces.bulk_action and references workspaces by tag:

# in CLAUDE.md

- Agency: yes
- Workspaces tagged "tier:premium" should be auto-refreshed on rank drops
- Workspaces tagged "tier:standard" should only alert, not auto-refresh

Claude reads the tags and behaves differently per client.

Gotchas

Always include workspace context in CLAUDE.md. Without it, Claude doesn't know which workspace to act on and either asks every time or guesses.
Skills are markdown, not code. They're free-form instructions; Claude follows them flexibly. Don't over-engineer.
Demo mode is great for development. Build your skills and CLAUDE.md against demo mode first; swap in the live key when ready.
Approval gates still apply server-side. Even if your skill says "publish," the workspace's approval policy can hold the invocation at WAITING_APPROVAL.

API: Agent · invoke
API: Inventory
MCP: Catalog
MCP: Setup — Claude Code
Concept: Agent
Playbook: Keyword research for a brand in 20 minutes

Build an MCP-powered SEO agent in Claude Code