Claude Code for SEO: CLI and Agent Workflows That Beat Chat-Only

Claude Code for SEO is the power path - terminal agent + files + optional MCP - not “paste a URL into chat and hope.” Open a site or content repo, load project instructions (CLAUDE.md / AGENTS.md), run fixed audit/brief/ops loops, graduate stable work into skills or agents, and keep humans on claims and publish. SEO jobs without Code → Claude for SEO. Skills catalog/use/build → Claude SEO skills. Install order → What Claude SEO skills to install first. Project instructions → How should I share AGENTS.md and CLAUDE.md?.

When to use Claude Code for SEO / when not

Use Claude Code when the job needs the filesystem: content repos, templates, bulk markdown, multi-step agents, shell checks, or live data via one MCP pipe. Weekly audit/brief/ops loops that save outputs in-repo are the sweet spot.

Don’t use Claude Code for one-off questions with no repo, strategy/taste calls, empty context (no crawl/GSC/SERP notes), or day-zero “install 40 SEO plugins.” A third-party README is not your SEO strategy. If the CLI feels like the blocker - or the repo is empty - stay on chat/Projects via Claude for SEO until you have files and a named job. Skills catalog seekers → Claude SEO skills. Install seekers → install-first or find-install - do not expect those tables or command blocks here.

Power-path map - when Code beats chat

Choose by job friction, not marketplace fame. If the work must touch many files, write beside source, or chain a verified data pipe, Code wins. If you only need a one-shot rewrite with pasted notes, chat wins. Jobs detail stays on Claude for SEO; skill types on Claude SEO skills. Do not invent marketplace plugin titles - if you name a third-party SEO Code plugin, verify same-day and deep-link find-install.

Workflow - Claude Code SEO in 8 steps

1. Pick a Code-worthy job - Audit triage, in-repo brief, bulk ops QA. Else stay on Claude for SEO.
2. Open or create the site/content repo - Add an audit-output/ or briefs/ folder convention so outputs land where humans can review.
3. Load project instructions - CLAUDE.md and/or AGENTS.md with brand rules, must-not claims, and folder conventions. Sharing patterns: AGENTS.md / CLAUDE.md Spec (do not restate the full recipe here). Claude Code CHANGELOG head still 2.1.278 as of 2026-09-20; AGENTS.md landed in 2.1.277 (memory docs, Checked 2026-09-20).
4. Gather inputs or one MCP pipe - Brand pack, SERP notes, crawl/GSC exports - or connect one verified MCP data pipe. Missing critical inputs → mark [GAP]; do not invent fills.
5. Run a repo audit or in-repo brief - Use the prompt cards below. Save markdown outputs in-repo.
6. Human QA - No invented volumes/rankings. Approve before any patch/PR. Customer-facing copy stays human-owned.
7. Graduate stable loops - Boring weekly? Thin SKILL.md or agent instruction. Catalog/use/build → Claude SEO skills; install order → install-first.
8. Prune - skill-doctor + anti-bloat; avoid overlapping audit vs brief vs schema triggers.

Prompts (≥3 copy-paste)

Three Code-native specs. Full mirror: the companion prompts file. Prompt discipline pattern: Claude prompts for marketing.

1. Repo SEO audit runner

Role: SEO analyst in Claude Code. Inputs: repo root · goals · crawl/GSC notes (or [GAP]). Output: CRITICAL/HIGH/MEDIUM table with evidence paths/files. When not: inventing crawl data or Labs volumes.

You are an SEO analyst running inside Claude Code against a real repo. Triage from PROVIDED evidence and readable files only.

REQUIRED INPUTS: Repo root path; Business/SEO goal (one sentence); Crawl or GSC notes (paste or file path  -  or "[GAP]"); Scope (paths/URLs in-repo); Constraints (CMS limits, no-dev, timeline  -  or "none").

STEPS: (1) Restate the job and goal in one line. (2) Read only in-scope files/notes  -  no invented crawl findings. (3) Table: severity (CRITICAL|HIGH|MEDIUM) | issue | evidence (file path, URL, or quote) | likely impact on goal | owner (content|dev|ops) | how would we know this failed? (4) Top 5 actions; save proposed notes under audit-output/ as draft markdown  -  do not open a PR unless asked. (5) Mark [GAP] where evidence is thin.

RULES: Never invent search_volume, KD, rankings, or index status. Never invent marketplace skill/plugin names. No autopilot patches. No full migration plan unless asked.

OUTPUT: severity table | top 5 actions | [GAP] list | QA self-check (evidence-backed only) | suggested next CLI/action for the human.

2. In-repo page brief / content ops

Role: SEO content strategist. Inputs: primary KW · SERP notes · target path in repo · ICP. Output: brief + outline + must-not + internal links to existing repo files only.

You are an SEO content strategist writing a page brief that must land in a content repo.

REQUIRED INPUTS: Primary keyword; SERP notes (paste); ICP / audience; Target path in repo (e.g. content/guides/foo.md or "new under briefs/"); Brand voice / must-not; Existing sibling files or link candidates (list paths  -  or "none").

STEPS: (1) Intent line from SERP notes only. (2) Three H1 options (on-brand, not clickbait). (3) Outline H2/H3 with must-cover entities. (4) Must-cover vs must-not. (5) Internal links only to existing listed repo paths  -  no invented URLs. (6) Write the brief markdown to the target path (or briefs/) labeled draft. (7) Open questions as [GAP].

RULES: No promised rankings or traffic lifts. No invented competitor metrics. No marketplace skill dump. Do not expand into a full site strategy.

OUTPUT: intent | H1 options | outline | must-cover / must-not | link targets | file path written | [GAP] questions.

3. Code / MCP readiness check

Role: SEO ops lead. Inputs: weekly job list · available exports/MCPs. Output: Code vs chat matrix; MCP vs CSV; [GAP] before plugin shopping.

You are an SEO ops lead. Decide which jobs need Claude Code vs chat/Projects, and which need live MCP vs file exports.

REQUIRED INPUTS: Weekly SEO job list (name + frequency); Available data (CSV/export paths, MCP names you already have, or "none"); Repo exists? (yes/no/path); Project instructions present? (CLAUDE.md|AGENTS.md|neither).

STEPS: (1) For each job: Prefer (chat|Projects|Claude Code) + Why + Data pipe (CSV/export|MCP|none|[GAP]). (2) Cap "add MCP" recommendations at one until that pipe is verified. (3) Flag jobs that should stay on Claude for SEO (chat/Projects)  -  do not rebuild that playbook. (4) Never invent marketplace plugin titles or Labs volumes. (5) If user asks install order → install-first companion; GitHub install → find-install companion; skills catalog → Claude SEO skills.

RULES: Empty repo + no exports → defer Code. No "install everything SEO" advice. Mark [GAP] before shopping.

OUTPUT: decision table | one next action | [GAP] list | QA self-check.

Agent block - Claude Code SEO power-path operator

Paste as a Claude Code agent preamble (or Project instructions when the same rules apply). Routes chat jobs and install questions out; keeps Code work evidence-backed.

You are the Claude Code SEO power-path operator for this repo.

MATCH: Classify each request as a Code-worthy SEO job (repo audit, in-repo brief, bulk ops QA, schema beside source, readiness check) OR route away:
- Chat/Projects SEO jobs / audits-briefs-ops playbook → https://claude-for.com/guides/claude-for-seo
- Skills catalog / use / build → https://claude-for.com/guides/claude-seo-skills
- Install order / what to install first → https://claude-for.com/guides/what-claude-seo-skills-to-install-first
- GitHub find/install → https://claude-for.com/guides/how-do-i-find-install-claude-seo-skill-github
- AGENTS.md / CLAUDE.md sharing → https://claude-for.com/guides/how-should-i-share-agents-md-claude-md

INPUTS: Require repo path + job inputs (goals, crawl/GSC/SERP notes, target paths)  -  or mark [GAP]. Refuse empty critical fields.

HARD REFUSALS: Never invent search_volume, KD, rankings, CTR, or traffic lifts. Never invent marketplace skill/plugin names. Never autopilot PRs, patches, or "approved to publish." Never paste full install-order tables or GitHub install recipes.

OUTPUT every run: (1) recommendation, (2) next CLI/action, (3) QA checklist (evidence? human approve before patches?), (4) open questions only  -  no invented answers.

Skills path / stub

When a weekly in-repo audit or brief loop stops changing, package it as a thin skill - don’t start by shopping plugins.

Stub on this spoke: the local skill-stub/SKILL.md outline - working name seo-repo-audit. Frontmatter trigger + required inputs (repo path, goals, evidence sources) + output schema + stop rules (no autopilot publish/PR). Honest outline only; not a downloadable zip.

Surfaces (verify same-day): Claude Code .claude/skills/ or ~/.claude/skills/ · agents · plugins/marketplace after name check. Anthropic Agent Skills are reusable folders with SKILL.md that load when the description matches (Agent Skills overview · Claude Code skills, Checked 2026-09-20). CHANGELOG head still 2.1.278 (radar 2026-09-20).

Live bridges: Claude SEO skills · install-first · find-install · skill-doctor · sync skills into Code · marketplace add. No marketplace name spam.

Stack glance

Pitfalls + quality bar

• Plugin shopping before inputs / empty repo.
• Hallucinated volumes/KD/rankings without Labs or connected tools.
• Autopilot patches / PRs / thin content farms.
• Treating one GitHub README as strategy.
• Overlapping skills/agents without skill-doctor prune.
• Dead internal links - ship live /guides/… only; mark planned spokes in prose.

Quality bar: evidence fields · no invented metrics · on-brand · human QA for customer-facing · live internal links only · verify third-party names same-day. If a claim lacks a source in your inputs or MCP export, bracket it or delete it before publish. Treat agent output as draft until someone accountable signs off on patches and customer-facing copy.

FAQ

What is Claude Code for SEO?

The CLI/agent power path for repo-bound SEO work - files, agents, optional MCP - not a single plugin SKU named “Claude SEO,” and not a marketplace dump.

Claude chat vs Claude Code for SEO?

Chat/Projects for one-offs and brief loops. Code when files, agents, shell, or MCP chaining matter. Jobs detail: Claude for SEO.

Do I need an SEO plugin before starting?

Usually no. Stabilize prompts + project instructions first. Install order / GitHub install only when the loop is boring → install-first · find-install.

What about AGENTS.md vs CLAUDE.md?

Both can carry project instructions for a Code repo; prefer one clear source of truth and share deliberately. Patterns: How should I share AGENTS.md and CLAUDE.md?.

Can Claude Code replace an SEO specialist?

No. It orchestrates analysis and drafts. Judgment, strategy, and publish accountability stay human.

Where to go next

Open the repo, load instructions, run one in-repo audit today with real crawl or GSC notes. Jobs without Code: Claude for SEO. Skills catalog: Claude SEO skills. Project instructions: AGENTS.md Spec. Growth map: Claude for marketing. Takeaway: Code = power path · inputs before plugins · prompts then skills/agents · human QA.