Skip to content
Agents tracked: 502 Releases (7d): 613 Agent status: 4 with issuesSince last limit reset: Claude 7d · Codex 5d For AI agents: Markdown Updated Nov 18, 2026

✓ Checked

Claude Code as an SEO Second Brain: CLAUDE.md, Rules, Skills and Memory

An X Article sets up Claude Code as an SEO workspace with CLAUDE.md, path rules, skills, auto memory and a hook. We checked each mechanic against Anthropic's and Google's docs: most holds, a few points are outdated or imprecise.

On November 2, 2026 @NicholasDulait published an X Article, "Claude SEO Second Brain: The Complete Guide". It describes how the author runs a company's SEO from Claude Code: a short CLAUDE.md as the site map, .claude/rules/ files for page and brief rules, skills for routines such as a page refresh, auto memory for corrections, and a PreToolUse hook for the one thing that must never happen. As reported in the article, the setup has a 128-line map, 17 skills and 171 memory files.

Almost every mechanic in it is a documented Claude Code feature, so it can be checked. We read each claim against Anthropic's Claude Code docs and changelog, and the Search Console part against Google's own documentation. Most of it holds. A few points are outdated or imprecise, and a few things a team should know are missing.

What loads when, per Anthropic's docs

The article's main rule is to put each piece of knowledge in the place that loads only when it is needed. That is also the advice in Anthropic's memory docs. Here is when each piece actually reaches Claude:

ContextCLAUDE.mdEvery session, the whole file (a file over 4 MiB is skipped). Files from your folder and every folder above it are joined, not overridden. The project-root file is re-read from disk after /compact.You write it
ContextCLAUDE.local.mdEvery session, next to CLAUDE.md. Personal notes for this project; the docs say to add it to .gitignore.You write it
Context.claude/rules/ without pathsAt launch, with the same priority as .claude/CLAUDE.md.You write it
Context.claude/rules/ with pathsWhen Claude reads, writes or edits a file that matches the pattern, not when a topic comes up in chat.You write it
ContextSkillsThe name and description are listed in every session; the full SKILL.md loads when you or Claude invoke it.You write it
ContextAuto memoryThe first 200 lines or 25KB of MEMORY.md, whichever comes first. Topic files are read on demand. Stored per repository, on this machine only.Claude writes it
EnforcedPermissions and hooksNot context at all: a deny rule or a PreToolUse hook stops a tool call whatever Claude decides.You write it
Our summary of Anthropic's Claude Code docs (memory, skills, hooks and permissions pages), read on November 6, 2026. The edge colour separates instructions Claude reads and tries to follow from rules Claude Code enforces.

What the article gets right

These claims match the current docs closely, several of them almost word for word:

  • CLAUDE.md is not the system prompt. The docs say its content is delivered as a user message after the system prompt, and that there is no guarantee of strict compliance, especially for vague or conflicting instructions.
  • Memory is context, not configuration. The docs say Claude treats CLAUDE.md and auto memory as context, not enforced configuration, and point to a PreToolUse hook for blocking an action.
  • Under 200 lines. Anthropic's target is under 200 lines per CLAUDE.md, because longer files use more context and reduce adherence.
  • HTML comments are free; imports are not. Block-level <!-- --> comments are stripped before CLAUDE.md reaches Claude. An @path import helps organize a long file but does not reduce what it costs in context. A path wrapped in backticks is left as plain text.
  • Files are joined. Every CLAUDE.md and CLAUDE.local.md from the working folder up is concatenated, not overridden.
  • After /compact. The project-root CLAUDE.md is re-read from disk; nested CLAUDE.md files and path rules load again on demand; instructions given only in chat are gone.
  • Conflicts. If two instructions contradict each other, the docs say Claude may pick one arbitrarily.
  • "Remember ..." saves to auto memory; "add this to CLAUDE.md" writes to the map instead. /memory lists the files, toggles auto memory and opens its folder.
  • /doctor prompt-audit exists and needs Claude Code v2.1.283 or later. It reports outdated or contradicting instructions and changes nothing until you ask Claude to apply its edits.

Where it is outdated or imprecise

The article says What the docs and changelog say
An auto memory index past the limit is cut silently, with no error Since v2.1.210, a memory write that leaves MEMORY.md over its limit produces an explicit error, and Claude is told to rewrite the index. Since v2.1.268 the truncation warning says how many lines were cut and where
A skill's steps take zero context on the days it is not used True for the steps, but each skill's description is in context in every session, and the listing has a size budget
Ask about /pricing and the page rules switch on A path rule loads when Claude reads, writes or edits a matching file. Mentioning a page in chat does not load it
Run /context and look under "Memory files" The docs say /context all; in fullscreen mode /context alone collapses that list
A hook is the only "never" that cannot be ignored A permission deny rule is the docs' tool for a hard deny, and covers more ways to change a file than an Edit and Write hook

The memory limit story is real, but older. As reported in the article, the author's index grew to 213 lines and 32.5KB and 52 lines stopped loading. The 200-line and 25KB limits are exactly as the docs state. What changed is the warning: the docs now say that after Claude writes to MEMORY.md, Claude Code checks it against both limits, reminds Claude to shorten it when it gets close, and returns an error when it is over, because everything past the limit is dropped on the next load. The author's advice still stands: one line per entry, details in topic files.

Skills are cheap, not free. The docs say skill descriptions are loaded into context so Claude knows what is available, and the full skill content loads only when invoked. With many skills, Claude Code may drop some descriptions to fit a listing budget of 1% of the model's context window, starting with the skills you use least. A skill you only run by hand can set disable-model-invocation: true, which keeps its description out of context. After auto-compaction, invoked skills are re-attached with the first 5,000 tokens of each and 25,000 tokens in total.

Path rules key on files, not topics. A rule with paths: loads when Claude uses Read, Write or Edit on a matching file, or a Bash command that reads one, such as cat. "What do you think about our pricing page?" loads nothing until Claude opens the pricing page's source file. Path rules and subfolder CLAUDE.md files also have a per-file size limit (40,000 characters with a 200K context window); past it Claude gets only the start of the file. The docs say this limit applies to on-demand files since v2.1.309.

Guarding robots.txt. The article's hook idea is sound, and the docs include a near-identical "block edits to protected files" example that exits with code 2. But a hook on Edit|Write only sees those two tools. Anthropic's permissions page says Read and Edit deny rules also apply to file commands Claude Code recognizes in Bash, such as sed and tee, and to redirect targets like > file. Neither a hook nor a deny rule stops a Python or Node script that opens the file itself; for that the docs point to the OS-level sandbox. The hooks guide, discussing its if filter, also says to use the permission system rather than a hook for a hard allow or deny.

What it leaves out

  • Auto memory stays on one machine. The docs say it is machine-local: shared by all worktrees of a repository, but not across machines or cloud environments. The "171 files of corrections" do not reach a colleague or a CI run. Anything the whole team needs belongs in CLAUDE.md or a rule.
  • Subagents do not get it. The main conversation's auto memory is not loaded into subagents (a fork is the exception).
  • AGENTS.md. By default Claude Code reads AGENTS.md only when there is no CLAUDE.md or CLAUDE.local.md in the folder or above it. If your site repository already has an AGENTS.md for other agents, see our AGENTS.md vs CLAUDE.md page for the @AGENTS.md import pattern.
  • Debugging loads. The InstructionsLoaded hook logs which CLAUDE.md and rules files loaded, when and why. That is the direct way to see whether a path rule ever matched.
  • Version checks. Several of the behaviours above depend on the version: v2.1.283 for the prompt audit, v2.1.210 for the memory error, v2.1.309 for the on-demand limit.

The Search Console part

The article's page refresh starts with "which of my pages lost clicks over the last 90 days?", answered through ChatSEO, a third-party product connected by its own MCP server. We did not check ChatSEO. The data behind it comes from Google, and Google's docs set the limits:

  • The Performance report shows clicks, impressions, CTR and average position. Its default view covers the past three months.
  • The Search Console API's Search Analytics query returns rows grouped by dimensions such as page and query, sorted by clicks. One request returns up to 25,000 rows (1,000 by default), and Google says the API returns the top rows, not necessarily all of them.
  • Unless you ask for fresh data, the API returns only finalized data, so the newest days may not be in the result yet.

"Lost clicks" means comparing two date ranges, page by page, so it is two queries and a diff. Any MCP server that wraps this API can answer it; our guide on what MCP is lists popular servers. On a large site, check that pages with few clicks are not simply missing from the top rows.

A minimal setup that matches the docs

A .claude/rules/pages.md that loads only when Claude opens a page file:

---
paths:
  - "src/pages/**/*.{md,mdx}"
---
# Page rules
- Never change a page's URL or canonical tag. Propose a redirect plan instead.
- Each page names one target query in its front matter. Keep it.

A skill at .claude/skills/page-refresh/SKILL.md; the description is what Claude sees in every session:

---
name: page-refresh
description: Refresh a page that already ranks. Use when asked to refresh or update an existing page.
---
1. Compare the page's clicks and queries for the last two periods from Search Console.
2. List queries that lost clicks, then the sections that answer them.
3. Propose edits as a diff. Do not change the URL.

And .claude/settings.json with a deny rule plus the hook from the docs:

{
  "permissions": { "deny": ["Edit(robots.txt)"] },
  "hooks": {
    "PreToolUse": [
      { "matcher": "Edit|Write",
        "hooks": [ { "type": "command", "command": "\"$CLAUDE_PROJECT_DIR\"/.claude/hooks/protect.sh" } ] }
    ]
  }
}

A bare file name in a permission rule follows gitignore rules and matches at any depth, so Edit(robots.txt) covers every robots.txt at or under the folder you start Claude Code in. protect.sh reads tool_input.file_path from its input, prints a reason to stderr and exits with code 2 for protected paths; the docs' example script is a good start. After setup, run /context all and check that CLAUDE.md is listed under Memory files.

Claude Code on AgentGid

Claude Code is #1 by Gid Score among all agents AgentGid tracks (84.6 on November 5, 2026), with 12,690,433 weekly package downloads and 154,578 GitHub stars. Its latest release is 2.1.318, from November 4, 2026, well past every version this setup needs. It is included in Claude Pro ($17 a month billed annually, $20 monthly) and higher plans; the Free plan does not include it. Gid Score measures adoption and attention, not quality. For the cost side of a large CLAUDE.md, see our guide on cutting token usage in Claude Code; for a first setup, how to start with a terminal coding agent.

The verdict

  • The placement rule is right. Facts in a short CLAUDE.md, page-specific rules in .claude/rules/ with paths:, procedures in skills, corrections in auto memory. That is what Anthropic's docs recommend too, and it works for any site, not only SEO.
  • Update before you copy it. The memory warning, the prompt audit and the size limits all changed in recent versions; on a current version the "silent cut" is reported to Claude.
  • Use a deny rule for the real "never". Add a hook if you want a custom check, but permissions.deny covers more ways a file can change.
  • Keep shared knowledge out of auto memory. It does not leave your machine.
  • Treat the Search Console step as a data job. Google's API returns top rows and finalized data; compare two periods yourself or through an MCP server you trust.

The Claude Code facts come from Anthropic's docs and changelog, read on November 6, 2026; the Search Console facts from Google's Search Console Help and API reference; AgentGid figures from our catalog. The author's numbers are as reported in the article. We did not set up this workspace ourselves.

Was this useful?

Agents mentioned in this guide

# Agent Gid Score Downloads 7d 7d 30d Stars Latest release
1
Claude CodeAnthropic
87 17.5M up 23.5% down 0.7% 156K+134/day 2.1.3262d ago

Live table, updated daily (figures as of Nov 18, 2026); numbers in the text above are as of the dates given there. "#" is the rank among products of the same kind by Gid Score, our 0-100 composite of adoption, community, attention and momentum (how it is calculated); a dash means no public data.