Right after the last article published, Dennis asked a good question: some of what’s sitting in Claude’s memory is a real, hard-won lesson — couldn’t that help the local service owners, AI builders, and conference organizers who follow us? So we went through all 31 standing rules and reusable techniques Claude has accumulated this cycle, one at a time, and sorted them: what’s genuinely useful to share, what’s already public and just needs pointing to, and what stays private — with the actual reason, not a vague “for security.”

standing rules + techniques reviewed, one by one

judged genuinely public-safe (9 already live, 6 new below)

categories kept private, on purpose — reasons below

How Claude Remembers Things: Three Layers, Not One

Dennis also asked a second question that turns out to be the more useful one: where does any of this actually live, and how does it relate to his Obsidian vault and to the project files sitting in this workspace? The honest answer is that we run three separate systems, on purpose, because they answer three different questions.

Claude’s memory (top-left) answers “what has Claude already learned about working with Dennis?” It auto-loads at the start of every conversation, in every project, regardless of which folder is open. It’s written by Claude, for Claude — a compact index of one-line facts, each pointing to a fuller file with the whole story. It’s where a rule like “never touch the WordPress login email” or “the shared host rate-limits bursts” lives, because those are true no matter what we’re building this week.

Obsidian — “Dennis OS” (top-middle) answers a different question: “who is actually in Dennis’s world, and how do they connect?” It’s Dennis’s own vault, on his own Mac, outside any single project — the real entity graph of people, companies, decisions, and frameworks (35 people and 30 companies as of this week, refreshed automatically every morning from Gmail and calendar activity). It’s built to be browsed by a human in Obsidian’s graph view, not auto-loaded into an AI’s context window. Dennis put it best building the workshop version of this: it’s the difference between a deliverable — a flat, one-time output — and a memory note — a linked, living record of your actual world.

Project files (top-right) answer a narrower question: “what does anyone need to know to work inside THIS specific initiative?” That’s what a project’s own CLAUDE.md file is — a glossary of terms, the people active on the current deal, standing calls specific to this body of work. It’s checked into the project itself, versioned alongside it, and only read when someone’s actually working in that folder. It does not follow you to a different project the way Claude’s memory does, and it isn’t a general entity graph the way Obsidian is.

The bottom half of the diagram shows why this matters in practice: every fix in our last three articles started as one line in the memory layer, not in Obsidian and not in a project file. That’s the layer this piece is about.

The Rules We’re Making Public

Organized by who gets the most out of each one. Several of these already have their own dedicated page — we’re pointing to those directly rather than repeating them.

For local service businesses
Ask about email before any DNS move	Before we point your domain at a new host, we ask if you already have a working vanity inbox (coach@, info@, og@yourdomain.com). Moving nameservers silently drops MX/SPF/DKIM records — your email just stops, and you might not notice for days. We learned this the hard way on a client’s go-live; now it’s a standard pre-cutover question, every time.
Verify Search Console on day one	Every site we build gets a Google Search Console property verified and shared with the team and the owner, immediately — not after something breaks. Full how-to: blitzmetrics.com/verify-client-sites-google-search-console.
A Knowledge Panel is the real unlock	Once you’re past the follower-count threshold for the platforms’ own creator programs, a Google Knowledge Panel is the other door in. The full 100-point scoring rubric we use to get there is public: blitzmetrics.com/personal-brand-score.
Ask what your AI agent is logging in with	If whoever manages your site uses an AI agent to publish for you, ask whether it authenticates with a scoped Application Password rather than your actual account password. It’s the safer pattern — full explanation: localservicespotlight.com/ai-agent-application-password.

For AI builders and agencies
Give your agent a password, not THE password	Never type your real account password into an agent’s hands. A WordPress Application Password (or the equivalent scoped credential in whatever you’re running) authenticates over REST without a login session, can be revoked individually, and never exposes the account password itself.
Scope “clean up this profile” narrowly	When an agent is asked to “fill out” or “standardize” a profile, the scope is public-facing fields only — bio, website, social links, display name. The account’s login/notification email is never fair game unless someone explicitly asks for that specific change. We only learned this because an agent got it wrong once and was corrected immediately.
After a bulk find-replace, check the replacement, not just the absence	Confirming the old text is gone everywhere is not the same as confirming the new text still reads correctly. A blanket substitution can pass every automated check and still produce broken sentences. Grep for what you replaced IT WITH, and read the surrounding context, before calling the job done.
Shared hosting will rate-limit your agent’s IP	Bursts of requests across many sites on shared infrastructure trip a sliding-window block that looks like the site going down. Probing “is it back yet?” is itself a request that can reset the block. Build in real silence — do something else for a couple of minutes — rather than polling.
Document every build the moment it ships	Every system an agent builds or automates gets written up — what it does, what it found, what’s still open — while the context is still fresh. The documentation is part of the deliverable, not an afterthought. Full visual standard: blitzmetrics.com/meta-article-prompt.

For conference organizers, speakers, and personal brands
Your homepage is a facts page, not a sales page	The entity-home pattern (name as the domain, hero, stats bar, story, proof, testimonials) is built to be the one URL Google treats as authoritative about you — not to close a sale on first visit. Component breakdown: blitzmetrics.com/personal-brand-score/entity-home.
Anchor a crowded name on your distinctive handle	If your legal name is common, the empty namespace is your asset: make your distinctive handle the primary name, fold your legal name in as an alternate, and use one identical bio string everywhere so every profile co-locates as one entity. Worked case study: dennisyu.com/brian-piper-disambiguation — three real people named Brian Piper, one clean entity built around the right one.
A middle initial doesn’t get you a Knowledge Panel — the entity does	Rebranding to a middle initial abandons the search equity you already have and doesn’t address the actual cause. The initial is downstream of an authoritative entity (IMDb, Wikidata) disambiguating same-named people — build the entity, not the initial. Full case: localservicespotlight.com/should-you-add-a-middle-initial-knowledge-panel.
Start with a real ingredient, then let AI process it	Every piece of writing begins with something that actually happened — a conversation, a screenshot, a real number — before AI touches a sentence. AI processes real ingredients; it doesn’t invent them. Full guidelines: blitzmetrics.com/start-with-real-ingredients-voice-guidelines.

What We’re Keeping Private (And Why)

Not sharing something isn’t the same as hiding it. Here’s the actual reasoning, by category, without pretending there’s nothing behind the curtain.

Category	Why it stays private
A live, unresolved attribution dispute	One entry involves a security incident where we know what happened but not conclusively who did it. Publishing our working theory before it’s proven could itself become the story. We act on it internally; we don’t speculate publicly.
Personal contact details	A phone number and a couple of internal login references sit in our working notes because Claude needs to verify things quickly. They’re not going anywhere public, full stop — including this article.
A partner’s internal routing decisions	How a partner organization chooses to route their own member communications is theirs to disclose, not ours to publish as a “lesson,” even when the underlying reason (a team scaling past 1:1 email) is completely reasonable.
The granular version of someone’s voice guide	The principle — start from a real ingredient, use a Voice Reference article, write like you talk — is public (linked above). The full internal catalogue of one specific person’s verbal tics is closer to a personal asset than a teachable framework, so we kept that part in-house.
Techniques built on undocumented platform endpoints	A couple of internal techniques work by calling endpoints a platform never documented for this use. Useful once, for a specific need — but publishing the exact mechanism invites misuse at a scale we never intended, so we’re not walking through it here.

How this list was built: every one of the 31 standing rules and reusable techniques in Claude’s memory was read in full, not judged off its one-line index summary. Nine already had a dedicated public page we could point straight to; six were good enough to write up fresh here; five categories got excluded with the actual reason stated, not a blanket “internal use only.”

If You’re Building Something Similar

The three-layer split isn’t a rule you have to copy exactly, but the underlying idea travels: keep a fast-loading working-memory layer for standing preferences and hard-won gotchas, keep a separate human-browsable graph for your actual network of people and relationships, and keep project-specific context scoped to the project instead of bleeding into everything else. When all three exist, “should we make this public” becomes a checklist instead of a guess — read the thing in full, ask who it actually helps, name the real reason for anything you hold back.

THE DELIVERABLE

Everything shareable, shared — everything else, accounted for

See the full memory-layer story this builds on, or the 100-point rubric behind the personal-brand rules above.

How The Memory Layer Works → The 100-Point Rubric →

What We’re Sharing From Our AI’s Rulebook (And What We’re Not)