Claude Training — for Dawn

An assistant you talk to in plain English

Claude is an AI model made by Anthropic. Claude Code — what's running this session — is the version that doesn't just chat back. It can read and write real files, look at live websites, run commands, and make actual changes to a real computer or server when you ask it to.

You don't write code to talk to it. You write what you'd say to a sharp new hire: "check what's on wholemagnetics.com right now," "build a page that explains X," "did that last change actually work." It asks when it isn't sure rather than guessing.

Two modes worth knowing

• Normal mode — Claude can take action directly: edit files, run commands, fetch pages.
• Plan mode — Claude can only look around and propose a plan; it can't touch anything until you approve it. Good for anything you want to think over first, or anything that reaches a real production system.

Same idea as WholeReach's own product, actually: software drafts, a human approves everything. That's the pattern to expect from Claude too — especially for anything that can't be easily undone.

The real engine lives at automarketingengine.com

Everything WholeReach and WholeSEO sell runs on one shared engine: enter a website, it runs a real audit, and stands up readiness scores, a fix list, and a 30/60/90 plan. When you're testing, ask Claude to run the actual flow against a real domain and show you the real output — not describe what it probably does. "Probably does" isn't a test.

Give Claude the specific thing: a domain, a URL, a report path — not "check the engine." The more specific the target, the less Claude has to guess, and the easier it is for you to verify the result yourself afterward.

The same rule we sell to clients applies to Claude

WholeReach's whole pitch is software drafts, a human approves everything — no auto-send, no invented metrics. Hold Claude to that when you're using it to market our own sites: have it draft the post, the page, the email — you read it, fix what's off, and only then does it go out. If Claude ever wants to publish or send something without showing you first, that's the moment to stop it.

Give it a memory instead of re-explaining

If you catch yourself explaining the same thing to Claude twice — a convention, a "we never do X here," a recurring test — that's a sign to save it (ask Claude to remember it, or write it down somewhere durable) instead of repeating it every session. The second time you write the same instructions, you've found something worth capturing once.

What it can actually reach and do

• Read & edit real files — including the very page you're reading, and anything mirrored from the live server.
• Run commands on this computer — the terminal, PowerShell. This is how sites get mirrored, pages get deployed, and things get checked.
• Fetch and read live webpages — how we confirmed every site description on this page actually matches what a visitor sees right now, not just what a file says.
• Search the outside web — for anything not already in our own network.
• Delegate to sub-helpers — for a bigger job (checking several sites at once, researching several things in parallel), Claude can fan work out and bring the results back together, instead of doing everything one slow step at a time.
• Plan before it acts — "Plan mode" lets it look around and propose what it would do, without touching anything, until you say go. Ask for this on anything that reaches the live server.
• Run something slow in the background — and keep working while it runs. The 380-site mirror behind this page ran that way; you can ask "check on it" any time.

Confidently wrong is the real failure mode

Claude can sound completely sure and still be mistaken. It won't announce its own uncertainty unless asked. Spot-check anything before it matters — open the real page, re-run the real report, don't just trust a summary of one.

Vague requests get the most obvious guess

Name the domain or path exactly when it matters. Earlier in building this very page, "add wholetech.com" turned out to already be there — a vague request got a reasonable but wrong guess at what was missing. Being specific up front saves a round trip.

Long sessions can lose earlier detail

If Claude seems to have forgotten something you covered a while ago, just remind it plainly — don't assume it's being difficult on purpose.

Tools have real quirks — verify big jobs finished clean

While mirroring the network for this page, a Windows quirk silently corrupted a file transfer, and a filter missed a batch of backup folders it should have excluded. Both got caught and fixed — but only because the results were checked, not just the "done" message. Treat any bulk or automated job the same way: verify the outcome, don't just trust that it ran.

If it acts on production without asking, stop it

Anything touching the live droplet should come with a plain explanation before it happens. If Claude just does something risky and tells you after, that's not normal — push back and say so.