For Mike · 10/15/26
Your 90-day plan,point by point.
My answers to your questions, in order,with the numbers on the next tab.
Hey Mike,
Thanks for going through all of it, and for saying the work is landing with the team. Here are my answers to each point, in order. The detail, and the source for every number, is on the next two tabs.
How this page works
- The tabs at the top: my reply, the numbers on one page, and a checklist of your points with where each figure comes from.
- The menu on the left: the sections of the tab you’re on.
- The menu on the right (wide screens): the parts of the section you’re reading.
- “Shown on The numbers” links: jump straight to the detail and the engineering behind a point.
Your place is saved in this browser only.
Section 2 of 6 · Your message
On the 6–7x
That number doesn’t surprise me. I’d estimate I’m running about 6–7 times as many projects in a given week, across several departments, at the same time.
What the lines count
Anthropic counts “lines accepted” as every line Claude Code writes in any file, and it never subtracts later deletions, so every rewrite counts again.
I measured my September from my own session logs: 394,000 to 408,000 lines, depending on how edits are counted. Your 402K sits right in that range.
What they are
| Kind of file | Share of September's lines |
|---|---|
| Scripts | About 46% |
| Documents: briefs, plans, handoffs, records | 25% |
| Web pages and site code | 16% |
| The code behind our videos | 9% |
| Data and settings | 4% |
Many of the scripts are short working scripts Claude writes to get a job done and then throws away.
What that means
So it measures volume across departments, not app-code complexity.
What you shipped in HQ, Clients and the RxCare integrations, is a different kind of work, and I’m not comparing the two.
Section 3 of 6 · Your message
Process, not output
And the collateral.
The engineering underneath
You’re right that a big share of my usage isn’t a finished ad or page. It goes into the engineering underneath, built so everything we put out is consistent, compliant and on-brand on every layer: words, compliance, voice, music, video, pictures and context.
Under that sits shared memory across chats, so I can run 7–10 chats at once without anything colliding or getting lost. It’s a factory that builds workflows that build more workflows, and the collateral is what comes out of it.
Closed, for example
Justin taps twice, and the Episode Builder researches the guest, builds the episode plan from our episode guide, checks it, and sends him a finished prep page and a guest site.
My side, the production, is largely automated, with minor edits and my approval at each publish: editing, music, trailers, thumbnails, YouTube and Spotify, 20 shorts per episode, subscriber ads and daily analytics. That replaced Descript, CapCut, encoding in DaVinci Resolve, and every manual step in between.
Section 4 of 6 · Your plan
Your five points
One step per point, in your order.
1. Hosting and DNS under DigitalOcean
No problem. Once you’ve set up company-managed access and any standards you want followed, I’ll move gethpm.com off the Hetzner server, put its DNS on our central DNS server, and move the rest over.
2. HPM code in HPM’s GitHub
Agreed. The CI, rollbacks and visibility are worth having. Once this week’s launches are out, I’ll migrate the HPM repos into company-owned repos in the Health Plan Markets org.
Is there a naming or branch setup you want me to follow?
3. Why Codex and Grok
Claude is still my lead for planning, judgment, writing and design. Codex does three jobs.
It builds and designs
It did much of the 2027 marketing materials and built the Closed Episode Builder and connector, HPM’s lead relay and the Medicare landing pages. I review all of it before anything goes live.
It reviews, blind
About 850 independent reviews ran from 9/8 to 10/13, about 735 of them on Codex. Where a review gave a clear verdict, about half needed changes before shipping.
It caught a Medicare page that would have said “your request is in” even when it failed, and a lead form that would have counted an unrelated checkbox as consent.
It saves Claude usage
Priced the same way, Codex’s work equals about 15% of my Claude use in September and about 21% so far in October, including the work I run directly in Codex, which is mostly design. Without it, I’d need about a fifth more Claude usage this month.
I’ve maxed out Codex’s own weekly limit four times. OpenAI also resets Codex limits often; in April its CEO reset them for everyone “to celebrate 3 million weekly Codex users.”
Grok is just an occasional third opinion.
3a. “These two calls”
Sending work to Codex costs Claude about 2,000–3,000 tokens per job, about 1–2% of my Claude usage counting the context it re-reads, and Codex does about 6–7 times as much work in return.
If the dashboard showed you something else, point me to it and I’ll dig in.
4. Local processing
Renders already run locally, daily and overnight, including on the laptop. The analytics come straight from the platforms’ own APIs, like YouTube’s, not Claude’s API; Claude’s part is reading them in our daily routines.
If you saw something specific that looks like cloud API use, point me to it and I’ll check.
What does use Claude is the thinking before a render (script, structure, copy, compliance), which is what keeps the output consistent.
5. 30-day review
Sounds good, happy to review it together.
I’m all for optimizing, and I already do it daily, for better token use and performance. But with the right amount of usage, most of this becomes a moot point.
Rationing rules like fewer reviews, shorter chats or juggling accounts would cost more in time and efficiency than they save. That’s penny-wise and pound-foolish.
Section 5 of 6 · The ask
The bigger picture
What the role covers
My role started as recruiting (lead generation included) and social media. It has grown into multiple departments, run solo: media and video production, design for the company and our agents, and the engineering behind all of it. HPM isn’t paying for those roles anywhere else.
What makes it work is the combination: capable AI, plus being able to organize the work, orchestrate the models and run it all at once. That’s what carries the Medicare campaign, Closed, the design work, and any other large project that crosses the desk, like the RxCare campaign and video, Cyndy’s customer service course, and the AEP kickoff video Justin asked for two days before he needed it.
Costs already cut
- Recruiting texting: about $32,000 this year, moving from WooSender to Salesgod.
- Subscribers: bought subscribers replaced by our own YouTube ads, at about 9.6 times as many per dollar.
- Tools: Descript, CapCut and DaVinci encoding are gone from the workflow.
- Assistant role: in the next month or two, the workflows I’m building should cover the assistant role, saving $1,100 a month. That takes the AI usage to build.
Section 6 of 6 · The ask
The account
Thank you for offering the temporary second license. Here’s why I’d ask for a different setup, using Anthropic’s published plans.
The plans
- Today: HPM pays $275 a month for my two accounts: my Team Premium seat ($175: $125 for 6.25× Pro usage plus $50 of extra usage, which ran out this month) and a Claude Max 5× plan ($100 for 5×). That’s 11.25×, split across two accounts.
- Two Team seats: $350 ($175 each) for 12.5×.
- One Claude Max 20× account: $200 for 20×.
Why one account
Max 20× is $75 a month less than today and $150 less than two seats, for almost twice the usage I have now. In any 5-hour window it’s more than three times what one account gives me, because two accounts can’t be pooled.
On efficiency: two accounts aren’t efficient, and one Team seat alone simply can’t do the job.
One account also ends the account juggling: today I switch midweek and again at the start of the next week, with handoffs, re-authorizing Google and app access, and a dashboard that only shows you part of the work.
Data and terms
Anthropic allows a personal and a Team account under the same email address. No member data goes into these tools, and my training setting is off.
If a personal plan raises anything on HPM’s side, like ownership, retention or offboarding, let’s confirm it together.
My ask
One personal Max 20× account at $200, instead of two accounts of any kind.
If the goal is solely to track usage, the current setup limits the very usage we’re trying to measure, and at an unnecessary premium.
Happy to walk through all of it when we meet.
Nick