For Mike · 10/1/26
What Nickis building.
HPM Marketing, 9/17/26 to 10/1/26The work, the system behind it,and the case for a second Claude seat.
This site is built from Nick’s snapshot document, for easier reading.
The content is the file’s, section by section, with one extra page at the end to see it live.
The file drops straight into Claude if you want it summarized.
How to read this: the first table is the request. Section 2 is what got built, Section 3 is the system behind it, and Section 4 is the evidence. Every number comes from files, git history, chat transcripts or the live usage meter, as of 10/1/26.
The request at a glance
| Ask | One more premium Claude seat for Nick, in HPM's Claude organization (Team plan) |
|---|---|
| Why | The account is at 86% of its weekly limit, with 3 days and 10 hours to go (live meter, 10/1/26). When the cap hits, the work stops until the reset. |
| Use | Continuity, not a split. When one seat hits its limit, Nick switches to the other and picks up where he left off. Every chat saves its state to the knowledge base, so nothing is lost. |
| Data | No member data goes into any of these tools. The knowledge base is HPM marketing and business material only. Training on submitted data is off for the outside review tools (Codex and Grok). |
| Cost | The existing Team plan seat price (IT to confirm) |
| Check-in | Review usage together after 60 days |
Section 2 of 8 · The short version
Summary
Two weeks. Seven projects. Mostly one person.
- A product launch campaign for Aiden, RxCare's AI CRM.
- A rebuilt production line for the Closed podcast.
- The 2027 print program and personalized agent pieces.
- A consumer-leads site for ACA and Medicare.
- A five-course customer service training site.
- A daily paid-media check.
- A weekly leadership report.
All of it ran at the same time.
The pace comes from a system, not longer hours.
- 71 skills: versioned instruction files that run a workflow the same way every time.
- 11 scheduled routines, several of which learn from their own results.
- Work routed to the right model for each task, including OpenAI's Codex, which runs on its own subscription.
The record shows it. HPM commits went from 54 to 99 a week in early August to 260 in the week of 9/21, then 204 in the first four days of the next week.
The request: a second seat, so the work keeps going when the first one hits its weekly cap.
Section 3 of 8 · 9/17 to 10/1
What got built
Each project, and the automation behind it.
Every project is published at labs.gethpm.com, the internal site where all of it is reviewed. The full list is at https://labs.gethpm.com/directory/.
Aiden (RxCare CRM) agent launch
A full launch campaign for Aiden, RxCare's AI CRM for HPM's Medicare and ACA agents.
What shipped
- A research dossier and a message house, rebuilt around Gus's answers.
- Five ad directions, static and motion ads, and a landing page.
- Six feature emails.
- A 3:48 agent-journey film, in a short cut, a 4K master and a 1080p version.
- A demo-booking QR code that counts scans.
- A feedback button on every page that sends partner notes straight to Nick.
Nothing reaches agents until RxCare confirms the claims. Mike saw this one in the 9/23 sign-off packet.
How it was built
- The layered research pipeline described in Section 3.
- Independent reviews on the feedback workflow before it went live.
- A measured render benchmark that cut the 4K render from about 21 minutes to 15.5.
Closed podcast production line
Closed is Justin Jacobs' podcast and a core recruiting channel.
In two weeks, an episode went from "edit and post" to a full studio production.
What every episode ships with
- The front end and back end.
- The marketing package.
- An episode toolkit site.
- A guest promo package with its own mini-site.
- Edited shorts, trailers and promo cards, all hosted on labs.gethpm.com.
The episodes
- S4E3 relaunched 9/29: a new cold open and outro, 29 face-safe shorts in four formats, a guest page, a Spotify trailer, and a screenplay guide site for Justin with 14 videos.
- S4E4 releases 10/1: thumbnail, cold open, a guest page with 31 shorts, and posting across YouTube, Spotify, two Facebook pages, LinkedIn and Zoho with tracked links.
- S4E5 and S4E6 are in production, with moment maps of 67 and 48 moments.
Around the episodes
- A YouTube API pipeline for click-through and watch time.
- A weekly channel report and a packaging playbook.
- A team Google Sheet that feeds the public episode schedule.
- 39 comment replies in Justin's voice, across 38 episodes.
- All 21 chapters of Justin's Journey to the Summit set up on Spotify.
Justin has given Nick standing authority to sign off on comments for him. Anything new and Justin-specific, like his AI voice or the guest packages, gets Justin's approval before launch.
See two episode packages
- S4E3 with Umer Khan: the episode toolkit and Umer's guest page.
- S4E4 with Romilda Pierce: the episode toolkit and Romilda's guest page.
The screenplay guide: https://labs.gethpm.com/closed-guide/
2027 marketing materials and agent pieces (built in Codex)
The print program for the year ahead, designed mostly in Codex.
The pieces
- Medicare AEP A-frames, rack cards, postcards, banners and coasters.
- A rebuilt, print-grade HPM logo.
- A Hispanic Heritage Month piece and social concepts.
Pieces for six agents
- Rebecca Kyle: rack cards.
- Mike Daniels: gym A-frames.
- Nadeem Khan: flyers plus a fill-in template.
- Jose Ouseph: a flyer.
- Kelly Harrington: a flyer with a tracked QR code.
- Brenda Powell: feather flags.
19 governed deliverable folders and 73 commits in two weeks. Claude ran about 20 blind reviews on Codex's work as the independent check.
ACA and Medicare consumer leads, HPM Studio, Link Studio
Three builds that sit underneath the marketing.
- Consumer leads: a strategy dossier and site for a new consumer pillar, researched against public sources (CMS, KFF, FTC, GAO). A dark v1 and a light v2 are set up as a split test. Internal only, with no ads running yet.
- HPM Studio: an in-house editor for the labs pages, so text changes no longer need a developer. Nick is happy to walk IT through how it deploys.
- Link Studio: tracked short links and branded QR codes, at linkstudio.gethpm.com.
Customer Service Training (for Cyndy)
A five-course training site for HPM's customer service team.
- Lessons, videos and a feedback page.
- Downloadable artwork.
- Justin's welcome video, re-voiced at a set loudness.
It was rebuilt from scratch after a Codex draft fell short in a side-by-side comparison.
See it: https://labs.gethpm.com/cyndy/
Paid media and the weekly L10 update
The reporting side, mostly automated.
- A daily Meta check: pulls Facebook ad stats every morning and writes a short brief. It never publishes or spends.
- Ads tools: a tracker, an analytics snapshot, a case study and an ad-test creative bench.
- A Google Ads Demand Gen subscriber test for Closed.
- The weekly L10 update: a leadership report, archived by year, month and week, with a weekly video trailer for Closed and Facebook.
See it: https://labs.gethpm.com/weekly/
Section 4 of 8 · How it works
The engine
The skills, routing, reviews and routines that make the pace possible.
This is where most of the usage goes. It's also why the list above was possible.
Content and creative pipelines
One command now runs what used to take a chain of separate chats.
/content-engine(created 9/19): a step-by-step conductor for any marketing build. It goes from audience pain to emotional angle, message pillars, beat sheet, channel copy and a compliance gate. Steps can't be skipped.- A context-free reviewer checks the emotion before Nick sees anything.
- Layered research: each campaign runs Research, then Beneath the Surface (what the research means emotionally), then Content Engine, then Ad Studio. Each layer feeds the next. The four tabs are visible on the Aiden and consumer-leads dossiers.
/hpm-compliance: a gate every outward asset passes. Company facts come from one source of truth. No invented stats, no named competitors, and no generation credits spent without sign-off. On 9/23 it gained a check for six kinds of implied claims.- Supporting skills:
/emotional-method,/copywriting,/humanizeand the Ad Studio framework.
Research, copy, compliance and creative, in one pass.
Video and audio
Every rule learned on the video work is now a required step.
The Remotion engine
/remotion-engine (created 9/21, v1.11 by 10/1) builds videos in code with Remotion. It forces:
- Voice-over generated one line at a time with Fish TTS.
- Whisper word-level timing, so each animation fires on the exact spoken word.
- Text always on screen, with animation as an optional layer.
- Video length locked to the voice-over.
- Per-brand color and type kits.
- A house loudness standard.
Around it
- Loudness targets live in a decision register (for example, -25 LUFS for Justin's voice films) and are applied by
finish-audio.py. - Whisper runs on separate GPU and CPU lanes, so two long episodes can transcribe at once.
- Music beds are generated with Suno, and HPM's library is catalogued by mood.
- Closed has its own skills:
/closed-episode-content,/closed-thumbnail, a face-safe shorts spec and a guest-package spec.
One transcript now produces every title, description, tag set and social post for an episode, in Justin's voice.
Orchestration and model routing
Every task goes to the right model, and new models get adopted on day one.
Routing by fit
An orchestrator skill (v1.12, 9/30) decides who takes each piece of work.
- Every model has a floor and a ceiling: the simplest work worth its cost, and the hardest work it still does well.
- Each task goes to the cheapest model whose ceiling clears it.
- Quality is the entry bar. Cost breaks the tie.
- The result: the highest quality at the lowest token spend.
Who does what
- Opus: design, first builds, video and the judgment calls.
- Sonnet: the lighter work Opus doesn't need to do, such as grunt work, checks, research and browser work.
- Codex: follow-up builds and most reviews, on its own subscription.
- A dispatch cap stops tiny jobs from being handed off, where handing off costs more than doing.
The Codex lane
Four models, matched to difficulty: Astra for hard problems, Sol for substantial builds, Terra for bounded fixes, Luna for mechanical work. Codex builds in disposable staging folders, and Claude reviews before anything is merged. In two weeks, Codex ran about 314 review and build sessions.
Day-one re-tuning
When a model ships, the skills and instructions are audited and re-tuned for it:
- Opus 5: re-tuned at launch, 7/24.
- GPT-6 Astra: routing approved 9/4, verified 9/5 after a CLI upgrade.
- Opus 5.5: 69 skills audited on launch day, 9/23, and installed the same night. That audit produced a
/new-modelskill, so the next migration is a procedure, not a project.
Independent reviews
Before anything goes live, it gets a cold, blind read from models that never saw the reasoning: Codex (OpenAI) and Grok (xAI), alongside a fresh Claude.
They look for errors, inefficiencies, bottlenecks, missed opportunities and failure modes. They share none of Claude's blind spots, so many of the best ideas and preventive fixes come from here.
From these two weeks:
- Two independent reviewers said "simplify", which retired a coordination layer with zero recorded use (9/24).
- Codex flagged a conversion-engine proposal as overclaiming before it shipped (9/22).
- Codex brief checks caught timing and overlap errors before a video was built (9/27).
- Competitor-ad research ran through Grok and Codex, alongside a 34-lens ideation skill.
- The daily evolution routine gets cold Codex and Grok reads on its proposals every morning.
A hash-chained receipt log records each counted review. About 529 were logged in two weeks.
Reliability and redundancy
The system protects its own work.
- The heartbeat runs four times a day, with zero model tokens. It audits the knowledge base, rebuilds indexes, and backs up to two local repositories, with restore drills.
- The save chain (
/save-all,/wrap-up,/close-out) commits work by explicit file path and writes a checkpoint. A new chat resumes from saved records instead of redoing work. - Several AI runtimes (Claude Code, Codex, Grok) work in the same knowledge base. Each files a claim before writing, so sessions never overwrite each other.
- Hooks are allow-listed and hash-pinned. Lint checks guard the instruction files.
- 57 Python tools, with 40 test files.
- 17 recorded rulings and 77 review records in two weeks.
Work is saved, backed up and checked automatically.
Routines that learn from their own results
11 scheduled routines. The most important ones are feedback loops.
The paid-media morning check (daily)
- Pulls Meta ad data through the API and writes a short brief. It never publishes or spends.
- Reads a change ledger first: what the last run learned, and what it still owes.
- Clears what it owes, then writes its own lesson forward.
- Keeps a list of beliefs it once published and later found false, so it doesn't repeat them.
- Gets challenged by cold Codex and Grok reviews. When they argued it might be tracking the wrong metric, Nick weighed it and ruled: cost per registration stays primary (9/17).
Source: projects/recruiting/paid-media/reference--the-recursive-loop.md and loop-change-ledger.md.
The YouTube analytics loop (weekly)
- The YouTube API pull feeds a title and thumbnail playbook built from the channel's own data.
- For example: named-guest titles get a 3.3% median click-through, against 1.9% for question titles.
- The playbook shapes each new episode. The next pull measures the result.
Early signal:
- Baseline click-through: 2.2% lifetime, 2.5% over the latest 28 days (9/1 deep-dive).
- S4E1 (Sergio), the first episode packaged under the playbook, became the channel's biggest episode ever, at 34,087 views.
- S4E2 (Brenda) reached 3.0%, the channel's best click-through at real reach.
Most views come from paid promotion, so click-through is the cleaner signal. Two episodes aren't proof, but the loop is moving the way it was built to.
Source: projects/closed/youtube-channel-report/.
The other routines
- Daily evolution (6:30am): researches improvements to the system, gets cold Codex and Grok reads, and stages at most one candidate a day. Nothing installs without Nick's approval.
- Daily troubleshooting (7am): sweeps the last day's failures and proposes fixes, so repeat problems become rules.
- Also: the weekly L10 update, a lead follow-up check, a morning report, housekeeping, a weekly tool scout and a disk sweep.
The routines recommend. Nick decides.
Section 5 of 8 · Output, pace and efficiency
The evidence
What the numbers show, and what a second seat adds.
What the revision data shows
The measure: how many correction rounds a piece takes.
How we counted
Every message from Nick asking to change something already made, from the first request to his first clear acceptance. Approvals, questions and new scope don't count.
The results
- Before the skills: a median of about 6 rounds per piece (visual and ad work, June to early September).
- Recently: about 5, and 4 for work made after the relevant skill existed.
- Closed thumbnails: 7, then 4, then 2 rounds, across three episodes in a row, as the method became a skill.
Bigger pieces, same rework
Rounds stayed about level. What changed is the size of each piece:
- July: one event ad plus a video took about 8 to 11 rounds.
- Now: a full Closed episode launch (thumbnail, cold open, launch, 29 to 31 shorts, a guest page and posting everywhere) took about 19 rounds in total.
The deliverables got several times bigger. The rework didn't.
Partner direction
Many of today's rounds aren't fixes, either. They're partner direction. Gus sent RxCare's detailed changes for the Aiden launch, and Cyndy is about to send hers for the training site. That's partners steering a draft that already holds up.
Why cost per deliverable drops
Each saving has a mechanism behind it.
- Skills store past corrections, so the same mistake is never paid for twice. Examples: positive-only underlines, a minimum hold before any fade, per-context loudness.
- The compliance gate and blind reviews catch errors, and suggest leaner approaches, before Nick sees anything.
- Codex brief checks caught timing and overlap errors before a cheaper model built the video.
- Routing sends each task to the cheapest model that does it well. Codex work uses none of the Claude cap.
- Checkpoints mean a new session picks up where the last one stopped.
Volume and pace
| Week of | HPM commits | Brain (infrastructure) commits | HPM sub-projects active |
|---|---|---|---|
| 8/3 | 54 | 47 | 5 |
| 8/10 | 99 | 70 | 3 |
| 8/31 | 173 | 245 | 4 |
| 9/14 | 60 | 221 | 3 |
| 9/21 | 260 | 431 | 7 |
| 9/28 (4 days) | 204 | 138 | 6 |
Method: git log author dates, weeks starting Monday, counted 10/1/26. The current week is still moving. Commits measure activity, not effort or quality. "The brain" is the shared knowledge base and infrastructure the HPM work runs on.
The order matters:
- The infrastructure surge came first (8/31 to 9/14).
- The jump in HPM output and parallel projects followed (9/21 on).
Commits alone don't prove cause. The projects in Section 2 and the revision data above are the stronger evidence.
Reliability
In a regulated insurance business, the gate matters as much as the speed.
- The workflow is built to require a documented compliance check before any external release.
- Company facts come from one source of truth.
- Final business and compliance approval stays with people.
- No member data enters any of these tools. None of this work touches member records.
Two catches from these two weeks:
- The 2027 coaster's QR destination was flagged for outdated company facts before it went to print.
- Three independent reviews held back the Aiden feedback workflow until a truncation defect was fixed.
An HPM asset
None of this lives only in Nick's head.
- The skills, rules, method docs, decision register and handoff files are all versioned in git.
- The content engine, Ad Studio, compliance gate and Remotion engine are built for reuse by any HPM team or brand.
- The Remotion engine already keeps a separate kit for each brand.
What a second seat adds
The weekly cap is now the constraint.
The cap
- On 10/1 the meter read 86% of the weekly limit used, with 3 days and 10 hours left. Extra usage: $0.00 of a $50.00 monthly cap.
- Over the last seven days, about 60% of usage went to interactive build chats (82 sessions) and about 40% to routines and reviews. These are relative shares, not dollar figures.
- The work runs in parallel. When the cap runs out mid-week, every project stops at once.
Why continuity works
- Each chat saves its state, decisions and next steps through the save routines.
- Any new session resumes from those records.
- Switching seats is the same handoff Nick already does between chats every day. It isn't a restart.
Why a seat, not extra usage
- Extra usage is metered and capped at $50 a month on this account.
- A second seat adds a full second weekly allowance, at a fixed, predictable price.
Next week alone
- The full Ad Studio build for the Medicare and ACA consumer-leads launch.
- Two more Closed episodes, each a full studio production: marketing, the episode toolkit site, the guest package and mini-site, trailers, shorts and promo cards.
- Cyndy's training revisions.
- The ongoing Aiden work.
What this page cost
- The research, this document and this site were built in one chat, in an afternoon.
- The cost meter puts that chat at about 0.8% of all Claude usage across the past week (a share of usage, not of the allowance, measured 10/1/26).
- The Codex review ran on its own subscription and used none of it.
The ask: a second seat, so the work keeps its pace all week instead of pausing at the cap.
Section 6 of 8 · Problem, fix, rule
How it evolved
Two from building this page, the same afternoon.
This is the loop working in real time. A note becomes a fix, and the fix becomes a rule.
- The rule applies to every build that follows, because the system is built for the future.
- It also goes back and fixes earlier work, so every new discovery reaches the builds already shipped.
Two from this afternoon
- Too hard to read. Nick's note: "this is so hard to read." Every section was rewritten into short, scannable blocks. Within the hour, that pattern became a standing rule for every site, document and piece of content. A blind Codex review checked it before it went in.
- Numbers cut off. The big section numbers lost their bottoms on this page and on the Closed site. One root cause sat in a shared style. It was fixed everywhere it appeared, and the rule went into the site guideline, so every new build starts fixed.
Six more
And six more, out of many. These two weeks alone logged 17 rulings, 77 review records and 12 evolution records, and most follow the same pattern.
- Text vanished in videos. Scenes without on-screen text felt empty. Fix: text is always on screen, and animation is an optional layer. Now a required step in
/remotion-engine(9/22). - Loudness kept drifting. Videos came out too loud, and Justin's side clipped. Fix: loudness targets per context, stored in a decision register the video skill reads first (9/29 to 10/1).
- A long partner note arrived cut off. Fix: feedback delivery was hardened and given a send budget, and three independent reviews cleared it before it went back up (9/24 to 9/25).
- Model names went stale in the rules. Fix: rules name models by role, one file owns model identity, and a
/new-modelskill runs each migration (9/23). - Two chats saving at once blocked each other. Fix:
wait-for-clear.pyqueues the second save, with 35 tests, including one that proves the race test can fail (9/27). It now runs in 17 chats. - Some builds never got used. A seven-day audit found that tools wired into an existing step get used, and tools that are only listed don't. Fix: a "born wired" rule, so every new build names who uses it, and unused builds come up for retirement (10/1, proposed).
Section 7 of 8 · Terms
Glossary
- Skill: a versioned instruction file Claude loads on demand, so a workflow runs the same way every time (for example
/content-engine). - The brain: Nick's git-backed knowledge base (
C:\SecondBrain), with the rules, records, systems and tools every AI session reads. - Orchestrator: the skill that routes each task to the right model and effort level.
- Codex lane: OpenAI's Codex, used as a parallel worker and independent reviewer, on its own subscription.
- Leaf review: a blind, single-level review by a separate model that never saw the reasoning it's checking.
- Heartbeat: the scheduled, model-free job that audits, indexes and backs up the brain.
- Save-all / close-out: commands that commit work and write a checkpoint, so the next session can resume.
- Decision register: the file that holds settled production choices, like loudness levels, so they're applied automatically.
- labs.gethpm.com: the internal site where every project is published for review.
Section 8 of 8 · See it for yourself
See it live
The projects, and Link Studio.
The projects
Everything above is published on labs.gethpm.com.
The directory lists it all.
Link Studio
Link Studio makes HPM’s tracked short links and branded QR codes.
Because it controls where every go.gethpm.com link points, access is by account only, with two-step sign-in.
Your account is ready. Nick texted you your sign-in details.
Sign in with your email and the temporary password.
The first time, you’ll set up two-step sign-in with an authenticator app.
Then change the temporary password in your account settings.