PowerUp AI Mastermind — May 28, 2026

“Mentally separate the conversation from the library of documents related to your project. The folder is the memory; the conversation is disposable.” — Lou

This Week in 30 Seconds

Dirk’s board letter — how to build a quarterly skill that improves each time it’s used; Lou demonstrated live the “interview to skill” build + self-update loop.
Don Back’s governance brief — RAG-style intake (bylaws, Act, transcript) + Socratic back-and-forth → structured board brief; secondary insight on AI adoption’s tool-first mistake.
Joanna’s legal documents — the conversation-vs-library mental model, NotebookLM as a zero-hallucination citation store, and grounding Claude Code to specific regulatory URLs.
Scott’s document freshness — a periodic sync command to keep local copies of changing external documents up to date.
Donald’s book app — shipped a Kindle-style reader with Opus planning + Codex executing; praised Superpowers methodology.
Lou’s ambient intelligence work — realization that complexity was building where simplicity was needed; switching to Claude + Codex tandem; here.now walkthrough.
Kasimir’s Council questions — profile update mechanics and vetting board members for complementary blind-spot coverage.
Opus 4.8 release — appeared in chat during the session; Mac app demanded three updates.

1 — Dirk’s Board Letter: Build Once, Improve Forever

Dirk is launching a startup and needs to write a board update letter quarterly. He’d already iterated twelve pages down to two with Claude and real-world board-member feedback, but felt uncertain about consistency, accuracy, and legal notes each time. His question: how would Lou approach this, and should there be separate writing and checking skills?

Lou’s answer came in two parts. First, the research foundation: feed in real board letters (public shareholder letters work well) so the model has a feel for the register. Second, rather than asking it directly, use the interview method — tell Claude what you’re trying to produce and say “ask me the questions I need to answer.” Go back and forth, cover the legal angles, refine the tone. When happy with the output, say: “Review everything we’ve done and turn that into a skill.” Then — crucially — after every quarterly use, add: “Learn from everything we did this session and update the skill.” The skill gets smarter each quarter.

On whether to split writing and checking into separate skills: Lou’s instinct was one focused skill for this specific purpose. Board letters aren’t copywriting — they’re a business register for a specific audience with specific expectations. One well-scoped skill handles it.

💡 What This Means for You

The self-update loop (“learn from this and update the skill”) is the difference between a tool that stays average and one that compounds. Add it as a closing ritual for any recurring task.

Deep Dive: Insight - The Self-Improving Skill Loop — Have the Skill Learn From Every Use

2 — Don Back: The Governance Brief and AI Adoption’s Tool-First Mistake

Don shared two connected stories. The first: he used AI to produce a governance vs. operations brief for a not-for-profit board facing a festering governance issue. The process was exactly what Lou described in Dirk’s case — load the bylaws, the governing Act, the meeting transcript, then run a Socratic back-and-forth (“interrogate me, I’ll interrogate you”) that surfaced both the framing and a few sub-issues Don hadn’t consciously named. Result: a structured, credible document ready for the chair and secretary. Don was clear he wouldn’t build a skill out of it — a one-off situation. “If I were still consulting with not-for-profits, yes. This one, no.”

The second observation was forward-looking. Across the companies and PhD grads Don advises, he sees a consistent pattern: 2024–25 was the year of tool-first AI adoption — pick a platform, push everyone to use it, see what happens. What’s now emerging is the corrective: document the process, optimize the process, then let that drive your selection and implementation of the tool. The tool shapes work around its defaults when it leads; the process shapes the tool when it leads. Don framed this as the information → knowledge → wisdom ladder: AI handles the first two rungs superbly; the third — deciding what the AI is for — remains the human domain, and that’s where the value sits.

Scott added a sharp observation in the chat: “I have a hammer — let’s find nails!” is the exact trap Don was describing.

💡 What This Means for You

Before deploying any AI tool, write the process first — even if it’s rough. You’re not ready to automate what you can’t yet articulate.

Deep Dive: Insight - Document the Process Before You Choose the Tool

3 — Joanna’s Legal Documents: Context Architecture for High-Stakes Work

Joanna is self-representing in formal complaints across Ontario Small Claims Court and several regulatory bodies. Her documents now exceed 150 pages. She’d been working through ChatGPT and Claude threads, building in citation-checking loops to catch hallucinations, but still felt 70–80% manual. She asked Lou whether there was a better workflow.

Lou’s first move was a diagnostic: “What are you trying to produce, and where do the citations exist?” From there he walked through two distinct problems and their solutions.

The context architecture problem: Joanna was uploading files into conversations, which filled context with content rather than leaving room for intelligence. Lou’s reframe: “Mentally separate the conversation from the library of documents.” Make a project folder on your desktop, drop every relevant document in, and write a CLAUDE.md at the root that tells Claude what each file contains and where the regulatory URLs live. Use Claude Code, point it at that folder, and let Claude manage its own reading — it pulls only the files it needs per question. When a thread gets long, start a fresh one, tell it to read CLAUDE.md, and you’re immediately back at full context without re-uploading anything.

The hallucination-in-citations problem: Lou offered two grounding mechanisms. NotebookLM: give it the URLs of the relevant regulatory pages (FSRA, Insurance Act, UDAP rules), and it answers only from those sources, returning the reference for every claim. You physically cannot hallucinate from outside the notebook. Claude Code with a ground-truth directive: add to CLAUDE.md — “Consider [regulatory URL] and its subfolders ground truth. Only provide regulatory information from that source. Do not use your pre-training or internet searches unless I request it.” The division of labor: let the locked source verify which citations are real, then hand those to Claude to write the argument.

Scott followed up with a related generalization: how do you keep locally downloaded documents in sync with sources that change occasionally? Lou’s answer: write a command that teaches Claude how to find the most recent version, then run it on a schedule. Compare, update if different.

Lou also closed with a useful reframe for Joanna about AI and thinking: “It can already think — you’re just not using it the way that causes it to think yet.” The shift is from Q&A to delegation: “Here’s my case, here’s what I’m up against. What do I need to know, what questions do I need to answer, what positions should I take?” Then: “Great — do that for me.”

Deep Dives:

4 — Donald Kihenja: Built a Book App with Opus + Codex

Donald is finishing a book on his coaching methodology and wanted a distribution format better than a flipbook reader (which broke on mobile text expansion). He had an extended conversation with Claude about options — which surfaced both paid alternatives (including Book Funnel, which lets him track exactly where readers are and trigger automations) and the possibility of building his own reader.

He ended up building both — a custom Kindle-style app and an evaluation of Book Funnel — and chose to implement the custom app because of the tracking capabilities. The workflow that made it possible: Claude/Opus for planning and architecture, Codex for execution, with Donald asking Codex to report back at checkpoint gates and Claude to review the work. “20–30 commits, all very well documented. For the first time I feel like a real programmer.” He credited the Superpowers methodology (the obra/superpowers GitHub repo, mentioned by Kasimir in a prior session) for much of his confidence in managing the scope.

Lou picked up the thread with his own recent discovery: pointing Codex at the same folder as Claude for a code review found half a dozen things Claude had missed. “I like the two of them working together.” His tentative conclusion: one $20 C l a u d es u b scr i pt i o n pl u so n e$ 20 ChatGPT/Codex subscription is better than two Claude subscriptions. Claude wins on knowledge, planning, and brainstorming; Codex may outperform Sonnet 4.6 on implementation once the spec is detailed enough — and comes with far higher usage limits.

Donald’s app is on Cloudflare Pages. He and Lou discussed subdomain routing (subdomain vs. folder path) for integrating it under his main domain.

💡 What This Means for You

Opus writes the spec; Codex implements it; the second model reviews the first’s code. Cross-model QC catches what self-review misses.

Deep Dive: Insight - Opus Plans, Codex Executes — A Cross-Model Division of Labor for Building

5 — Lou’s Ambient Intelligence Work + Here.now Walkthrough

Lou shared two things he’d been building. The first was a realization about his ambient intelligence framework: he’d been building an agent harness when all he actually needed was to reuse skills and commands more easily. The complexity had grown in the wrong direction. He simplified to a command-line program and did a cross-model code review (Claude + Codex) that found additional issues.

The second was a walkthrough of brave-quiche-rax4.here.now — his current temporary distribution channel for AIM member assets. Here.now is a static file hosting platform originally designed for inter-agent communication; Lou is using it as a website. Each item is an HTML file. The catalog at the moment includes: a Gmail multi-account MCP setup guide, two variations of an article (“Your AI is Lying to You” and “The Most Dangerous Thing AI Does is Agree With You”), the Grounded Challenger voice profile (a blend of Harv Ecker’s grade-6 motivational register with the Experienced Insights Guide structure — vivid but without the wealth-blueprint values), and several teaching blocks from recent sessions.

How the publishing cycle works: at the end of any productive session, Lou presses the teaching-block button in Claude Code. It creates a teaching block, uploads it to here.now, and updates the index. “Speed of completion — I can get knowledge and experience out to you instantly.”

Lou also described the Gmail MCP workflow that replaced GoHighLevel for AIMM emails: exported member emails from GoHighLevel → created a Gmail contact group → tell Claude “send to this contact group” → uses his support@coachlou.com send-as alias. Now his mastermind ingest skill can trigger the email automatically when the process completes, without a separate scheduled automation that could fire on a bad run.

Scott asked about the here.now index ordering (no timestamps visible). Lou explained it organizes by type — setup guide first, then articles, then teaching blocks — and flagged a planned improvement: group by teaching-block thread so members can read a session’s evolution in order. Scott noted: “The article led to the skill development — knowing how to read them in order would let us follow how your brain worked.” Lou confirmed the repo’s Teaching Blocks folder is already grouped by conversation.

6 — Kasimir: Council Profile Updates + Member Vetting

Kasimir asked two questions about Lou’s Council skill. The first: does the Council have a process for updating member profiles when their views change significantly? Lou’s current implementation: a folder of expert profiles, one per member; the moderator dynamically selects which experts to include, reads the relevant profile (or creates it on first use), and updates it at session end with meaningful shifts in perspective or expertise. It’s not deep episodic memory — it’s a one-to-two-page profile of capabilities, behaviors, and general principles.

Kasimir’s version: he uses a 20-point profile as a system prompt, running council members across business decisions in a Claude Project. He has 31 members. His question: is there an easy way to automatically update the profile based on their latest publications or evolving public positions?

Lou’s answer: that’s a different scope — not conversation-level updating but world-level updating. The pattern: a cron job or periodic skill that goes out and researches whatever defines each expert’s profile (their publications, positions, recent interviews), then updates the file with new findings. It would need to be expert-specific since each has different places to look. “A whole different skill unto itself.”

The second question: is there a vetting process for selecting council members that accounts for blind spots — ensuring the council’s perspectives complement each other and don’t all lean into the same gaps? Lou’s honest answer: he’s not working on the debate quality; he’s working on AI-managed state machines using the Council structure as the testbed. The event management and agent coordination is the research concern, not the deliberation quality. “But if you wanted to do that, you’d probably want each one of these as a little agent with some kind of inter-agent communication with a moderator who can spot blind spots.” Complex, but not impossible.

Kasimir also gave Joanna a useful tip for NotebookLM: if you’re hitting the 50-source limit, combine PDFs using any free merger tool before uploading. One combined PDF counts as one source, and the size headroom per notebook is large.

7 — Opus 4.8 Appeared

Donald noted in the chat: “Opus 4.8 has appeared this morning.” Scott added: “That may explain why the macOS Claude app has demanded to be updated three times this morning already.” The group briefly noted this is the current session’s model — the same model used throughout this write-up.

Community Corner

Joanna returned after technical difficulties last week; still building toward using Claude Code for her legal documents and planning to implement the context-folder approach this week.
Don Back contributed a second clean framework this session (process-first AI adoption) in the same crisp three-step form he tends to produce.
Donald Kihenja noted in the chat: “When you talk to Claude, assume it knows 10X more than you on any topic — that POV will guide your questions.” Simple, true, and worth repeating.
Scott Delinger added sharp running commentary throughout: the “I have a hammer” framing for tool-first adoption and the “messy middle is where learning lives” point from Dirk’s thread.

Links Shared in Chat

here.now (Lou’s asset repository): https://brave-quiche-rax4.here.now/
Donald Kihenja’s 77-Day Protocol app: the77dayprotocol-reader.pages.dev
Opus 4.8 release — noted in chat; no link, just the observation it dropped this morning.

Try This Before Next Session

Run the interview-to-skill loop on your most repeated task. Pick the thing you do most repetitively with AI (a weekly report, a client follow-up, a research brief).

Open Claude Code, point it at the folder for that task.
Say: “I want to produce [X]. Interview me — ask me every question you need me to answer to get this right.”
Work through the interview until the output is good.
Say: “Review everything we did and turn it into a skill. Save it as [skill-name].md in this folder.”
Use it once for real. Fix anything that’s off.
Close with: “Learn from the corrections we made today and update the skill.”

You now have a skill that’s already better on its second run than it was on its first.

Open Threads

The Codex + Claude tandem workflow — Lou is planning to open an OpenAI subscription; Donald’s book app is the proof point. Would be worth a full demo session once Lou has run several projects with the split.
Kasimir’s council member vetting for blind-spot coverage — the moderator-as-gap-detector pattern is a genuine open design question, probably a bigger project than it appears on the surface.
Donald’s book app subdomain/folder routing — Lou and Donald to coordinate DNS changes offline.
Scott’s here.now grouping request — Lou plans to reorganize the index by teaching-block thread so the evolution arc within a session is readable in order.

Next session: 2026-06-04

← Previous

Derived Artifacts

(none this session — patterns and prompts discussed are absorbed into existing commands and the six new insight pages)

PowerUp Coaching — Living Knowledge Base

Explorer

PowerUp AI Mastermind Recap — 2026-05-28