PowerUp Coaching — Living Knowledge Base

May 7, 2026 — AIMM Mastermind

16 min read

“If you use Obsidian to look through all this, you’re going to get retrieval. But if you point your Claude on here, now you’re going to get inference on your knowledge.” — Lou

This Week in 30 Seconds

  • The Mac Disaster Lesson — Lou wiped and reinstalled from scratch after Apple blocked recovery from a moved home folder. Lesson for everyone: create a second admin account today, back up hidden config files (especially .claude/).
  • Spec-Driven Work Trees — Lou’s new development workflow: you write the spec, AI runs the full cycle (plan → code → test → PR). Work trees isolate experiments so anything that fails gets deleted, not merged.
  • Your Second Brain Is for Inference, Not Search — Kasimir’s Obsidian demo crystallized the key distinction: pointing Obsidian’s search at your vault gives you retrieval. Pointing Claude Code at it gives you inference. The links, tags, and taxonomy are the navigation map Claude follows to reason across your knowledge.
  • The Three-Layer Architecture — Optimal retrieval requires three layers: keyword search, knowledge graph traversal, and semantic similarity. Together, they approach 96.6% of RAG performance while adding temporal search that pure vector retrieval can’t do.
  • Scott’s One-Month Automation — Scott built an automated contact verification system in 2 days using a local Mistral 7B cron job. His EA went from a week of manual work to 10 exceptions to review. One automation a week = potentially saving a year of capacity by year-end.
  • Brand Before Publish — AI-generated HTML slides come in two defaults: purpley-blue or black-and-gold. Dirk’s workflow (Markdown → HTML) produces beautiful output. Lou’s extension: feed it your CSS and stand out in a homogeneous feed.

Topic 1: The Mac Disaster — And What To Do Right Now

Lou opened the session with a cautionary tale: he tried to reassign his Mac’s home folder to an external drive to free up internal storage. Apple’s macOS increasingly locks this operation down, and his system became unrecoverable — even in recovery mode with terminal access. The only fix was a clean reinstall.

Two lessons he wanted everyone to take away immediately:

Create a second admin account on your computer today. If your primary account becomes inaccessible (corrupted profile, wrong path, permission error), a second admin account lets you log in and fix the issue without a wipe. Lou didn’t have one because he’d never needed it in 20 years. He needed it this week.

Back up your app config files, including hidden ones. When Lou restored, he knew to copy his .claude/ folder back in — Claude was instantly at full capability with all memory and preferences intact. Most apps store their configuration in hidden files in your home directory or in ~/Library. Make sure your backup system includes these. If you ever need to reinstall, knowing where the config lives is the difference between a 10-minute restoration and starting from scratch.

Don’s observation in chat: he’d always maintained two admin accounts for exactly this reason. Donald: “That’s how I have my Linux set up.” One of those quiet best practices that feels like overkill until it isn’t.

Topic 2: Agentic OS + Spec-Driven Parallel Work Trees

Lou’s session prep was cut short by the Mac incident, but he described where his thinking is heading this week: a new development workflow built around what he called spec-driven, parallel work trees.

The approach: you write the spec — architecture, functionality, what success looks like — and then Claude runs the full implementation cycle inside an isolated Git work tree. Plan, code, test, retry if it fails. When it passes, issue a PR. If the whole attempt is a failure, delete the work tree. Your main codebase is untouched throughout.

What makes this compelling is the parallelism. Because work trees are sandboxed, you can have 5, 8, even 10 of them running different features or different approaches simultaneously. You are functioning as a product manager and architect — setting direction, evaluating outputs, making decisions — not as a coder executing line by line.

The model allocation Lou described: Opus 4.7 for spec-writing and architecture deliberation; Sonnet for implementation and testing. Expensive thinking at the high-leverage stage; efficient execution at the high-volume stage.

On the broader Agentic OS zeitgeist: Lou is deliberately not adopting OpenClaw or Hermes wholesale. Instead, he sends Claude to read their repos, asks for an architecture spec, and then selectively cherry-picks what fits his own ambient intelligence model. “If there are 3 or 4 different ways of doing something, I can sort of look at how everybody’s doing it and Frankenstein the functionality I want.” The goal: stay in control of the intelligence layer, avoid dependency on frameworks that will rev without you.

→ Deep Dive: Insight - Spec-Driven Parallel Work-Tree Development — Become the Spec Writer, Not the Coder

Topic 3: Agent APIs and the Coming Commerce Layer

A thread Lou wove through the opening discussion: the agent-to-agent commerce infrastructure is being built now, and knowledge entrepreneurs should be thinking about it.

Stripe has launched agent-to-agent communication infrastructure. The implication: AI agents will soon be able to purchase on behalf of users. In preparation, Anthropic and others are encouraging developers to add API-level purchase interfaces to their products — not just the human-facing checkout flow, but a structured API that a purchasing agent can call.

For coaches and course creators: Lou’s framing was “think in terms of making an agent for everything that you can think of.” The schema work you’re doing for GEO and AI citation — adding structured information about your services — is the same layer that agent commerce will read. If your schema says “here’s how a human can buy this,” it will soon need to say “here’s how an agent can buy this too.” The Shopify + Stripe pathway is probably the most accessible starting point.

Lou’s practical note: he’d rather have this work through a subscription than through API fees. Watching, not immediately acting — but worth structuring products with API-accessibility in mind as things evolve.

Topic 4: Claude Rate Limit Update

Lou flagged a platform update relevant to everyone doing intensive work: Anthropic increased the per-5-hour rate limit by 5x. This doesn’t change the monthly cap, but it means that mid-session interruptions — “come back in 3 hours” — should be much rarer for Pro and Max users.

Kasimir confirmed the API speed also improved noticeably, which makes Claude-powered applications smoother. Lou’s note: Enterprise and higher API tiers got the biggest bump, which makes sense from a business standpoint (“the more money you spend, the more benefit you get”).

For those who regularly hit rate limits: this mostly addresses the frustrating “mid-thought” interruption problem. If you were burning through limits on long-context, multi-step work, the window is now wider.

Topic 5: Kasimir’s Obsidian Knowledge Graph — A Live Demo

The session’s longest and most substantive section was Kasimir’s live screen share of his Obsidian second brain, prompted by Dirk’s request after a discovery call earlier in the week.

The architecture Kasimir has built:

  • 450+ raw source files (PDFs, books, articles — including Jay Abraham’s entire archived work)
  • 650+ temporal records (conversation extractions, archaeology from past Claude sessions, project logs)
  • A six-node ontology: entities, concepts, sources, synthesis, index/navigation, and ops
  • A five-pillar taxonomy anchoring all content: consciousness (sovereign identity), thought processes (strategic clarity), feeling/values, AI-augmented leverage, and doctrine/process

The taxonomy is the key structural move. Rather than an arbitrary subject index, Kasimir organized everything against his five professional pillars — the areas where he genuinely claims authority and where his clients’ challenges live. 78% of tagged content fell under pillars 1 and 2 (consciousness and strategy). That distribution wasn’t designed; it emerged from the data, and the data confirmed what his practice is actually about.

The quality filter: every raw source gets a score from 1-10. Only 7-10 scores make it into the vault. The rest are acknowledged and discarded. This stops the second brain from becoming a hoarder’s attic.

The tiered memory system. Two output streams from each conversation: one goes to Obsidian (structured notes with frontmatter), one goes to Pinecone (vector embedding for semantic retrieval). Kasimir built a wrap-up command that runs at the end of each session to populate both streams automatically.

The “archaeology” project. Kasimir went through his old Claude conversation history — Chat, Code, Cowork — and imported the most valuable material into Obsidian. His second brain now contains not just new learning but preserved institutional memory from months of working sessions.

Lou’s addition to the discussion is worth capturing separately as its own insight:

The fundamental distinction isn’t between different search tools. It’s between retrieval and inference. Obsidian native search gives you retrieval — keyword matching. NotebookLM gives you retrieval — semantic similarity. Pointing Claude Code at your Obsidian vault gives you inference — reasoning from your knowledge graph, following relationships, synthesizing across time and source.

The taxonomy and links aren’t metadata for your benefit. They are the navigation scaffolding Claude follows when it reasons on your behalf. Better taxonomy = better inference.

And the practical performance claim: the hybrid of knowledge graph + structured Markdown, when the taxonomy is well-designed, approaches vector RAG performance at roughly 96.6% while enabling temporal and relational queries that semantic retrieval can’t do at all.

Community tips from chat:

  • Donald Kihenja: Native Obsidian search tip: line:("whatever you need to search") — for exact line-level text retrieval
  • Donald on linking: “I let Claude Code sort the linking unless I want to make sure something is specifically linked.”
  • Donald to Elizabeth on getting started: “You don’t even need any plugin. Just open Claude Code, switch to your ‘mother folder’ aka Obsidian vault, then start chatting.”
  • Kasimir confirmed the linking mechanism is bidirectional; Claude Code built it

→ Deep Dive: Insight - Your Second Brain Isn’t a Search Engine — It’s an Inference Engine → Deep Dive: Insight - The Three-Layer Knowledge Architecture — Keyword, Graph, and Semantic Retrieval

Topic 6: Scott’s Local LLM Automation Win

Scott came with a concrete, real-world problem and a two-day solution — the kind of story that makes everything discussed in this group feel immediately actionable.

The problem: Scott’s CEO’s executive assistant manually checks a spreadsheet of 132 contacts (provosts, VPs of Research, CIOs across 22 Ontario universities and 22 colleges) against 44 institutional websites. People change roles constantly, especially in July when 5-year academic terms roll over. The task took her a week of focused effort annually.

The solution Scott built in 2 days:

  1. Started with Claude to understand the problem structure and identify edge cases (many colleges don’t have a formal CIO — just a Director of IT)
  2. Hit his 5-hour rate limit window mid-project and pivoted to a local Mistral 7B model (via Ollama) for continuation
  3. Built a Python script wrapped in a bash shell that: checks if Ollama is running, scrapes each institution’s website, compares current info to the spreadsheet, flags anything that changed
  4. Created a cron job that runs it monthly at 3:01am (sysadmin tip: never between midnight and 3am due to DST transitions)
  5. The script adds a new tab to the spreadsheet with confirmed/changed status for each contact

The EA’s response: “This is going to save me so much time. I just saw that last tab, and I can ignore anything that says Confirmed.” 111 of 132 entries came back Confirmed on the first run.

Scott’s takeaway: “It’s matching the resources to the problem. This is not solving world hunger.” A 7B Mistral model is good enough for structured web scraping and comparison. Using Claude Opus for this would be wasteful overkill.

Lou on the compounding math: “Imagine if you did that once a week. By the end of the year, you’d have 52 automations, each potentially saving you a week. You could save yourself an entire year. Just by creating one automation a week.”

Lou’s additions:

  • OpenRouter — if you need a larger open-source model but can’t run it locally, OpenRouter provides standardized API access to most major open-source models, often free or at-cost. Useful when the local box doesn’t have enough RAM.
  • Batch processing API — both Claude and OpenAI offer 50% off their regular API fees for batch processing (non-real-time jobs). If you’re running scheduled automation, use the batch API and cut your costs in half.
  • Local LLM note: On Mac, always use the MLX-optimized version of any model (Ollama defaults to it). It’s roughly 2x faster than the non-optimized equivalent.

Topic 7: HTML vs. Markdown — Brand-Consistent AI Output

Dirk shared a workflow improvement he’d discovered: draft content in Markdown (readable, editable, easy to revise), then once satisfied with the content, ask Claude to “produce me something beautiful” in HTML. The difference in output quality is striking.

Lou extended it into a brand differentiation argument. The problem with default AI-generated HTML: everyone’s using it, and it looks the same. Claude’s default palette is purpley-blue or black-and-gold. NotebookLM presentations have their own recognizable look. “Your entire feed looks like it comes from one provider.” The signal is becoming as detectable as the AI em-dash.

The solution: force your brand CSS into the output. Ask Claude to read your website’s CSS and use your color scheme and typography. Or give it your hex codes and font stack directly. The output will match your existing brand look. Bally confirmed she’s been doing this consistently with strong results.

Lou’s note on using Claude’s Design feature (web-only): you can design a custom template in Claude’s Design mode on the web, then reference that template in Claude Code for consistent branded output across multiple documents.

Donald: “Take the hard road: differentiate yourself.”

→ Deep Dive: Insight - Brand-Consistent AI Output — Differentiate in a Homogeneous Feed

Topic 8: Canadian AI Funding — SCIP and SR&ED

Bally flagged that she’s seeing consulting demand from Canadian business owners who know AI funding is available but don’t know what to do with it. Two programs came up:

SR&ED (Scientific Research and Experimental Development) — A Canadian federal tax credit program that reimburses qualified R&D spending. Don Back noted it’s reimbursement-based (money already spent), not a grant. Tax credit, not direct funding.

SCIP (Sovereign Cyber Infrastructure Program) — Scott’s domain. 700M+ in hardware. The goal is to host a national-level AI compute cluster. Multiple mega-organizations are bidding: Bell, Simon Fraser University + Queen’s University, and the Digital Research Alliance of Canada. Nothing committed yet.

Lou suggested there’s a consulting opportunity here for members who are Canadian-based or well-networked in Canada (Bally, Scott, Don being the natural candidates). Don’s note: “Chasing money, then trying to figure out what to do with it is the fast road to heartache and frustration” — meaning the money should follow a real business case, not the other way around.

Open thread for anyone interested in Canadian AI consulting opportunities.

Community Corner

Scott Delinger built a local LLM automation that saved his EA a week of annual work — in two days. First real-world deployment of a local model for a structured research task. Proof that 7B parameter models are more than adequate for deterministic, structured jobs.

Kasimir Hedström gave a live demo of his Obsidian knowledge graph — 450+ sources, 650+ temporal records, five-pillar taxonomy, tiered memory to Obsidian + Pinecone. An architecture other members can learn from directly. Kasimir will share the architecture document in the group; if the repo integration is figured out, it’ll go there too.

Bally Binning confirmed the brand-consistent HTML workflow is live in her practice. Also flagged a Canadian AI consulting market that’s emerging as SCIP and SR&ED awareness spreads.

Elizabeth Stief had the session’s most relatable moment: “Now I am off to learning how to better use my Obsidian. 🤯” A breakthrough in understanding what the tool is actually for, plus two practical how-tos from Donald (Claude Code directly in the vault folder, no plugin needed).

Dirk Ohlmeier kicked off the Kasimir demo with a discovery call share. Also contributed the Markdown-to-HTML workflow that launched the brand differentiation discussion.

Lou is rebuilding this week with a new spec-driven, parallel work-tree development workflow. Will share when stabilized. Also reinstalled his Mac from scratch and is now two admins wiser.

Jamie W dropped in briefly — had to leave early for another call.

  • Obsidian native search syntax: line:("whatever you need to search") — Donald Kihenja
  • Kasimir shared screenshots of his Obsidian linking mechanism in chat (Screenshot2026_05_07_230653.jpg, Screenshot2026_05_07_230734.jpg)
  • Claude Max plan rate limit increase (5x on 5-hour block): mentioned by Lou
  • OpenRouter (openrouter.ai) — open-source model access via standardized API — Lou
  • Batch processing API (50% off regular API fees) — Claude + OpenAI — Lou

⚡ Try This Before Next Session

Pick one repetitive task in your business that takes meaningful time but doesn’t require your judgment — updating a spreadsheet, pulling data from websites, formatting a document, verifying information. Build one automation for it this week. It doesn’t have to be elegant. It just has to work. Scott’s took 2 days and saved a week per year. What’s yours?

Open Threads

  • Lou’s spec-driven work-tree workflow — will share with the group once stabilized
  • Kasimir’s Obsidian architecture document — going into the group; repo integration TBD
  • Canadian AI consulting opportunities (SCIP, SR&ED) — Bally, Scott, Don to connect
  • Heather and Joanna extended trial access — group approved
  • Claude rate limits 5x increase — implications for work session planning

Next session: Thursday, May 14, 2026

[← Previous: 2026-04-30_Mastermind] · Sessions Index

Derived Artifacts

Graph View

Backlinks