Briefing #3
Weekly Intelligence Briefing: May 14–21, 2026
For AIMM Members — This report covers Lou’s active R&D sessions from the past week. Emphasis on projects that produced tools, frameworks, and assets you can put to work immediately. Sessions marked ⭐ have high direct member value.
At a Glance
| Session | Topic | Member Relevance |
|---|---|---|
| May 14–21 | LKB-Vault scale-up + /mastermind-triage command | ⭐ The AIMM knowledge base — your inputs are the substrate |
| May 19–21 | LCM-vs-LLM — Anti-Modal Ideation bundle (2 releases) | ⭐ Original-thinking architecture you can install |
| May 14–19 | GIC LEAP — web app shipped, then pivoted to Claude-native skill | ⭐ “Build it as a skill” reframe with a working example |
| May 19 | ”Map the Mind, Not the Voice” — new diagnostic skill | ⭐ Voice profiling that targets cognition, not surface style |
| May 19 | WD18TB drive dedup — pipeline shipped + teaching block drafted | ⭐ Publishable IP from operational housekeeping |
| May 19–20 | T. Harv Eker voice profile + Sakana coherence-brief pattern | Voice library now at 3; multi-article series get a spine |
| May 19 | Sakana 7-article tournament publish (leveragingAI) | Voice-conditioned series production |
| May 19 | AIMM teaching delivery 2 — “Borrowed Fluency / SakanaSpine” | Teaching arc continues |
| May 14 | Subagent fanout demo against accounting folder | Operational pattern study, aborted as demo |
| May 16 | AIMM May 14 mastermind — teaching block produced (no formal recap yet) | Recap gap flagged |
This Week’s Big Themes
1. Modal subtraction is the architecture of original work. The LCM-vs-LLM release codified the move: generate broadly, strip the consensus center, then skeptic-test what survives. It came out of inverting Meta’s Large Concept Model objective — instead of predicting the next sentence-level embedding, predict the cluster the consensus would land on and explicitly produce something else. Lou applied it to his own writing pipeline this week and shipped it as a vault insight. It is a generalizable claim about how to compete in a world where everyone gets the same median answer from the same models.
2. When you catch yourself spinning up infrastructure, ask “could this just be a Claude skill?” The GIC LEAP app shipped Phase 1 of a Next.js + Supabase + Cloudflare Worker stack — and then the question landed mid-stream. A working SQLite + uv run Python + React-artifact skill replaced the hosted version in days, with no auth, no hosting, no DNS. The skill version is live in ~/.claude/skills/gic-leap/ against the same database that drives the dashboard. This is the prototype to point at when someone asks why their internal tool needs to be a web app at all.
3. Triage is the missing loop in a compounding knowledge base. At 147 insights and 6-7 candidates per ingest, scoring alone stopped working — too many borderline cases, too many hub-overload candidates, too many supersession decisions. /mastermind-triage shipped this week as the first vault command that holds a discussion with Lou before promoting, dropping, or deferring. First run: 1 promoted (rescored 65 → 78), 1 deferred, 1 dropped to raw/trash/ — with two cap-evictions documented in log.md for the audit trail. The lesson transfers: any compounding system above a certain size needs a human-in-the-loop review surface, not just a scorer.
4. Voice is a typed library; series need a coherence spine. A third forensic voice profile landed (~/.claude/voice_styles/t-harv-eker-millionaire-mind.md) and got applied to all seven leveragingAI articles plus three bonuses. At the same time, the Sakana writing tournament gained a new primitive — the coherence brief — that holds a multi-article series together as a load-bearing anchor instead of letting each article drift. Voice profiles + coherence briefs + the Sakana tournament now compose into a content stack where each layer is callable as a typed resource.
Project Deep Dives
1. LKB-Vault Scale-Up + /mastermind-triage Command ⭐ HIGH MEMBER VALUE
Sessions: May 18–21 — multiple Claude Code sessions against LKB-Vault/, two process-transcript runs (May 7 and May 21 mastermind transcripts ingested)
Outputs:
wiki/mastermind/.claude/commands/mastermind-triage.md— new interactive review-queue command- 11 new canonical insights across the two ingests (see list below)
- Two new mastermind-derived commands:
wiki/mastermind/commands/brand-html-output.md,wiki/mastermind/commands/knowledge-base-audit.md - Vault state: 45 sessions · 147 insights · 60 article briefs · 20 skills · 43 commands · 5 entities
The LKB-Vault is the AIMM community’s collective knowledge substrate. Every mastermind session you attend, every insight Lou drops in chat, every framework that lands in a Cowork — they’re ingested here, scored, atomized into individual insight files, and cross-linked. This week the vault crossed an operational threshold: scoring rubrics aren’t enough at this scale. Three items from the May 21 ingest flagged for review, two hub insights ready to overflow their inbound-link cap, four old hub-split candidates still pending — the queue was getting wider than the autopilot could handle.
/mastermind-triage is the answer. Lou specified it mid-session:
“Have the triage process surface each of the things that need to be reviewed. A little discussion goes back and forth about how to maybe improve the score or how to make it so that it’s worth approving or promoting. Once some changes are made, then the score can be re-evaluated. If it looks like the score is 70 or better, then I can go ahead and ask for it to be promoted.”
First run cleared all three flagged candidates from the same-day ingest:
| Action | Item | Detail |
|---|---|---|
| Promoted | Recursive Ambient Folders | Rescored 65 → 78 after sharpening |
| Deferred | ROI-Bounded Improvement Loops | Returned to candidate queue with notes |
| Dropped | Dunbar’s Number for Wikis | Moved to raw/trash/ |
Two side effects worth noting. First, the promotion forced cap-evictions on two hub insights (Ambient Intelligence, Design AI Systems for Composability) per the vault’s core-principles.md, both documented in log.md for audit. Second, the new insights from May 21 form a conceptual spine for the vault’s “AI-first knowledge architecture” arc — the seven that landed:
- The Hot Cache, Wiki, Semantic Memory Stack for AI-First Workflows
- Recursive Ambient Folders — The Folder Tree Is a Rule Inheritance Hierarchy
- MCP vs API vs Skill — Three Patterns for Exposing Your Knowledge to Clients’ AI
- Modal Subtraction — Generate, Strip the Median, Skeptic-Test the Outliers
- Ship the Folder, Not the Polished Article — Your R&D Trail Is the Deliverable
- The Universal Audit Phrase — Errors, Omissions, Oversights, Duplications, Contradictions
- Audit From the Avatar’s Eyes — Persona-Bound Quality Review
A verbatim quote from the May 7 capture that sets the orientation for the whole vault:
“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.”
How to use: If you maintain any kind of growing knowledge base (a Notion workspace, an Obsidian vault, a Drive folder), the pattern transfers directly. Score what comes in, but build a triage surface for the borderline. Don’t let “needs review” become “stays in the candidates queue forever.” The mastermind-triage.md file is readable as a template — copy the structure and adapt it to your own scoring rubric.
2. LCM-vs-LLM — Anti-Modal Ideation Bundle ⭐ HIGH MEMBER VALUE
Sessions: May 19 (research) + May 21 (meta-craft pass that detected one-article-trying-to-serve-two-audiences and split it)
Outputs (under /Volumes/Extreme Pro/users/loudalo/GitHub/aimm-shared-repo/LCM vs LLM/):
sakana-lcm-release/article/predicting-ideas-not-tokens.md(3,574 words — foundation)sakana-lcm-release/article/how-i-read-the-lcm-paper.md(1,647 words — craft walkthrough)expert-mind-release/article/how-sakana-lcm-produced-this-architecture.md(3,027 words — application demo)chat3/teaching-block-one-article-to-two-bundles.md(3,962 words — meta on the split)- Three SVG diagrams:
ambient-folder-structure.svg,non-modal-architecture.svg,five-component-architecture.svg
The starting move was a paper read: Meta’s Large Concept Model, which predicts sentence-level SONAR embeddings instead of next tokens. Lou inverted its objective. Where the LCM predicts what the consensus would say next, anti-modal ideation predicts the cluster the consensus would land on — and explicitly produces something else. The pattern decomposes into three stages:
- Divergent generation — produce many candidate ideas across the space
- Structural-analogical retrieval — pull patterns from neighboring domains
- Modal subtraction — strip the consensus center; skeptic-test what survives
The modal floor — the consensus you subtract — must include the author’s own prior work, or you self-plagiarize. This is the part that makes the architecture honest. You’re not just dodging what everyone else says; you’re dodging what you already shipped.
The May 21 session caught a structural fracture in the first draft: a single teaching block was trying to serve both a research audience and a craft-process audience. The fix was a split into two bundles plus a primer, with the meta-teaching block (one-article-to-two-bundles.md) documenting the split itself as the meta-asset. Each bundle ships with its own ambient .claude/CLAUDE.md folder so an AIMM member can drop a Claude into it and immediately understand the architecture.
How to use: The bundles are installable. The predicting-ideas-not-tokens.md foundation article is the lens; the how-sakana-lcm-produced-this-architecture.md is the worked example. If you’re producing thought-leadership content and finding that your drafts feel “smart but unsurprising,” the diagnosis is probably modal: you’re predicting the consensus continuation. Run the three-stage process explicitly on your next piece. The skeptic test in stage 3 is the part most writers skip — that’s where the original survivors get separated from the contrarian-for-its-own-sake noise.
3. GIC LEAP — Web App → Claude-Native Skill Pivot ⭐ HIGH MEMBER VALUE
Sessions: May 15–16 overnight autonomous build (web app, Phase 1 shipped) → May 19 pivot to Claude-native skill
Outputs:
- Web app:
/Volumes/Extreme Pro/users/loudalo/GitHub/code experiments/GIC LEAP Growth Strategy/— Next.js + Supabase + Cloudflare Worker stack, 13 routes, 0 TypeScript errors, deployed to Vercel + Cloudflare - Claude-native skill:
~/.claude/skills/gic-leap/(deployed and live) - Database:
~/gic-leap/data.db(11 tables) - Worktree:
.claude/worktrees/claude-native-skill/withSKILL_PLAN.md(9 phases) andskill/scripts/scanner.py - Live dashboard:
~/.claude/skills/gic-leap/artifacts/dashboard.tsxrendering real positions through the artifact panel
The first build was straightforward — an overnight autonomous orchestration (“while I sleep”) with 5-hour wake-up recovery produced a working web app: GIC (Guaranteed Investment Certificate) yield funding LEAP call options, with brokerage-style position management. Phase 1 deployed cleanly to Vercel and Cloudflare.
Then the pivot. The question — could this run inside Claude as a skill? — produced a working alternative in days. Same underlying data, but now:
- SQLite instead of Supabase (file lives at
~/gic-leap/data.db) - Python scripts via
uv runinstead of Node API routes — including Black-Scholes delta/theta in stdlib only, no scipy - React artifacts in the Claude artifact panel instead of a hosted Next.js dashboard
- CBOE volatility indexes (
^VXN,^VIX,^VIX9D,^VIX3M) for IV Rank, no paid data feeds
The entry gate logic encodes the strategy: 5 criteria (VIX, expiry, OTM%, budget, 200-day MA) with a circuit-breaker rule (exit at -70% gain) and a roll-candidate trigger (≥2.0× multiplier with ≥6 months expiry). Same logic, two implementations. The skill version requires no auth, no hosting, no domain, no DNS — just an MCP-aware Claude.
How to use: This is the answer to “I’m building an internal tool — do I really need a web app?” Most of the time, no. If your tool is for you (or for one client you’re sitting next to), a Claude skill with SQLite, a few Python scripts, and a React artifact is the cheaper, more portable, faster-to-ship build. The skill version doesn’t replace the web app for shared multi-user scenarios — but it absolutely replaces the web app for the case most knowledge entrepreneurs are actually in: building tools to run their own work. Ask Lou for the SKILL_PLAN.md if you want the 9-phase build sequence as a template.
4. Map the Mind, Not the Voice — New Diagnostic Skill ⭐ HIGH MEMBER VALUE
Session: May 19 — full skill built from spec
Outputs:
~/Downloads/map-my-mind/SKILL.md— the skill itself~/Downloads/map-my-mind/article.md— companion article framing the skill’s purpose~/Downloads/map-my-mind/prompts.md— 5 diagnostic prompts~/Downloads/map-my-mind/references/prompts/01-05-*.md— individual reference files
The premise: most “find your voice” work targets the wrong layer. Voice is surface — word choice, cadence, register. Mind is structure — how you decompose problems, what you treat as evidence, where you place stakes, which contrasts you reach for. A forensic voice guide (like the lou-aimm-voice.md shipped two weeks ago) captures voice. Map the Mind captures the cognitive shape underneath.
The skill ships 5 diagnostic prompts that probe different layers of how someone thinks. They’re designed to be run against transcripts of you working (Cowork sessions, recorded calls, even a long Slack thread) — the same source material a voice profile uses, but the extraction targets cognition rather than register.
The skill landed as part of the same week as the T. Harv Eker voice profile, and the contrast is the point. Eker’s voice is highly distinct — contrast triplets, wealth-files structure, declarations, action closes. But Eker’s mind is different: it’s hierarchical, evangelist-shaped, monocausal. A writer who learned Eker’s voice without his mind would produce convincing surface that doesn’t make decisions the same way. The two artifacts are complementary, not redundant.
How to use: Run the 5 prompts against your own working transcripts before you try to “find your voice.” A mind map you can hand to a writer (or a Claude in a voice session) is more useful than a style guide alone. The artifact this produces — a structured map of how you think — also doubles as input to AI systems that need to operate as you on a specific task: it’s the substrate a high-fidelity Lou-clone would need to run consultatively, not just imitatively.
5. WD18TB Drive Dedup — From Watch List to Publishable Article
Sessions: May 19 — pipeline build-out + teaching-block drafted
Outputs (under ~/wd18tb-dedup/):
- Pipeline:
analyze_v2.py,build_keepers_v2.py,find_backup_chains_v2.py,consolidate_backup_chains.py,clean_keepers.py,rerank_keepers.py - Tracking:
PROGRESS.md,PROJECT_README.md - Article:
article/teaching-block.md+ 3 SVG diagrams (dedup-decision-tree.svg,openrsync-trap.svg,pipeline-overview.svg) - Build artifact:
dist/article/teaching-block.html
The WD18TB dedup project was on last week’s watch list as “operational housekeeping.” This week it crossed the line into publishable IP. The pipeline is now a proper toolchain: context-aware analyzer, keepers builder (v2), backup-chain finder and consolidator, re-ranker, cleaner — with an rmlint run complete and the keeper-selection logic still iterating.
The teaching-block article names a specific gotcha — the openrsync trap — that surfaced during the dedup work and is the kind of discovery that other people will hit. The diagram set makes it concrete (openrsync-trap.svg shows the failure mode visually). This is the pattern: operational projects that look like housekeeping often produce the most teachable artifacts, because the gotchas are real and named after you survived them.
How to use: When your “tedious housekeeping” thread runs into a gotcha worth naming, write the teaching block while it’s fresh. The pipeline scripts themselves are reusable as a template for anyone facing a multi-terabyte drive consolidation. Ask Lou for the WD18TB folder if you have a similar problem.
Operational Sessions
Content Production Stack
Sakana 7-article series — leveragingAI tournament publish. Shipped seven published articles plus seven Lou-voice variants and the coherence brief that holds the series together. Files at ~/Downloads/leveragingAI_zi_typed_donut/article-{1-7}-published.md and /lou-voice/article-{1-7}.md. Coherence brief is the new structural primitive — it doubles as a series anchor and a quality gate.
AIMM teaching delivery 2 — “Borrowed Fluency / SakanaSpine.” Frame + map-dimensions evolution shipped at ~/aimm-teaching/delivery-2/aimm-frame-winner.md and aimm-map-dimensions-final.md, plus three voice-variant drafts (prosecutor, storyteller, cartographer). Continues the teaching arc; ready as a member release.
T. Harv Eker voice + Eker-styled article variants. Voice profile at ~/.claude/voice_styles/t-harv-eker-millionaire-mind.md. Applied to all 7 leveragingAI articles + 3 bonuses at ~/Downloads/leveragingAI_zi_typed_donut/eker-style/.
Personal Infrastructure
Home folder backup automation. ~/scripts/homeback.sh + ~/Library/LaunchAgents/com.loudalo.homeback.plist. Daily cumulative archive on Extreme Pro fed by rsync. Shipped and scheduled.
Statusline overhaul. ~/.claude/statusline/statusline.sh got 5-hour + weekly usage-limit bars below the existing context/cost line, with a brighter grey contrast pass after readability was flagged in-session.
ambience-claude/ reorganization. Audit and inheritance-ordered consolidation of scattered .claude/ folders and CLAUDE.md files across ~/loudalo/. Outputs at ~/ambience-claude/AUDIT.md, build.sh, ACTIVITY_LOG.md. Reorg pass complete; dedup pass still pending.
Daily AI news brief. Scheduled run shipped to ~/Downloads/AI-Digest-May-20-2026.md. Continues as a recurring scheduled task.
Mastermind Processing
May 21 mastermind recap. Produced via /zoom-mastermind-recap at ~/Documents/Zoom/2026-05-21 15.00.48 AI Mastermind/AIMM-Recap-2026-05-21.md. Explicitly not published to Notion per Lou’s instruction. Bot intrusion logged — unknown bot joined, refused to identify, was kicked. First occurrence.
May 14 mastermind — teaching block only. A May 16 session produced teaching-block.md (22KB) plus two SVGs (context-vs-disk.svg, receipts-pipeline.svg) from the May 14 Zoom folder. No formal session recap exists yet at wiki/mastermind/sessions/ for 2026-05-14 — only May 7 and May 21 are present. Flagged as a vault gap.
No GEARS client work surfaced this week. Don Back, Kasimir Hedström, James Wheaton — none appeared in any transcripts in the window. Flag in the watch list.
Assets Available This Week
| Asset | Description | Status |
|---|---|---|
/mastermind-triage command | Interactive review-queue workflow for vault candidates | Live in wiki/mastermind/.claude/commands/ |
| LCM-vs-LLM Anti-Modal Bundles (2) | sakana-lcm-release/ • expert-mind-release/ — installable ambient folders | Shipped in aimm-shared-repo |
| GIC LEAP — Claude-Native Skill | ~/.claude/skills/gic-leap/ — SQLite + Python + React artifact | Deployed; ask Lou for SKILL_PLAN.md |
map-my-mind skill | SKILL.md • 5 diagnostic prompts + companion article | In ~/Downloads/map-my-mind/; ready to publish |
| T. Harv Eker voice profile | t-harv-eker-millionaire-mind.md — 3rd named voice | Live in ~/.claude/voice_styles/ |
| Sakana coherence-brief pattern | Multi-article series structural anchor | Pattern established via leveragingAI run |
| WD18TB dedup teaching block | Article + 3 SVGs + working pipeline scripts | In ~/wd18tb-dedup/; ready to publish |
| 7 new vault insights | The May 21 ingest — AI-first knowledge architecture arc | Live under wiki/mastermind/insights/ |
| Home-backup automation | homeback.sh • launchd plist | Daily scheduled |
| Statusline (rev. 5h/weekly bars) | Updated statusline.sh with usage-limit bars | Live |
Watch List — Developing Threads
- May 14 mastermind recap missing. Teaching block exists; no formal session recap in
wiki/mastermind/sessions/. Needs to be backfilled before the May 14 source material rots. - GEARS client work paused this week. Don Back, Kasimir Hedström, James Wheaton — none touched. Prior-week threads (MindMastery Phase B awaiting Kasimir’s 11-term approval) still open and now further past their natural cadence.
- Hub-overload splits still pending. Four hub insights flagged 2026-05-10 remain unresolved: Process Architecture (17 inbound), Skills Encode Judgment (16), Trust Before Automation (16), Teach One Era Ahead (15). Ambient Intelligence projected to cross 15 inbound on the next ingest — fifth split candidate.
- May 7 review-queue items still pending triage. Five-pillar-taxonomy-pattern, match-model-to-task, one-automation-a-week — flagged but not yet run through
/mastermind-triage. - GIC LEAP Claude-native skill phases 2–9.
SKILL_PLAN.mdis laid out; phase 1 deployed. Next phases include alert engine integration, position-roll suggester, scenario simulator. ambience-claudededup pass. Reorg pass complete; the actual rule-deduplication across the consolidated hierarchy still pending.process-transcriptrefactor deferred. Lou flagged he “may modify the process and decompose some of the parts” — memory-saves on insights paused until the refactor lands.- Bally / EU AI Act series — candidate skill spin-off. If a V1 ships,
walk-any-business-through-disclosure-requirementsflagged as candidate skill in the vault. - Zoom bot intrusion. First occurrence on the May 21 session. No pattern yet, but flagged as a security thread to watch.
_backlinks.jsonregeneration deferred. Bulk bidirectional updates queued for ~28 inbound-link targets touched by the May 21 ingest — runs at next/mastermind-cleanup.
Generated from active Claude Code sessions — May 14–21, 2026.
Compiled for AIMM members.