Briefing #2
Weekly Intelligence Briefing: April 24–30, 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 |
|---|---|---|
| Apr 24 | Folder as Harness — Concept Articulation + Memory | ⭐ Foundational ambient intelligence architecture |
| Apr 24 | Ambient Intelligence Framework (AIF) v0.1 Spec | Architecture specification — 16 sections |
| Apr 24 | AgentLibrary / Resource Library Build | AI capability management infrastructure |
| Apr 24 | AIMM Teaching Block Orientation | ⭐ Teaching series kickoff for AIMM members |
| Apr 24–30 | Ambient Folder Template — Deployable Scaffold | ⭐ Ready-to-use harness template |
| Apr 25 | robots.txt AI Bot Reference | GEO / AI visibility infrastructure |
| Apr 29–30 | /teaching-block Skill Build + Article | ⭐ Content production methodology |
This Week’s Big Themes
1. The “Folder as Harness” is the implementation pattern behind ambient intelligence. The concept that crystallized this week: a harness is everything an LLM needs except the LLM itself — identity, memory, skills, tools, governance, data. When a folder is intentionally built with this architecture, dropping an LLM into it produces a domain-aware agent immediately. Do this for every folder on your system and intelligence becomes ambient — structural, not session-bound.
2. Portability forces you to find the true core. When you ask “what if this folder needs to work on OpenAI, Gemini, or a local model, not just Claude?” — the exercise forces precision. The answer this week: intelligence lives in universal plain-text files; runtimes are thin adapter shims. This separation produced the cleanest version of the scaffold yet.
3. AGENTS.md is the open standard — and that matters for your builds. Confirming that AGENTS.md (agents.md.org) is a vendor-neutral, cross-tool convention — recognized natively by OpenAI Codex, Cursor, Aider, and others — means the operational driver you put in your folders works across every agentic tool without adaptation. The Claude-specific CLAUDE.md becomes a thin shim, not the brain.
4. Show-and-tell beats narrative for teaching. The /teaching-block skill took three iterations to land, and the gap between iteration 1 and iteration 3 is a transferable insight: optimizing for narrative quality works against replicability. The format that produces teaching with transfer value is first-person past tense, wrong-path-first structure, and verbatim artifacts over description.
Project Deep Dives
1. Folder as Harness + Ambient Intelligence Framework ⭐ HIGH MEMBER VALUE
Session: Ambient Intelligence — Founding Session + multiple iterations
Output: AmbientIntelligence/Ambient-Intelligence-Framework-v0.1.md (v0.1.1, 16 sections)
The session started from a question: “Do you remember my ambient intelligence concept?” — and the answer was scattered across memory. This week, the concept was formally articulated, named, and stored.
The core logic chain:
- A harness = everything but the LLM (context, memory, tools, skills, identity, data)
- A folder built intentionally = a harness for a specific domain
- Drop an LLM into it → it becomes a capable, domain-aware agent immediately
- Apply this pattern to every folder → intelligence is no longer session-bound
- The computer itself becomes the cognitive architecture — intelligence is ambient because it lives in the filesystem structure, not in any single conversation
This is distinct from “organized folders” by intentionality: you’re designing for LLM instantiation from the start.
The framework spec (v0.1.1) resolved 12 open design questions and landed on six governing principles:
- Locality — intelligence lives where the work lives
- Inheritance Without Repetition — context flows downward; child agents define only their delta
- Bounded Memory — each agent’s memory is scoped to its folder; high signal density, not comprehensiveness
- Ambient Invocation — navigating into a folder is the act of invoking an agent
- Composable Without Coupling — folder-agents collaborate via orchestrator without knowing each other’s internals
- Progressive Complexity — structure earns its existence; a simple folder needs only
AGENTS.md
The 5-layer model (Russian Doll):
| Layer | File(s) | When it earns its place |
|---|---|---|
| L1 — Container | harness.yaml, memory/, adapters/ | Always |
| L2 — Operations | AGENTS.md, skills/, tools/ | Any folder that does work |
| L3 — Identity | SOUL.md | When the folder has stable “who” across operations |
| L4 — Use-case shape | inbox/drafts/published/ or communications/sessions/work-products/ | When work has lifecycle stages |
| L5 — Continuity | HEARTBEAT.md | Long-lived autonomous orchestrators only |
The portability decision: Everything intelligent lives in universal plain-text files. Runtime-specific files (CLAUDE.md, system prompts) are thin adapters that point at the core — they don’t contain intelligence, they load it. This lets you swap LLM providers without touching the harness.
The Resource Library (AgentLibrary): Built alongside the spec — five typed agents (Researcher, Copywriter, AudienceValidator, TechnicalWriter, Editor) and five skills (research-brief-engineer, brand-writing-team, cognitive-operations, aif-init, agent-builder). Each has typed interfaces and registry entries. The AgentBuilder is a meta-agent that extends the library by learning from every build.
2. Ambient Folder Template — Deployable Scaffold ⭐ HIGH MEMBER VALUE
Session: Ambient Intelligence — multiple iterations through Apr 30
Output: ambient-folder-template/ — 21 files, fully documented and deployable
This is the usable artifact — not a spec, but an actual folder template you can copy-paste and inhabit immediately.
What’s in the template:
| File | Purpose |
|---|---|
README.md | Human-facing guide: concept, architecture, quick start |
ARCHITECTURE.md | Full technical spec: layer model, interface schema, cascade resolution |
harness.yaml | Manifest — name, version, layers, runtimes, memory-curation mode |
SOUL.md | Identity layer (optional L3) |
AGENTS.md | Operational driver (agents.md open standard) |
memory/ | Index + 4 typed example entries |
skills/, tools/, data/ | Capability and content layers with READMEs |
adapters/bootstrap.md | Universal portable activation — paste into any LLM |
adapters/claude/CLAUDE.md | Claude Code shim |
adapters/openai/, gemini/, opensource/ | Runtime-specific activators |
SETUP_AGENT.md | LLM-facing intake script — “read this if you’re setting up a new agent” |
HEARTBEAT.md.example | Orchestrator extension — rename to activate |
The SETUP_AGENT.md distinction: The README was human-facing and wouldn’t work if you told Claude Code to “follow the README” — it has no intake flow, no way to know what the new agent is for. SETUP_AGENT.md is the LLM-actionable version: it conducts intake (name, domain, purpose, voice, capabilities, runtimes), copies the template, fills in the files from intake answers, and confirms operational. Two separate documents for two different readers.
Cascade resolution: Capability is named, not located. When a folder declares includes: voice-guide, the resolver walks upward through parent folders looking for a registry containing that name. Local scope beats global — the same pattern Claude Code uses for CLAUDE.md inheritance. Promote a local tool to a global registry and every folder that names it picks it up automatically.
How to start: Copy ambient-folder-template, open it in Claude Code, and say: “Read SETUP_AGENT.md and help me create a new ambient agent.” Four minutes later you have a working harness.
3. AIMM Teaching Block Orientation ⭐ HIGH MEMBER VALUE
Session: Apr 24
Output: aimm-teaching-block-00-orientation.md (orientation for a 3-block teaching series)
A teaching series was started for AIMM members — specifically, the architecture knowledge entrepreneurs need to build durable AI capability. The orientation document sets up the three blocks:
- Block 1 — Folder design and context economy. How to organize AI infrastructure so it scales rather than choking on its own weight. Solves: why you keep re-explaining the same context to Claude.
- Block 2 — Resource taxonomy and interface discipline. How to prevent accumulated capability from becoming a swamp of overlapping, ungoverned skills. Solves: why systems that grow enthusiastically become hard to use.
- Block 3 — Ambient intelligence. How to remove the conversation as the activation surface — folders that activate on environmental signals and do work without you. Solves: why your attention is the bottleneck.
The orientation is written for knowledge entrepreneurs (coaches, consultants, course creators, advisors) who are using AI seriously enough to feel friction. Its argument: the practitioners who build this architecture in the next year are establishing compounding advantages. By the time it becomes table stakes, the early movers will have years of system refinement that can’t be shortcut.
“The gap between ‘I have a bunch of prompts I reuse’ and ‘I have a governed system of typed resources with stable interfaces and ambient activation’ is two to three years of compounding work, not two to three weeks.”
The series totals ~19,000 words. Each block is independent. The compoundingArchitectureSeries.zip is ready to distribute.
4. /teaching-block Skill — Show-and-Tell Methodology ⭐ HIGH MEMBER VALUE
Session: Apr 29–30
Output: /teaching-block skill (v3), article: “Three Rounds of ‘Not Quite’ Built the Skill I Needed”
The problem: every productive session with Claude produces a good conversation and nothing shareable. Screencasts and manual write-ups take longer than the original session. What was needed: a skill that reads the conversation and produces a teaching article in one command.
Three iterations. Each one named a specific failure mode:
Iteration 1 — Well-written, wrong framing. The spec optimized for narrative — “discovery memoir,” “narrative arc.” Output read like Medium articles. They were engaging without being useful. A reader could appreciate them without being able to do anything with them.
Iteration 2 — Borrowed structure from a gold standard. Added a style guide reference and extracted the structural pattern from a real exemplar article. Found a 4-beat formula: setup → what was tried → what didn’t work → what did. The key constraint: every section opens with the wrong path before the right one. Never lead with the answer.
Iteration 3 — Show-and-tell, first-person past tense. The final reframe: show the artifact — the actual prompt, the actual output — then explain it. Don’t describe what the reader could see. Plus: first-person past tense throughout the body. “I tried X” instead of “try X” changes who owns the experience.
The five concepts that transfer:
- Describe outputs by the reader’s experience, not the document’s structure
- Don’t conflate engaging with useful — narrative quality can work against replicability
- Steal structure, not just voice — where the wrong path goes, where the generalization lands
- Show the artifact, then explain it — paraphrase loses the specificity that makes process transferable
- First-person past tense is structural, not stylistic — it determines the reader’s role (following vs. executing)
To use: Run any session where the process matters as much as the answer. At the end, type /teaching-block. Works for: client discovery sessions that produced a framework, research sessions where you figured out a new tool, debugging sessions where the fix revealed something non-obvious, strategy conversations that worked through a real decision.
5. robots.txt AI Bot Reference
Session: Apr 25
Output: robots-txt-reference.md — comprehensive AI crawler reference, audit date Apr 24
A full audit of every major AI company’s crawler tokens — prompted by a template that raised more questions than it answered. Investigating Claude specifically revealed it has three distinct bots, not one. That triggered a complete vendor-by-vendor audit.
Key finding: Every major AI company now has multiple distinct crawler tokens for different purposes (training, user-initiated fetch, search indexing). A robots.txt that only lists the most commonly cited tokens is leaving significant coverage gaps.
Structure decision: By-company with stacked User-agent lines, not by type. For a uniform Allow: / policy this is the least error-prone structure — you can verify completeness per vendor at a glance.
What was added beyond the standard template: Claude-User and Claude-SearchBot (Anthropic), OAI-SearchBot (OpenAI), four Meta tokens, Applebot, Googlebot, bingbot, Amazonbot, DuckAssistBot.
This reference document is available for any GEARS client or AIMM member who wants a complete, audited robots.txt configuration.
Operational Sessions
GEARS Client Work
Don Back — PhD Career Academy — Ongoing. Brand guidelines, content strategy, core search queries, ideal client handbook, URL structure, keyword topic map, testimonials collection all in place.
Kasimir Hedström — MindMastery / Sovereign Performance — Phase B continuation. Canonical brand copy assets at four length tiers (150w, 300w, long-form, story asset system). MindMastery ontology activation pending Kasimir’s final term approvals from prior week.
James Wheaton — Platform setup templates from the prior week in place. Ongoing.
Assets Available This Week
| Asset | Description | Status |
|---|---|---|
| Ambient Intelligence Framework v0.1.1 | 16-section spec — full architecture | In AmbientIntelligence/ folder |
| Ambient Folder Template (21 files) | Deployable harness scaffold — copy + run | In ambient-folder-template/ |
| AgentLibrary | 5 agents, 5 skills, registry | In AgentLibrary/ |
| compoundingArchitectureSeries.zip | AIMM teaching series (orientation + 3 blocks) | Ready to distribute |
| /teaching-block skill | Converts any session into a teaching article | Live command |
| Teaching block article | ”Three Rounds of ‘Not Quite’” — publishable | Ask Lou |
| robots.txt AI reference | Complete vendor-by-vendor crawler audit | In Downloads |
Watch List — Developing Threads
- AIF v0.2 — Deferred items: background/event-driven invocation, multi-user/shared folders, worked end-to-end example, A2A pipeline communication design
- AGENT_LIBRARY_PATH global placement — Needs manual edit to
~/.claude/CLAUDE.md; blocked by write permissions in session, requires user action - AIMM Teaching Block Series — Orientation done, 3 blocks ready as zip. Potential to run as a structured AIMM release sequence
- MindMastery Phase B — Awaiting Kasimir’s approval on 11 term definitions + 7 proposed psych nodes before any database activation
- Cognitive Twin — Run 3 calibration at ~0.78. Run 4+ unlocks consultative mode
Generated from active Cowork sessions — April 24–30, 2026.
Compiled for AIMM members.