Brain (202)

Thought Topics Date
Model routing for client work: local models (Ollama) are viable for mechanical and bulk tasks — routing decisions, transcription, brand voice with a good system prompt, status updates. Keep capable frontier models for domain expert work where errors are costly — clinical reasoning, Jyotish interpretation, nuanced synthesis. The consulting differentiator isn't picking the best model for everything; it's knowing which tier each task belongs to and routing accordingly. Clients who run everything through GPT-4 are leaving money on the table and paying for it in latency. consulting,ai-agents,model-routing,positioning 2026-04-05
Two-tier processing architecture: mechanical pipeline (yt-dlp, faster-whisper, cron) handles dumb work autonomously; AI analysis (Fabric, Scout, Claude) handles judgment work on demand. Never mix them in the same trigger — cron does not call agents, agents do not run on cron. The tier boundary is the design decision. Applies directly to client agentic systems: separate the mechanical from the intelligent, and make the boundary explicit. architecture,ai-agents,consulting,systems-thinking 2026-04-05
CLAUDE.md as a per-turn tax — every token in CLAUDE.md and MEMORY.md is charged on every single interaction before Kyle types a word. Treat them like a rate card, not a reference manual. Pointers to files beat inline content every time. The skeleton+links pattern (5 rules, 3 file paths) scales indefinitely; the 2000-line handbook pattern compounds cost with every message. Applicable to any system prompt design for clients. claude-code,consulting,architecture,cost-optimization 2026-04-05
Single task system, multiple entry points — lazlo.db on Citadel is the source of truth for all tasks. Web app (remote capture), Claude Code/Lazlo (session work), and future RemoteTrigger are all entry points into the same DB. todo.md is a legacy pattern: a local flat file that creates parallel systems and forces manual sync. For client work, this is an antipattern worth naming: any task system with more than one source of truth is already broken. architecture,consulting,systems-thinking 2026-04-05
Sovereignty principle for content infrastructure: local-first beats API-first for personal systems. Clone once, read forever. No runtime dependencies, no rate limits, no vendor lock-in, no daily API calls. ACIM (365 lessons), Stoic quotes (krasimir/daily-stoic), Catholic saints (cpbjr/catholic-readings-api) — all cloned to Citadel, all read from local files by date. Update clones quarterly with git pull. The same principle applies to client work: any content source that can be self-hosted should be. Reduces dependencies, improves reliability, keeps data inside the building. infrastructure,sovereignty,consulting,architecture 2026-04-04
The Merlin move in AI consulting: clients don't see code, they see a team. Kyle did the systems thinking — stakeholder analysis, constraint mapping, spec, architecture decisions. Bob wrote the Python. The consulting differentiator isn't being able to build everything yourself — it's hiring the right agent, scoping their authority correctly, and verifying their output. If you can define the problem, write the spec, and know when the output is right, you have the whole game. That looks like magic to people who think they need to write the code themselves. consulting,ai-agents,positioning,systems-thinking 2026-04-04
Put the database near the writer. The write path determines where the DB lives — everything else is a read replica. If Claude Code on Bohica is the primary writer, kybrain.db stays on Bohica. If the web app on Citadel is the primary writer, lazlo.db stays on Citadel. Don't fight the write path. Applies directly to client work: when you see scattered data across systems, find who writes what and where — that's where the DB belongs. Everything else is a sync problem, not a storage problem. architecture,data,consulting,systems-thinking 2026-04-04
idea → convo → spec → build → test. ADDIE translated to AI consulting methodology. Keeps Kyle in the discernment lane (ideas, analysis, verification) and agents in the execution lane (building, syntax, implementation). The methodology clicked as a lived experience today — not just a concept. The shift: you don't need to be an expert developer to direct one. Knowing what to build and why is the leverage. Bob can write Python. Kyle can't out-Python Bob and shouldn't try. consulting,methodology,ai-agents,addie,systems-thinking 2026-04-04
Two modes of AI work — never mix them. Gather mode: exploratory, multi-source, disposable sessions, light and loose. Execute mode: pre-gathered context, single objective, clean handoff, scoped brief. Mixing them burns tokens, confuses the model, and buries the original instructions. The mode-shift signal: when 'this is interesting' becomes 'I know what I want to do' — that's the /wrapup trigger. Fresh session for execution. — 2026-04-03
Production agentic systems are 80% non-glamorous plumbing and 20% AI. The boring stuff is the whole game: session persistence, workflow state vs conversation state, permission tiers, token budget tracking, verification loops. Most clients think they have an AI problem when they actually have a software engineering problem. The 12-primitive framework from the Claude Code leak is a consulting diagnostic checklist in disguise — use it to name exactly what's missing when a client's AI system is breaking. — 2026-04-03
The future isn't fitting into someone else's workflow — it's building exactly what you need for where you are now. Prompt engineering is dying because models are getting smarter; the craft shifts from "how do I phrase this" to "do I actually know what I want." That's harder. Generic apps are becoming obsolete; custom-built workflows calibrated to your exact context are getting more valuable as models get good enough to execute on them reliably. The bottleneck was never model capability — it's clarity of intent. Jin didn't ship claw-code because of better API access. He shipped it because he could decompose the problem so cleanly the agents didn't need him. That's the skill. Clients don't need another SaaS subscription. They need someone who can extract what they actually want (which they can't always articulate) and build the custom system that does exactly that. The app store is a compromise. The consulting value is the alternative to the compromise. — 2026-04-03
Hybrid cloud+local AI architecture is a consulting differentiator. Most AI consultants default to cloud-only. The consultant who asks "what data can and can't leave your building?" immediately signals a different level of expertise. The value isn't in being a deployment engineer — it's in the discovery question. Triggers: data residency/compliance, IP protection, cost at scale (volume inference), latency requirements. Lead with the business problem, not the architecture. Watch the edge AI market ($25B → $143B by 2034) — this shapes what clients will start asking for. — 2026-04-03
AEIOUY — a coaching framework Kyle developed, grounded in Wayne Dyer, Tao Te Ching, Meditation, and Stoicism: A=Attention/Awareness, E=Expectation, I=Intention, O=Outcome Independent, U=Unique, Y=YOLO (live fully, present). The O/I consulting methodology is rooted here. The framework encodes the operating philosophy: awareness before action, intention without attachment to outcome, own your uniqueness, stay present. It connects Kyle's inner work to his outer consulting identity. aeiouy,coaching,framework,philosophy,wayne-dyer,tao,stoicism,valis 2026-04-01
Kyle's core superpower is Discernment. He sees 2nd and 3rd order effects before others have finished describing the problem. Systems thinking is the mechanism — he builds a mental model of the whole system quickly and spots where it will break, where the costs compound, what question nobody's asking. This is not a learned skill in the traditional sense. It's the observer stance applied to complex systems. In consulting terms: he finds the arche (the originating question) before proposing any solution. discernment,systems-thinking,superpower,consulting,valis 2026-04-01
Kyle's inverted triangle consulting framework: start at the bottom point where all foundational questions must be answered — purpose, functions, rules, modification, interface, timeline. As development moves up the triangle, the cost of changes scales exponentially. Getting the point right is everything. This is why Kyle starts every engagement at the vertex before anything gets built. Same principle applies to AI system design: architecture decisions made late cost 10x what they cost made first. valis,consulting,systems-thinking,framework,architecture 2026-04-01
Kyle's consulting methodology is O/I — Outcome Independent. Flip of the classic I/O (input/output) from programming. The entry point is the Apollo/archer allegory: you do all the prep, put in the reps, control everything you can control. Once you release the arrow, you accept the outcome. Kyle's role in the system: use systems thinking (his superpower) to maximize the probability of hitting the mark before the arrow leaves the bow. O/I is both a coaching framework and a consulting operating principle. valis,consulting,oi,outcome-independent,methodology,apollo 2026-04-01
Kyle's AI consulting identity is VALIS — named after Philip K. Dick's Vast Active Living Intelligence System (1981). The name works on multiple layers: PKD's concept of a vast intelligence cutting through noise and delivering signal directly maps exactly to what Kyle does for clients. In 2026, VALIS aged into relevance — it's not science fiction anymore. The name filters clients naturally: people who recognize the reference are the right clients. People who don't still find it short, punchy, and memorable. valis,consulting,identity,brand 2026-04-01
Lazlo web app scaffolded (Mar 31, 2026): Flask + SQLite app at ~/Desktop/Kyle/app/. DB at ~/Desktop/Kyle/lazlo.db — outside the app folder intentionally so todo_migrate.py and other tools can share it. Seeded with Kyle and Heather. Mail config tab is live. Tasks/Scripts/Brain tabs are stubs for phases 2-4. Run with: cd ~/Desktop/Kyle/app && python3 app.py → http://localhost:5000 — 2026-03-31
The stack test cycle audit (Mar 31, 2026): built the org, hired the team, stood up Citadel, ran the pipelines. What landed: KyBrain, morning email, FDN knowledge base, the team structure. What changed: Citadel scope, SSH/FTR priority. What needs to die: org complexity sized for higher SSH/FTR ambition, Citadel expansion mindset. The Stoic filter: what do I actually use vs. what's in the realm of sophistry? — 2026-03-31
Tool stack settled (Mar 31, 2026): Kate for markdown and quick file edits. Vim for terminal work (scripts, config files). Obsidian retired — vault archived, KyBrain is the mind. VSCodium shelved — option anxiety with no payoff at current stage. Re-enters only if LSP matters in a real codebase. Three rounds of vimtutor — trust the investment. — 2026-03-31