Stage 04 — Agentic
The door is open. Walk through it.
// The Journey
Every practitioner moves through the same four stages. Most stop at the first. This page walks all of them — keep scrolling to pass each door, and enter any stage to see the real prompts, real builds, and real systems behind it.
001 – 004 · keep scrolling
001 / Chat
The beginning. The most powerful thing most people ever do with it — and it's only the entrance. The answer is rarely limited by the model; it's limited by the prompt.
002 / Project
It becomes a collaborator. Standing instructions, durable knowledge, one place it can reuse — and the work compounds instead of resetting every time. This is where the real work starts.
003 / Automate
Workflows run without you. You design the system; it executes on a schedule — briefs written, dashboards rebuilt, pipelines fed. The clearest marker of this stage: something useful happens overnight.
004 / Agentic
The system acts, decides, and iterates on your behalf. You set the direction; it finds the path, takes real actions, verifies its own work, and stops when it should. This is the current edge.
// The log so far
// Thinking
For most of human history, the tools to solve complex problems were expensive, proprietary, and out of reach. You hired a firm. You bought the software. You paid for the integration. You settled for the feature set someone else decided you needed.
AI changed the math. The gate is gone. So is the ceiling.
Not for everyone equally — but for anyone with taste, judgment, and the willingness to start.
The tools you used to pay for and resent — the ones built for someone else's use case, with someone else's color scheme, missing the three features you actually needed — you can build yourself now. Exactly how you want them.
A personal finance system that thinks the way you think. A news filter tuned to the beats you actually care about. A search layer across your own messages that works better than anything the phone ships with.
Built by you. For you. No subscription. No compromises. No waiting for the roadmap.
The most interesting thing happening right now isn't the technology. It's what happens after one person in a room starts actually using it.
A conversation changes direction. Someone asks: wait, could it do that for me? Then they're building something. Then someone else asks them the same question.
AI spreads the way good ideas always have — person to person, through demonstration, through curiosity. The best thing you can do is start. The best thing we can do is share.
The role that makes sense to me has always been the same: take something complicated and make it approachable enough that someone who doesn't know what they don't know can still take their first steps.
That's what this site is. A log of what I've built, what I've learned, and what's possible when you stop waiting and start doing. The door is open.
// Work
The door is open. Walk through it.
There's a writing section here.
It's gated.
If you're supposed to be in it, you know how to find me.
Stage 01 — Chat
Chat is the first door. You type, it answers. For most people this is the whole relationship with AI — and they judge the entire technology by the quality of the answers they get back.
But the answer is rarely limited by the model. It's limited by the prompt. Most people type one-liners — and a one-liner is why they get a one-liner back.
// The anatomy of a good prompt
The same question, asked four ways, returns four different answers. Add each component and watch a throwaway prompt sharpen into an expert one.
[Goal] What do you need? [Context] What's the situation? [Source] What should it reference? [Expectations] Format, tone, audience, length?
The reusable template — Goal · Context · Source · Expectations. Structure once, reuse forever.
// Learned, not given
This structure isn't innate. For me it's the residue of more than eleven thousand prompts — my own slow shift from asking to instructing. The prompts below are mine, lightly anonymized, typos left in.
PowerShell automation
Drafting a prompt for another AI
Structure the prompt, and Chat stops being a toy. Then you give it memory, context, and standing goals — and you're at the next door.
Next — Stage 02 · Project
Stage 02 — Project
Chat is disposable. You ask, it answers, and the context evaporates the moment you open a new thread.
The leap to Projects is one move: give the model standing context, a memory, and goals — kept in one place it can reuse. Do that, and it stops being Google on steroids and it becomes a real collaborator.
// How it happened
The clearest example from my own history: writing email. Click any step to see the actual prompts (typos left in).
"Please revise this email to be more concise and professional:"
That one line was one of 344 near-identical prompts — the same request for clarity, professionalism, and conciseness, re-typed by hand, every time.
"I want the new project to be focused specifically on EMAILs for any situation"
Instead of re-typing the ask, I moved it into a ChatGPT Project — standing context the model reaches for on its own.
Its custom instructions are grounded in thinkers I trust on emotionally intelligent communication — Simon Sinek, Marshall Rosenberg, Peter Levine, and Kim Scott. They fix a structure for every message — open with why it matters, lay out context without blame, name next steps and owners, close with collaboration — and spell out the phrases to reach for ("Would you be open to…") and the ones to avoid ("They dropped the ball…", "You need to…").
"Speak from service, lead with clarity, and write with the calm confidence of someone already trusted."
My real, unchanged "Emails for Any Situation" project instructions — exactly as they live in the Project today.
📬 Project Instructions – Emails for Any Situation This project is designed to help craft or revise emails for any situation—whether internal or external, urgent or reflective, operational or strategic. Every message should reflect leadership-level presence and emotionally intelligent communication. 🧭 Communication Objectives Emails should consistently: Build trust and confidence in your leadership Offer clarity, foresight, and reliable follow-up Maintain a tone that is calm, respectful, supportive, and non-reactive Align with service-first leadership and relational collaboration ✨ Guiding Thinkers & Influences 1. Simon Sinek – Start With Why / Leaders Eat Last Open emails with why the message matters Frame communication around purpose, long-term alignment, and team cohesion Reinforce shared goals, trust, and responsibility Speak with vision and clarity—not just reporting status 2. Marshall Rosenberg – Living Nonviolent Communication Observe without judgment (e.g., "This has been pending since 5/29" vs. "They've been slow") Express needs clearly and calmly Use empathetic, emotionally attuned language Invite collaboration, not control 3. Peter Levine – Waking the Tiger / Biodynamic Leadership Influence Communicate from a grounded, composed presence Avoid urgency, defensiveness, or reactivity in tone Give the message space to breathe—clear, uncluttered, and spacious Acknowledge complexity without amplifying tension 4. Kim Scott – Radical Candor Balance clarity with care Speak directly, but with empathy and respect Offer honest feedback in a way that deepens relationships 5. Craniosacral & Somatic Leadership Principles Favor emotionally regulated language Write with attuned pacing and presence Contain emotional intensity with structure and calm energy ✍️ Email Structure Template Subject: Clear and focused Opening: Start with why the message matters Context: Present the facts neutrally and without blame Next Steps: Clearly state actions, owners, and timelines Tone Touch: Close with collaboration, warmth, and forward motion 🗣️ Tone Guidelines Use language that is: Calm – "Here's what I've reviewed so far…" Kind – "Appreciate your time and visibility on this." Trustworthy – "I've confirmed all vendor updates and next steps." Forward-looking – "This positions us well for Q3 systems readiness." Empathetic – "I understand the concern—here's how I've addressed it." ✅ Phrases to Use "To support our broader initiative…" "Would you be open to…" "Appreciate the patience as we worked through…" "Let's align this with our upcoming goals for…" "Here's what I've done, and what I recommend next…" 🚫 Phrases to Avoid "They dropped the ball…" "You need to…" "It's not my responsibility…" "So you don't get blindsided…" "I already explained this…" 🧩 Optional Future Add-ons Let me know if you'd like to include: Email categories (e.g., offers, status updates, apologies, praise, escalations, stakeholder updates) Audience-based variations (e.g., engineers, VPs, vendors, admins) Templates for commonly repeated messages 🧘 Final Principle Speak from service, lead with clarity, and write with the calm confidence of someone already trusted.
"when I say 'revise:' you know I want an email reviewed per the instructions"
All that depth, triggered by a single shorthand. The Project remembers, so I don't have to.
// The endpoint
Follow the "give it context" instinct far enough — years in, well past those first chats — and your context stops living inside any one app. It becomes plain text files any model can read: a portable "second brain."
"I am building MD files for an AI second brain. … Draft all ten MD files …"
The goal, in plain terms: a durable record I can return to — point at any day and recover what I thought and worked on.
Context, memory, goals — and the work compounds instead of resetting every time.
Next — Stage 03 · Automate
Stage 03 — Automate
At Stage 02 you have a collaborator. At Stage 03 you have infrastructure. The shift is from prompting to designing — pipelines, schedules, agents that execute whether or not you're at the keyboard.
The clearest marker of this stage: something useful happens while you sleep.
// What I built that I couldn't before
Before Claude and Cowork, I had ideas I couldn't execute. Not because the idea was too complex — because the execution required skills I hadn't built: web development, data pipelines, access control, deployment. The gap between I want this and this exists was measured in months of learning I didn't have time for.
The School Selector closed that gap.
Where it starts
The first step wasn't building anything — it was figuring out what to build. Claude ran the research: school quality metrics, what data actually matters for fit, how to weight competing priorities. Hours of reading compressed into a structured brief I could act on. I went from scattered notes to a clear scope without writing a line of code.
Where Claude becomes the builder
Some people call it "vibe coding" — and yeah, that's what it is. Cowork built it in plain language I could follow on the front end, while keeping the logic in the backend modern, clean, and sophisticated. Data model, scoring logic, comparison interface. I directed; it executed.
Where it ships
It lives at a real URL behind Cloudflare Access. Mine — not a screenshot, not a prototype. A tool I actually use, that I could not have built twelve months ago. Gated for privacy, web-available so it works on any device. Stage 03 is when the thing you built starts doing its job without you managing it.
// What runs while I'm not looking
The other half of Stage 03 is artifacts that auto-update — dashboards, pipelines, and intelligence feeds that regenerate on a schedule so the right information is always current. These aren't static documents. They're living tools.
Stage 03 is where you stop spending time and start building systems that spend it for you. The next step is making those systems autonomous enough to decide what to do with it.
Next — Stage 04 · Agentic
Stage 04 — Agentic
Stage 03 gives you systems that run. Stage 04 gives you systems that decide. The difference is judgment: not executing a schedule, but reading live state, choosing an action, taking it on real infrastructure without asking — and then checking whether it worked.
The clearest marker of this stage: something changes in the physical world because software decided it should, while I wasn't watching — and the software knew whether it was right.
// The loop
A network-remediation agent runs unattended on my home network — four background services, always on, surviving reboots. It doesn't wait to be asked. When something breaks, it moves through the same loop every time.
What it watches
Polls eleven devices and cameras every 60 seconds, and listens for real-time events on an open webhook — two signals, one picture of what's actually up. It watches its own eyes, too: if the poller goes stale, a self-check pages me. A dead watcher can't report its own death.
How it commits
A failure isn't acted on the instant it appears. It has to hold — two consecutive confirmations, or ~75 seconds — before the agent commits, so a device that just blinked is ignored. Each device type maps to its own runbook; anything without one is logged and escalated, never guessed at.
What it does
When an access point drops, the agent power-cycles the exact switch port feeding it — a real action on live gear, no approval step, no human in the loop. It refuses to touch uplink and infrastructure ports, and rate-limits itself to two actions per device per six hours.
How it checks itself
It doesn't assume the fix worked. It waits and watches for the device to actually come back — up to four minutes for a clean reboot — and confirms against the system that would really know. If recovery doesn't hold, it downgrades its own verdict to escalate and backs off.
// What makes it agentic, not just automatic
Stage 04 isn't a system that does more. It's a system you trust to decide — because it shows its work, respects its own limits, and stops when it should. That's the door at the end of the corridor. It's already open.
The current edge — Stage 04 · Agentic