A working
notebook
Experience, know-how, and thoughts from solo dev with AI.
I almost gave up, then reopened myWiki from the requirements with Grok
Facing the mountain left by unmanned work, I chose repair over rebuild
I asked to fix the title and the body vanished
Why a wiki you cannot hand-edit needs rewrite validation and undo
I tried to run it like second and cost broke me
Why I dropped full-analysis summarize, and merge only asks title and summary before stitching bodies
The body was sitting in the database
Why I ran one unmanned full loop, then dropped Postgres body columns for a two-layer md layout
Index and keywords came out worse than I expected
Splitting vertical classification from cross filters, then capping counts all the way into merge
Chat felt slow because of wait time, not unit cost
We cut full document-map injection and set 4k/10k caps, but short answers, search, and synthesis are still unfinished
I only saw the mountain of fixes after shipping unmanned
myWiki finished via dispatch after requirements and screens — cost of skipping mid-flight review
Why only now for us
First day on Grok after Claude quota friction — speed and back-and-forth without waiting
Why I pin a model on each todo
How dispatch injects model tags into subagents, and why longer todos are design not clutter
Subagents start blank
When a todo line and detail block are the whole interface — including a 140-turn failure
The Actual Point of Using AI - Loop Engineering
Where Agent Development Starts - Prompt Engineering
Making It So You Don't Have to Say It - Harness Engineering
Just Start Already
Two Token-Saving Tricks I Found the Hard Way
Set a lower default and lock each skill's model: two fixes that cut my cron token usage.
A Skill Is Just a Well-Made Prompt
Constraining behavior, reuse, auto-invocation, easy authoring — the four properties of Claude Code skills
The Code I Wrote by Hand Was Already in Claude Code
As prompt engineering gets built into the tools, what a builder should hold onto is an understanding of how LLMs behave
Every Session, the LLM Meets You for the First Time - Context Engineering
Context engineering and CLAUDE.md — how to ship project context to an LLM that has no memory
You Don't Need to Code to Build an App
What non-developers actually need to get right, from a 10-year developer who builds with natural language
AI Writes All My Code — So Why Do I Still Blog?
The real reason I blog was never the content. It was the review.
I Automated a Spot That Shouldn't Have Been Automated
Safe automation is knowing what to leave to humans
I Ditched the Claude Deploy Skill for GitHub Actions — and My Token Usage Hit Zero
How a self-hosted runner rabbit hole led to appleboy/ssh-action for home server deploys
How to Write Instructions for an AI Agent
Why "do B2" fails: handing your agent the full context with the build-plan format
From Idea to Build Plan: Automating 8 Design Docs with AI
Revealing the docs-spec workflow validated through myWiki's design process
Why You Can't Close Your MacBook While Running an Agent — How I Started Using Amphetamine
Why You Don't Have to Read Every Raw File to Get an Answer — 4 Levels of Wiki Index Priority
[Blog 3] Jumping from Seed Straight to Draft Kept Missing the Point
Why a two-stage seed→draft pipeline drifts off-target, and how a brief checkpoint fixed it
The Well Wasn't Dry
How I built blog-mine when daily blogging left my queue empty
Every Skill Pipeline Converged to the Same Shape
I built separate pipelines for blogging, apps, and tasks. They all ended up identical.
Why I Dropped Codex and Wired Claude CLI Into Hermes
Getting daily AI news via cron without spending a single credit
Why digest Wasn't Growing My Wiki
How 1,861 raw files went unfound by digest, and how a second-layer index cache fixed it.
Skills Grow As You Use Them
A real record of building the triage skill through conversation
Put Logic on the Model and Duplication Disappears
Avoid copy-pasting update rules to three places: gather them into a single model method.
Routers Do HTTP, Logic Lives in Services: Separating Layers
Why I split router, service, and repository responsibilities one line each
Common Modules and Base Classes to End Code Duplication
Around the third domain, lift the repeating code up into common
Your Folder Structure Is Your Design
Why I organize by domain instead of function, and what changes when you do
Data Access Through Repositories Only
Lock your queries inside repositories and there's only one place left to change
Everyone Was New to ORMs, So I Wrote the Guide
A 17-section development guide written to stop copy-paste duplication
Validate at Every Boundary: Schemas and Layered Defense
Don't ship your models as-is; validate at each layer with a different responsibility
Only Catch the Exceptions You Can Handle
Catch only what you can recover from, and let unrecoverable exceptions rise to the caller.
I Automated Todoist with MCP, Starting with Sorting
From triage to execution: how I automated Todoist by connecting it to Claude Code via MCP
The First To-Do App That Had Everything I Wanted
Why I chose Todoist: global quick add, free tier, and an MCP that makes it programmable
The Justice League Once Lived in My Stock App
BaseSqlModel, BaseRepository, a 15-chapter guide, and making the AI stick to one pattern.
The AI Was Fine. I Just Didn't Know the Tools.
Five automation mistakes in my stock app and what finally changed how I work.
Storing Is Only Half the Job: I Built a Search Skill Too
If the wiki has the answer, shouldn't it be able to find it when asked?
The Hard Part Wasn't Collecting, It Was Digesting
How I set up Claude Code to run the second wiki: digest rules, pin, pipeline, and hooks.
If It Doesn't Pile Up Automatically, You Won't Collect It
CLI only worked at my desk, and remote kept trying to edit my wiki, so I moved to Hermes
[Blog 2] AI Hallucinated From Examples I Intentionally Made
I removed the markers, and AI filled them with things that looked like my own story.
[Blog 1] Markdown Only, No Database: A Blog AI Manages End-to-End
The technical architecture of slowloop lab: Astro, text files, and a second brain wiki
[Blog 5] What It Means for AI to Manage Your Whole Blog
With just text files, AI can read and rewrite your entire blog without a database
[Blog 4] How Conversation Becomes a Post: One Sentence to Published
AI-prompted posts always sound generic. I built a skill pipeline to fix that.
The Person Building Alone with AI Agents
Without a second brain, I don't think I'd have gotten this far
How I Built My Mom's Online Art Museum in 6 Hours
Planning on the commute, AI that started without asking, and how we got it shipped.
I Read the Terms of Service After Building the Whole Thing
A stock portfolio app dev story: KIS API regulation traps and a last-minute pivot
Before Building the App, I Tried CLI and Obsidian First
I fed 130 scattered chat histories into a CLI to validate the hypothesis in a single day.
I Read a Stock Book and Decided to Build an App Instead
Zero stock knowledge, built it anyway. Back after a year, and done in two weeks.
Turning an Idle PC into a Zero-Cost Server
Hosting costs hurt, I turned on a home server, and hit SSL certs and domain config walls.
A Token-Burning Experiment Became the Foundation of Atlas
AI generates, humans evaluate: what the template experiment left behind.
AI Stamped Out Templates, But I Couldn't Judge Them
Auto-generation worked. Self-verification didn't: the gap that blocks autonomous loops.
Tokens Were Going to Waste, So I Made a Screen Every Week
The starting point of a design experiment born from leftover tokens
I Couldn't Pick the Design, So I Let Users Vote
Outsourcing the design call through voting, then piping that signal back into generation.
Neither of Us Speaks English, But the Museum Is Multilingual
Running a multilingual site without a translator: filling the DB with an LLM pipeline
I Didn't Build It All at Once
Why I split requirements, mockups, and code into separate phases, and what that saved.
Five Years Later, I Built My Mom's Art Museum Right
The failed shop that taught me the difference between selling and showing
Why I Break Every Big Task into Phases
How I structure every large task with AI: phases, gates, and why both matter.
Why You Should Not Integrate Your Tools
Context evaporates at every handoff. The answer wasn't tool integration, it was a spine.
It Didn't Exist, So I Started Building It Myself
Dozens of PKM apps, nothing fit, so after an OAuth headache I decided to build my own.
I Built a Second Brain So My AI Chats Wouldn't Disappear
What a month of AI chats left behind, and why I split raw files from wiki pages.
AI Wrote It, But It's My Story
How raw conversations stored in a wiki become blog posts
My Planning AI and Coding AI Were Working in Silos
The problem of context breaking between agents, and how I solved it with a second brain
4 Control Levers When Your Agent Won't Listen
Why it's not the model's fault, and what each of the 4 control levers actually does
I Quit Documentation Three Times. Then /docs-init Fixed It.
Complex structure kills documentation. The survivors: just index.md and log.md
Why "Build Me an App" Is Actually the Slow Way
One session = one item, dependency-aware splitting is the mechanism that creates speed
Decide First, Then Code: docs-spec Forces Specs Up Front
Even solo developers can avoid the "I have to rip this out" moment
3 Painful Lessons Deploying a Stock App at Home
iptime CAA block, DB port conflict, stale checkout: things you only learn the hard way.
Running Multiple Apps on One Home Server
How I run a stock app, art gallery, and bot on the same server, full setup shared.
Building a Loop Where Discord Messages Land in Your Wiki
How ideas thrown from your phone pile up in a wiki, built with a single ops repo.
How I Run hermes and Claude Code Together
hermes for chat, Claude Code for code and wiki: the split I found running both at home.