Notes on agents, AI products, and the way software gets built when the tooling writes back.
The Mom Test with agents: pick positive and negative user cohorts from data, prep each call with a one-page user card, ask questions that get facts instead of compliments, and turn interviews, support tickets and cancel reasons into one ledger of user facts. Copyable prompts included.
App store optimization taught from a real 20-market pass: how a store search turns into an install, how to read a keyword, where the words go on the App Store and Google Play, what makes people tap, what twenty countries taught me, and a 20-minute weekly routine with Sonar.
Agents are fast, but they don't know your company. It becomes queryable when its product data and the decisions its people already made sit where agents can read them. How each person gets an agent that works like them, what that looks like in real skills from a product company and an app studio, the team layer that shares it, and the file structures to copy. With prompts to paste.
When agents plan work, pick it up and report back, your tracker stops being a board you click and becomes a to-do list that also has to remember. How to write issues an agent can continue from, where people keep the decisions, how to close work with evidence, and the guards that keep agents from filling your board with noise. With prompts to paste.
A routine is a Claude Code session that starts on its own, on a schedule, from an API call or from a GitHub event, and works while your laptop is closed. What to give it, which connections and secrets it can use, where a person stays in the loop, and the mistakes a real one made in its first weeks. With prompts to paste.
Some jobs are too big for one Claude conversation, and some answers shouldn't be checked by the agent that produced them. You can split the work across agents with one job each, including one whose only job is to disagree. Three setups you can ask for today, with prompts to paste.
If you keep typing continue while Claude Code stalls halfway, you can hand it a finish line instead: a result it has to show, a check against the real thing, and a rule for when to stop or ask you. Three loops you can set up today, with prompts to paste.
Once you can make Claude finish a task and split a job across agents, the next step is work that repeats: the weekly review, the support queue, the monthly report. Two questions tell you which setup a job needs. A ready-to-run recurring review, with prompts to paste.
Most people re-explain their work to Claude Code every session. You can write it down once instead: skills that hold what you do, decide and know, pick themselves when a task fits, work with each other, and sync to your whole team. A ready-to-use setup, with a prompt to paste at every step.
It has been immensely helpful for me, it is easy to use and to the point.
My father had the internet. I have an agent in a terminal. I think we are feeling the same thing, one generation apart, and I wanted to write down what that feels like from the inside.
An agent harness is the program that wraps a model, which only reads and writes text. I read DeepSeek's open-source one and kept a checklist.
RuFlo bets that coordinating many agents is a systems problem, not a prompt problem. I think the bet is right, and I also think most teams reach for it two months early.
Every agent I run authenticates with a human's credentials or a shared service account. The audit log records the wrong actor, and I don't think I'm unusual.
Context rot is a memory problem, not a model problem. Here is the spec-driven loop I use to work around it as a solo builder.
The Ralph loop wraps Claude Code in a script that re-runs it until your tests and checks pass. How it works, the guardrails, and what it changes.
The people getting the best work out of coding agents aren't writing better prompts. They're writing SOUL.md, STYLE.md, SKILL.md. Here's the mechanism I think is actually doing the work, and the part of the idea I don't buy.
Jesse Vincent's Superpowers turns prompts into files with rules. I opened the copy installed on my machine and read every line.
Gemini 3 Flash scores 78.0% on SWE-bench Verified vs 76.2% for Pro, at 75% lower input price. Why distillation explains it, and the trade-offs.
Claude Code has the best agentic experience but costs ~$1k/month for heavy use. I built a Claude Skills-compatible system for Antigravity that makes AI coding preferences portable across platforms.
Cross-platform frameworks solved expensive labor. When AI makes labor cheap, the abstraction overhead remains but the savings vanish. Native is coming back.