Last week I caught myself about to make a one-time fix.
I was building Multi with AI agents. One of them added a high-priority task to the product backlog. It landed at the bottom of the list. The backlog sorted by recency, so the newest task always sat last regardless of importance.
My instinct was to drag it to the top. Five seconds. Done.
But tomorrow another agent adds another task and it lands at the bottom again. I’m back to manually sorting. Each fix is a 1x return.
So I wrote a protocol at the top of the backlog file. Priority tiers. Ordering rules. Instructions for evaluating new items against existing ones. Now when any agent touches that backlog, it reads the protocol and places the task where it belongs.
I did the work once. It compounds across sessions.
I’ve started seeing this across my whole workflow. I work with AI agents on writing, product development, research and distribution. Each conversation is a chance to build something disposable or something reusable. The gap between the two over time is enormous.
Three examples from the last month.
Skills over instructions. I write for my newsletter and publish through WordPress. The first time I asked an agent to publish a post, I gave it step-by-step instructions. The second time, I wrote those instructions into a skill file. Now any agent in any session can take a draft from my editor to a live post on the site: formatted, linked and indexed. I explained the process once. It runs on autopilot.
Process design over process execution. My daily input processing used to be a conversation. “Here’s my voice note, extract the drafts, update my context.” Now it’s a single trigger. Voice notes go in, finished drafts come out, my context stays current. The entire pipeline runs without me repeating a word.
Structured backlogs over ad hoc lists. The backlog story I opened with goes deeper than sorting. The protocol forces each agent to compare new items against existing ones in the same tier, check for overlap and re-evaluate whether current priorities still hold. The backlog gets smarter each time something is added to it.
Each time you solve a problem with an AI agent, ask: am I making a withdrawal or a deposit?
- A withdrawal is a one-time action. You get the result. The knowledge stays in the conversation and dies when the session ends.
- A deposit is a reusable component. You get the result and you get it again next time without repeating yourself.
Skills, protocols, structured documents, management rules embedded in files agents read. These are deposits. They accumulate.
Raw conversations, one-off instructions, manual corrections. These are withdrawals. They evaporate.
The compounding math is brutal. If you make 50 decisions a week about how agents operate and each one is a withdrawal, you make 50 decisions again next week. If 20% of those become deposits, by month three you’re making 10 decisions a week. The system absorbed the other 40.
Most people using AI right now are making withdrawals. They ask a question, get an answer, move on. The next session starts from zero. The AI gets more capable with each model upgrade but the system around it stays the same.
The leverage is in the assets you build around the model. A skill file is a tiny asset. A backlog protocol is a tiny asset. A style guide agents consult before writing is a tiny asset. Stack enough of them and the system operates with less friction each week. The context got richer. The model stayed the same.
That’s the compounding return.
🔮
What’s one process you repeat with AI that could become a reusable component?
