Mistake #1: Running Your Agent Without a Mission
For the first few days, I was treating my agent like a search engine. Ask a question, get an answer, close the chat. What I didn’t realise is that an agent without direction is just an expensive chatbot. It can’t self-direct. It can’t prioritise. It waits for you to tell it what to do next.
The fix is embarrassingly simple: give your agent a mission statement. One sentence at the top of your SOUL.md or mission control file that defines what the agent exists to do. Now every prompt includes purpose. The agent isn’t just completing tasks — it’s working toward something.
Mistake #2: Not Setting Up Persistence First
Day 2, I had a fully configured agent. Great personality, custom briefing prompts, specific instructions. Then I restarted the session. Gone. Everything I’d told it — vanished. It came back as a blank slate.
Memory is not automatic. You have to build it intentionally. My agent now opens every session by reading four files: memory.md, mission.md, context.md, and voice.md. The first 10 seconds of every session are silent — it’s reading. The next 10 years of every session are better because of those 10 seconds.
But here’s what I learned on Day 5: persistence isn’t just about remembering facts. It’s about remembering rules. I wrote the same instruction in four different files and the agent still forgot to follow it under load. Persistence of information is solved. Persistence of behaviour is the harder problem.
Mistake #3: Ignoring Skills
Skills are the mechanism that expands what your agent can actually do. The base agent is like someone who graduated university — smart, articulate, capable of reasoning. Skills are the specialist training you bolt on top.
Mistake #4: Paying Claude Opus Prices for Everything
I covered the full cost story in my Day 1 journal — the short version is $43 in two weeks on Opus, which would have been $0.86 on MiniMax M2.5. Fifty times cheaper for nearly identical quality.
The caveat: cheaper models need better prompts. When I switched, the first results were shallow. The fix was specifying everything — seven sections, exact sentence counts, required data sources, output format. More upfront work. But the prompts are reusable, and the cost dropped by 95%.
The 3 Cron Jobs You Need This Week
Set these up before you build anything else. They’re not features. They’re the foundation.