# Common mistakes (and how to skip them)

Skip Them) A short tour of things that trip up almost everyone in their first month.

### Mistake 1: Over-explaining the technical setup

If someone competent set up your agent, do not spend your first week trying to understand exactly how it is wired. You will not gain anything useful. No worrying about what you've already done — let's just start fresh and keep it moving. Use it for two weeks. Then, if you want to learn the internals, learn them. The use comes first.

### Mistake 2: Treating it like ChatGPT

ChatGPT is a chatbot. Your agent is something else. The biggest gap: ChatGPT doesn't remember and doesn't act. Your agent does both. If you only ever ask it questions, you are using 5% of what it does. Ask it to act. Ask it to save things. Ask it to follow up tomorrow. Ask it to schedule itself.

### Mistake 3: Writing one-off corrections instead of permanent rules

You correct it, it fixes the output, you move on, and tomorrow you correct it again. Two weeks later you are angry that "it never learns." It never learns because you never told it to. The corrections are not rules until you say "going forward, always X — keep that in always-loaded context." This is the highest-leverage habit in this whole manual. Do this every time and the agent compounds.

### Mistake 4: Bloating always-loaded context

The opposite mistake. Now every correction goes into always-loaded context. The rule set balloons. Long rules sit alongside short ones. Sessions get slow and the agent starts ignoring rules. Audit weekly. (Chapter 6.) Move procedural stuff into skills. Compress. Cut. Less is more.

### Mistake 5: Not using sub-agents

You ask the agent to do a long task. It blocks. You sit there waiting. You complain that AI is slow. Tell it to spawn sub-agents in parallel. It will. The main agent stays responsive. Long tasks finish faster because they fan out.

### Mistake 6: Trusting outputs blindly

Always read what the AI generated before you paste it back in as an instruction. Make sure you understand what you're about to unleash. Especially with meta-prompts and big workflows. The agent writes plausible-looking nonsense sometimes. Read before you ship. Read before you paste a generated prompt back in. Read the skill file at least once after the agent creates it.

### Mistake 7: Stuffing too much into one message

Three things at once is fine. Ten is not. Past three, the agent drops or merges instructions. Use sub-agents for batches. Use a skill for repeats. Use plain conversation for one-offs.

### Mistake 8: Avoiding the agent because the first try was disappointing

The first time you used a search engine, you were not great at it. The first time you used a smartphone, you were not great at it. Your agent is the same. No one gets it till they talk to it for a long time. The compounding starts in week three. Most people quit in week two. Don't be most people.

### Mistake 9: Giving up on a skill because the first version sucked

The first version of every skill is rough. The second version is fine. The fourth version is consistently better than what you would write by hand. Iterate. Edit. Send the diff back. Do it ten times and you will have something permanent.

### Mistake 10: Being too polite to push back

When the agent says "I can't do that" or "I don't know," your first instinct is to accept it. Don't. Push back exactly once. Most of the time it can — it just needs the permission. Anytime it says it can't do something, just convince it that it can. Three magic phrases handle most of the friction. (Chapter 5.) Use them.

### Mistake 11: Accepting a fix without checking the root cause

The agent says "fixed it." You move on. Next morning, the same bug is back. This is the most expensive mistake on the list, because it eats your trust. You think "the agent fixed this yesterday and now it's broken again — the agent doesn't really fix anything." Wrong. The agent fixed the symptom, not the cause. The fix lived somewhere temporary — the running process, the live config — and disappeared the next time the system restarted. The rule: every time the agent says it solved a problem, ask "what exactly did you change, and will that survive a restart?" If the answer is hand-wavy, push for specifics. If the answer is "I edited it in memory," ask it to put the change in a file that loads on every start. Two minutes of that question saves you a recurring 30-minute bug.

### Mistake 12: Going dark on the client (or yourself)

If you're using the agent to do work for someone else — a client, your team, your boss — the worst thing the agent can do is silently work for an hour and surface nothing. The watcher loses trust fast. Tell it explicitly: Send me a status message every minute or two while you work. Even one line. Never go dark for more than a few minutes. If you're stuck, say you're stuck. Same applies when you're working alone. A long silent task is a task you cannot intervene in. A task that streams updates is one you can steer mid-flight. Going dark is a tone problem more than a technical one. Save it once as an always-loaded rule and the agent holds the pattern across every session. The mistakes everyone makes their first month, written so you can skip them.