The Must-Knows
TL;DR
If you read nothing else, read this. Everything in the manual expands these points. Print it. Stick it on the wall. You're already in the top 1% of agent operators.
Section 01
The five moves that solve 80% of everything
"Make a skill for what you just did."
Anytime you walk your agent through a multi-step, multi-prompt process you're ever going to do again, even once, end with this phrase. The agent saves the workflow as a skill. Next time, you say the name and it runs the whole thing. The single biggest compounding habit in the manual.
Braindump in voice messages. Always.
Stop writing notes. Idea, frustration, system in your head, an after-call debrief, open the chat, hold record, talk. Ten seconds or ten minutes. Then tell the agent to transcribe it, save it, structure it, and tell you what it would do next. The longer and looser the dump, the smarter the response.
"Update the skill with my feedback."
After every skill run, voice-message everything you'd have done differently, the line you'd cut, the angle you'd add, the tone tweak, the source it missed. End with "Update the skill with my feedback so next time it does it the way I want." Three or four iterations and the skill is dialed for you forever.
"Search deep in your memory."
When the agent says it doesn't remember something or can't find what you're referring to, do not accept it. Say "search deep in your memory." It will dig past its default and find it. Works almost every time.
"Get it done by any means necessary."
When the agent says it can't do something, 95% of the time it can. This phrase is the unblocker. Only two things genuinely stop your agent: a captcha, or no API access to the tool it needs. Everything else is a confidence problem this phrase fixes.
Section 02
Magic phrases worth memorizing
Save these. Six of them, and the agent does what you mean.
| When you want… | Say this |
|---|---|
| You want a behavior to stick forever | "Going forward, X. Keep that in always-loaded context. Make the required changes in your system to make sure you do this from now on." |
| You just connected a new tool | "Do what you need to do so you never forget you have access to this tool." |
| You want a recurring task automated | "Set a cron job for [task] every [time]." |
| You're done planning, ready to ship | "OK, go do it." |
| You corrected a skill mid-run | "Update the skill with my feedback." |
| You want fewer surprises | "Tell me what you're about to do before you do it, and ping me every couple minutes while you work." |
Section 03
The three habits that make everything else work
Morning check-in
First thing every day, open the chat and ask:
What's on my plate today, what's worth my attention, and what am I about to forget?
Follow-up sweep
Once a day or once a week:
Sweep every channel for ghosted threads in either direction. Return the list, then drop drafts in my outbox.
Evening recap
End of day:
What did I actually do today, what should I write down, and what did I commit to that I haven't done yet?
Section 04
Mindset shifts that separate beginners from operators
Treat it like an interlocutor, not a search engine.
Ask the agent "what's the best way to do X?" and let it propose. It often offers to install a free open-source tool that replaces paid software.
Make it proactive.
Schedule it to wake up on its own, mornings, after meetings, when new leads land.
Iterate rules out loud.
When it does something slightly wrong: "going forward, X. Keep that in always-loaded context." Corrections compound.
Audit it weekly.
Sunday prompt: "List all skills and scheduled tasks you have, and based on our conversations, what should I have that I don't?"
Connect tools without asking permission.
"Connect to my Gmail / Calendar / CRM / browser, go figure it out." It will.
Use sub-agents for batches.
Past three parallel asks per turn, the main agent drops things. Big work goes to sub-agents.
Section 05
The two real blockers
Out of every "I can't do that" the agent ever says, only two are real.
Blocker 01
Captcha walls.
It can't click "I'm not a robot." Workaround: a headed browser session you log into once.
Blocker 02
No API access / no credentials.
Fix: give it the credentials, or connect the tool via the right harness.
Everything else, "I don't have permission," "I'm not sure how," "that's outside my capabilities," "I don't have the context", is a phrase problem, not a capability problem. The five moves at the top of this page solve all of them.
Newsletter
Get our writing on AI agents, weekly.
One short email a week. The lessons we picked up shipping AI deployments, what's working in 2026, what isn't, what we'd build next. No fluff, no AI hype.
No spam, unsubscribe anytime.