Your agent skips steps because your checklist is a suggestion in a prompt, not a gate in a system. An agent has no sense of a step being unfinished, only a sense of the work looking plausibly done. Move the checklist into a file the agent has to read, write to, and prove against before it can continue, and the skipping mostly stops.
Why Does My AI Agent Skip Steps When the Prompt Looks Fine?
One of my agents provisioned a client account last month without verifying billing details first. Step 3 of 9, marked "verify billing," got summarized into "billing looks standard" and skipped entirely. The fix took four minutes. The cleanup took a morning.
That is the pattern I see over and over. It is almost never the model being lazy. It is the checklist being written in a way that only a human who already knows the job could follow. "Review the call and update the CRM" is three steps wearing one sentence. The agent picks the easiest interpretation, usually "update the CRM with a summary," and the review never happens.
There is a second reason, and it is the bigger one. Nothing in the run marks a step as complete. A checklist sitting in a prompt is text competing with the rest of the context window. Twenty turns in, your nine steps are buried under transcripts and tool outputs. What is left is a vague goal and a strong pull toward the finish line. Agents are really good at finishing. They are not naturally good at finishing in order.
What Does a Checklist Look Like When an Agent Can Actually Follow It?
Four rules, and the first one does most of the work.
One verb per line. "Pull the call transcript" is a step. "Review and summarize the call" is a wish. If you cannot picture the single action, the agent cannot either.
Every step names an input and an artifact. The transcript goes in, a file called objections.md comes out. If a step produces nothing you can look at, it is not a step, it is a mood.
Every step has a done test the agent can check without asking you. "Artifact exists and lists at least three named objections" is checkable. "Thoroughly considered" is not.
No branching inside a step. Branching happens at gates, and gates get their own line.
I keep these in a plain markdown file in the repo, one file per workflow. In Claude Code or Cursor it sits in the working directory, so the agent reads it on every turn without being reminded. For no-code AI agents, the same idea works as an Airtable table with a status column, read by n8n or Make at the top of each run. The tool does not matter. The file living outside the prompt is what matters.
I have 12 of these files now. The content pipeline one has 9 steps and takes an agent about 14 minutes to run clean.
How Do I Force the Agent to Prove Each Step Before It Moves On?
Two mechanisms: a run log and a gate.
The run log is a file the agent appends to after every step. Step number, timestamp, artifact path, one line on what changed. Then the first instruction of the next step is "read run-log.md and confirm step 3 has an artifact." Not assume it is done. Confirm it. If the artifact is not there, the agent stops and tells me.
That single instruction cut skipped steps in my own workflows by more than half. It is annoying to write the first time and it is the cheapest reliability you will ever buy.
Gates are hard stops where a human or a separate check signs off. I only use two: a facts gate before anything goes to a customer, and a money gate before anything spends. Everything else is agent-judged, because a checklist with nine human approvals is not a workflow, it is a queue.
In n8n, the gate is an If node on the status column that routes to a Slack message. In Claude Projects, it is a pinned instruction plus me actually reading the log. It is boring and it works.
The rule I care about most: the agent has to fail loudly rather than skip quietly. Skipping quietly is the most expensive thing an agent does, because you find out three weeks later when a client gets invoiced for the wrong plan.
Can I Still Let the Agent Improvise?
Yes, and you should. The goal was never a robot that follows instructions like a fax machine. Improvisation inside a step is where the useful stuff lives. Improvisation across steps is how you get a mess.
So I write scope corridors into the checklist. Something like: "You may choose how to source the transcript, from the call tool, the calendar invite, or the recording folder. You may not skip having one." The agent gets freedom on method and zero freedom on sequence.
What's becoming clearer to me is that this is where most people get the tradeoff backwards. They write a loose prompt, get a creative output that skipped the fact check, then clamp down so hard the agent becomes useless. The answer is a tight sequence with loose execution inside each box. My best writing agent has a voice I never specified and I love it. It has never skipped a fact check, because the fact check is a gate, not a request.
How Do I Stop the Checklist From Drifting Across Dozens of Runs?
One source of truth, read at run time, never copied into a prompt.
The failure mode is familiar. You paste the checklist into an agent's instructions, then you improve it, then you have nine versions across nine agents and no idea which one is current. That is checklist drift, and it is why a workflow that worked in March quietly stopped working in June.
I keep all 12 checklists in one folder, and the 45 agents on my team read from that folder at the top of every run. When I change a step, every agent picks it up on the next run. No retraining, no re-pasting, no conversation about which version is live.
That is what an AI agent operating system is at the least glamorous level: a shared checklist library, a state file, and a couple of gates. None of it demos well. All of it is the difference between an agent you trust with a client and an agent you babysit.
FAQ
How many steps should a single checklist have?
Between 5 and 12. Anything under five is really a prompt, so the agent fills the gaps with guesses. Anything over twelve should split into two workflows with a handoff artifact between them. My longest is 11 steps and I have been trying to break it in half for a month.
Doesn't a stricter checklist make my agent slower?
Per run, yes, by maybe 10 to 20 percent. Per week, no, and this is the part people miss. When I added run logs and gates, my average number of reruns per task dropped from about 2.4 to 1.1. I will take a slower first pass over doing everything twice.
Do I need to code this, or can I use no-code AI agents?
No-code is fine. Airtable for the checklist, n8n for the gates, a Claude or GPT step for the actual work, and you have everything I described. What you cannot skip is the state file and the gate. If your tool can write a row and read it back at the start of the next run, you have enough.
None of this is hard. It is just not the part anyone demos, because demos show the magic and nobody shows the run log. I wrote the full playbook for this. How to Build Your Own AI Agent Operating System walks you through the exact architecture I use, step by step. You can get it at a.mastermindshq.business/ai-os-book.