Patience and Humility

AI is an incredible force multiplier for execution—for making things happen—but a dangerous substitute for your vision.
— Tiago Forte

One of my 2026 goals is to invest in AI literacy so I can better understand its usefulness in both my personal and professional life. To that end, I applied AI to a real workflow in May as part of a paid AI course I completed in June. This post is the third and final milestone of that goal: publishing one concrete use case or lesson learned.

I consumed quite a bit of AI content over the past eight months, much of it skewed to the extreme. I joined two Circle communities. I read books and articles. I watched YouTube and listened to podcasts. I collaborated with people who thought I was an expert and people who thought I was a novice—sometimes in the same week. I talked with people who believe there is almost nothing AI can’t do and with someone who was so emotionally triggered by AI they would not read a document I drafted with it. I still consider myself an AI hobbyist, but I’m now also a certified AI operator.

I’m also much more comfortable with my opinion about AI and how I use it. It is not a person; it is a tool. As a tool, it is great at some tasks and not at others. I am by no means an expert, but I now understand my position in the “information supply chain” between true experts and true novices.

My preferred LLM

It is an understatement to say that AI is a rapidly evolving technology. One frontier lab releases a new feature that gains popularity, and the others quickly develop their own. I experienced some FOMO as I learned about Claude Cowork, but then ChatGPT released its own version. When we refer to “AI” in the singular, we lose the nuance of the specific tool used:

1. Platform: I use OpenAI—not because of any philosophical or ethical disagreement with Anthropic or Google, but because of the current architecture of Apple Intelligence. I like the simplicity of one platform, one subscription, and native integration into my devices. There are certainly nuanced differences between platforms, but they aren’t important enough across the major ones for me.
2. Harness: I use ChatGPT Work. Harness engineering is the source of many recent AI gains. The models do continue to improve, but Codex/Work and Code/Cowork can orchestrate multiple agents over longer periods of time than a simple chat, and interact with local files. I didn’t think I needed this capability until I tried it, but it changed how I use AI.
3. Interface: I use the macOS/iOS apps. However, I do shift to a web browser when editing custom GPTs. Another neat feature of using ChatGPT Work is that I can run it on files streamed from my Google Drive to my Mac via my iPhone.
4. Model: I use 5.6 Sol with medium thinking, but don’t have a strong opinion. I haven’t done comparison testing and never worry about running out of tokens; I just chose what was recommended by Ethan Mollick.

My focus on context

I created a Venn diagram about a year ago while trying to understand how model, personal, project, and chat context overlap. I sent it to Tiago Forte and perhaps I inspired him, but his persistent, project, and perishable framework is much better.

I personalized my ChatGPT (and Notion) profile using what he calls a master prompt. AI doesn’t know me, so this document constantly reminds it of my personal identity and life context; professional role, expertise, and work context; thinking, learning, and communication style; values, goals, and operating principles; and relationships, boundaries, and sensitive areas. I then curate context for active projects, with a narrow focus on purpose, status, people, constraints, decisions, open questions, and voice for each. Lastly, I follow Rachel Woods’ MASTER method for individual prompts. This ensures the model has the specific context, success criteria, and examples for this interaction.

Why go through all this? Repeatable, high-quality outputs.

Persistent and project context are baked into how I use AI personally but, alas, have to be manually added at work. Sometimes I forget one or the other, which creates an opportunity for comparison. My persistent context connects meeting participants to their role in meeting summaries and without it, the AI doesn’t understand the plethora of acronyms and jargon used. My project context is used for quick leadership updates and without it, we revisit settled issues and don’t focus on the harder open questions.

My team is rapidly improving our organizational knowledge, aligning stakeholders, and making measurable progress.

My use case

My AI operator certification consisted of nine self-paced modules and a capstone project over twelve weeks. I was originally drawn to the focus on platform-agnostic playbooks (“own the playbook, rent the tech”) that an LLM (or college intern) could follow step-by-step. I then learned that a playbook is actually the artifact that follows from visualizing and designing a process before applying any AI or automation.

I pitched improving my company’s monthly status report to the owner and executive vice president. From my view, the current process did not integrate input across employees, did not compare input to our statement of work, and the reports had diverged from the agreed-upon format over time. Pages of information were provided each month, but the labor involved provided little value to either party.

Leadership approval took about 30 minutes. Interviews with subject matter experts totaled about two hours; in between, I iterated on the playbook over the course of a few weeks. The playbook would run in a tightly controlled domain, so the APIs and automation tools I had learned were not an option. Nonetheless, I built three custom agents that were ready for testing.

They all failed.

Sometimes the agent got stuck on a step, and sometimes it ran without stopping between steps as instructed. Sometimes the agent truncated the data tables and sometimes the agent got too creative with the report. With my limited permissions, however, there was no way to debug the agents. I could only use trial and error.

I tried simple instructions with the playbook attached. I copied and pasted the playbook into the instructions. I tried an orchestration agent with subagents for each step. Each failed in the same way.

I finally tried a standalone chat—and it worked.

The testing issues were a setback with the supervisors who would use the playbooks. Senior leadership understood the business development value and were willing to work through the bumps. The supervisors, however, already viewed the new process as more prescriptive and difficult than it was worth. Why were we fixing something not broken? Why does the AI make little changes I don’t want? Why does it then make other changes when I try to fix it?

As I often relearn, it was time to refocus on the user experience.

I developed a visual job aid that included expectations and troubleshooting. I helped employees sign up for half-day AI/ML training. I held one-on-one walkthroughs. Ever so slowly, the new process gained traction. I started to receive useful feedback on the output, instead of the concept, that I quickly incorporated into the playbook.

One supervisor commented that they had no idea how far beyond requirements their team was performing. Another used the new report to separately highlight a major strategic initiative their team accomplished. The customer was delighted with the new summaries that aligned with the agreed-upon format.

Conclusion

AI did not fix our process. It helped us see it more clearly, test a better version, and produce a more useful output with less effort once the people involved understood what problem we were solving. AI is powerful for execution, dangerous as a substitute for judgment.

Here are my three takeaways from completing this journey:

1. Wrestle with the work. AI-assisted work still requires judgment, context, and validation. Find the right amount of botsitting: enough oversight to ensure quality, not so much that the tool becomes a burden.
2. Own the output. You would not blame spellcheck for a bad sentence. Do not blame AI for a bad product. If your name is on it, the work is yours.
3. Use AI, but do not lead with it. Some people are curious, some are skeptical, and some are honestly afraid. The goal is not to prove you used AI. The goal is to produce better work.