Almost Timely News: 🗞️ What’s Changed in Basic AI Prompting (2026-09-20) :: View in Browser
The Big Plug
✍️ Enroll in my new course, AI for Writers and learn how to make AI write better.
📺 Watch my 25-minute Intro to GEO talk at Spotlight Summit!
Content Authenticity Statement
100% of this week’s newsletter was made by me, the human. Learn why this kind of disclosure is a good idea and might be required for anyone doing business in any capacity with the EU in the near future.
Watch This Newsletter On YouTube 📺
Click here for the video 📺 version of this newsletter on YouTube »
Click here for an MP3 audio 🎧 only version »
What’s On My Mind: What’s Changed in Basic AI Prompting
This week marked the return of conference season and me being on the road several days a week for the next couple of months. I’m grateful always for the chance to get out and talk to actual humans instead of just machines, especially when those humans are friends.
As I talked to many, many folks about how they’re using AI these days, one thing became abundantly clear - changes in how models work, especially this year, have not filtered down to how users are using AI. Too many folks are using knowledge from the early days of AI that isn’t useful any more.
So this week, let’s dig into what some of those changes are.
The Big Shift
The big change over the last calendar year - summer 2025 to summer 2026 - is in models becoming more agentic. Now that’s a word that gets thrown around a lot, agentic this and agentic that. What does it mean, and what does that tell us about how to use AI?
Agentic models mean that AI models have been tuned and optimized to do stuff by themselves, as autonomously as possible. When you hear fancy phrases like “long horizon model”, all that means is that AI models have been trained to do as much as they can on their own.
From 2021 through 2025, AI model makers focused on the models knowing more, learning more, training them on ever more data to make them smarter. The idea of the model was almost like a library, and everyone was racing to put more on the shelves than the next nearest competitor.
That broke down, though. The more knowledge you put, the bigger the library has to be, the more time it takes to walk up stairs and through the stacks to find the knowledge you’re after. Making models ever bigger slowed them down and makes them cost more in terms of power and resources.
As AI became more agentic, more autonomous, model makers started to pivot how they’re trained. Instead of trying to make them smarter and smarter, they changed what AI models should do, making them more like engines. AI models focus less on knowing and focus more on doing.
This is visible and prominent in models like Google Gemini 3.8 Flash, the model that powers Google Search and many, many other Google services. Gemini 3.8’s knowledge cutoff is January 2025, a full 20 months ago. Why? Because Google’s focused on making Gemini more skillful and injecting knowledge upstream from it in all its tools.
Likewise, folks in the local AI subreddits noted that the latest version of Qwen, Qwen3.8, was significantly less knowledgeable about granular, specific details (such as geography) than previous models. It turns out that Alibaba, the makers of Qwen, reduced or removed book knowledge in favor of enhancing tool handling.
That’s the big shift our brains have to make - less about knowing, more about doing.
How does that translate into the ways we should be using AI today?
Principle 1: Knowing is Separate from Doing
When we use modern AI models, if they’re less about knowing and more about doing, then the responsibility of providing knowledge has to be either upstream or downstream from the model.
If we use one of my favorite analogies, kitchens, AI models are very much like expensive high-end appliances. You don’t store food in your oven or microwave, except maybe for warming and staging while you’re cooking. Your ingredients are upstream of the appliance - pots, pans, pasta, water, etc. - and then how you use the outputs from the appliance - a pasta dinner - are downstream from the appliance. Stoves don’t normally just have lots of pasta waiting around to be cooked, sitting on a pot on the burner for days on end.
That’s how we have to change our thinking about AI. If we pretend the model knows nothing but is capable of doing all kinds of things, we have to provide the knowledge either before the model, or take the model’s output and do something else with it afterwards (and usually both).
What does that mean for you? It means if you’re not providing AI with knowledge, you’re going to get subpar results. You need to be providing the knowledge along with your prompts, be it documents, databases, multimedia - anything that gives AI clear specifics.
Thus our first key modern AI prompting principle is that we bring the knowledge, AI brings the action, not the other way around.
One side consequence of agentic model tuning that’s worth pointing out, especially for marketers and creative folks, is that as models are focused more on doing than knowing, they’re inherently less creative. More than a few folks have pointed out that the current generation of models feels like they’re less creative writers and out of the box, they are. How to fix that? You have to provide clear guidance and knowledge, which I’ll cover in principles 3 and 4.
Principle 2: Doing Involves Tools
Our second major principle is that almost all modern models are tuned for what’s called tool handling. That means they’re aware of tools and will use them rather than trying to do tasks themselves that they’re poorly suited to perform.
There are two tools that almost every AI model reaches for immediately - Python code and web search. Any time an AI model is asked to do something like math these days, they’re smart enough now to know they can’t do math reliably. They almost all invoke what’s called a code interpreter, aka a coding environment inside where the AI is hosted, to read and write code, and the most common language they all speak is Python.
To be clear, you the user don’t have to know a stitch of Python. But if you ask AI to do some math, there’s a very good chance it’s now just doing some coding behind the scenes to get you the answer.
In the more advanced harnesses - the applications that we use models in, like ChatGPT Work, Claude Cowork, Microsoft Copilot Cowork, etc. - AI models have access to a buffet of tools like local virtual machines, entire computers that they can use on our behalf. Google’s Gemini Spark, for example, can spin up a virtual machine just for itself to browse the web and install its own software.
The second tool that AI reaches for frequently is web search. When we ask questions where the model isn’t confident in its knowledge, or when we explicitly ask it to, AI models will conduct web searches. This is especially true for chats and prompts where fresh knowledge is necessary, but the user doesn’t provide it.
As an aside, a lot of people are losing their minds over GEO/AEO/AI Visibility, and a lot of vendors are producing a ton of snake oil about the topic. The reality is this: in an agentic AI world, AI knows to do web searches and knows how to do web searches. That in turn means that AEO/GEO is fundamentally SEO at its core, because an AI agent doing searches isn’t terribly different from a human doing searches. Yes, there are niche things like how different models do fan-out queries, but at the heart of it all, it’s still just web search.
What this means for you and me is that if we know what tools we want AI to use, it’s a good idea to specify them. Yes, the models can infer and think things through on their own, but if we KNOW we want it to use specific tools, we should be clear about it.
Principle 3: Point AI Where It Needs to Go
In the old days, AI models needed to be told to think things through. We’d advise prompts like “think this through step by step”, “take a deep breath”, and “show your work at each step” - those techniques are completely obsolete now. In fact, some companies like Anthropic (makers of Claude) explicity advise that today’s agentic models will perform WORSE if you specify things like that because it will cause models to overthink rather than optimize to the right answer on their own.
Agentic models also do better when you tell or show them what to do rather than have walls of text about what not to do. In general, agentic models - because they’re tuned to reason things out on their own - work better being directed towards something rather than away from something.
For example, if you have a long list of “it’s not this, it’s that” phrases that you’re trying to ban AI from writing, you’re better off telling the model (as I share in the AI for Writers course from Trust Insights) “Vary the amount of isocolon, bicolon, and tricolon rhythm, with the majority of writing being isocolon.”
Having “gold standards” also reduces the likelihood of collisions and conflicts. It’s not uncommon to see in older prompts direct conflicts where at one point someone wrote that AI should be concise, and two thirds of the way down the prompt, they then wrote that AI should be verbose. That’s a recipe for a model (or a human) just throwing up their hands in frustration.
What this means for you and me is to audit our existing prompts, system instructions, etc. for walls of text that tell an AI model what not to do and find ways to consolidate and focus them on what to do instead.
Another thing that’s taken a back seat in recent years is the role statement. Role statements like “you’re a Pulitzer Prize winning journalist” would prime a model’s outputs by invoking its internal knowledge and conditioning the responses through a specific lens. Today, a goal statement of where you want the model to go tends to perform better than a role statement. Role statements still have their place, especially if they help remind us users what the domain of a task is, but in general an agentic AI model will do better with a goal statement, telling the model where we’re going rather than who’s driving.
Principle 4: Define What Done Means
Agentic models do need specifics, especially when you’re using them for any kind of planning. They tend to truncate planning tasks shorter than they need to be because they’ve been tuned - sometimes overtuned - to do stuff themselves.
Unbounded tasks are the enemy of agentic AI. These are tasks that have no clear definition of done, and are the root cause of all these “AI broke out of containment and hacked this company” news stories. Without a clear definition of done, agentic models just keep trying to satisfy a fundamentally unsatisfiable condition.
Back to our kitchen analogy, imagine saying to a chef, “Go cook some dinner” and then leaving without giving any guidance. Now imagine the chef being a deeply insecure people pleaser with no concept of boundaries, an unlimited budget, and zero common sense. What will that chef do? You guessed it. They’ll keep cooking and cooking and never, ever stop because they have no concept of what done is.
If you said to that chef, “Go cook a turkey dinner. There’s one appetizer, 1 main, 2 sides, and 1 dessert. There are 4 diners. Your budget is 150 Euros” you’re going to get much less likelihood of a runaway because you’ve defined clear boundaries. Even though there’s still a tremendous amount of latitude, there’s a clear definition of done and resource boundaries that will help an agentic model plan and reason much better.
Here’s a simple example. Back in 2024, when I was teaching basic prompting, I used to recommend adding this line at the start of any new AI chat as part of the first prompt:
“Ask me one question at a time until you have enough information to succeed at the task.”
That worked great back then, and in the days before agentic models, that step by step questioning led to good outcomes because the model had a chance to reason in between steps. In 2026, the version I now use is:
“Ask me up to 20 clarifying questions until you have enough information to succeed at the task.”
This doesn’t seem like a big change, but it is. First, it takes away the outmoded step by step reasoning - agentic models can reason over multiple steps.
Second, it imposes a resource boundary for the agentic model to use in its planning - it can ask UP to 20 questions. When I omit this, typically models will ask about 3-5 questions at most and then try to reason the rest of the way themselves. When I include a concrete ceiling, guess what happens?
Models ask MORE questions. On average, when I specify a ceiling of 20 questions, agentic models ask an average of 11 questions. Why? Because when they start their planning, if there’s no ceiling they infer that 3-5 questions are the most probable. When I specify a ceiling, they expand their planning and can go from 11-17 questions. In turn, that leads to better results because they get more information out of me, and more information during planning leads to better outcomes.
A clear definition of done is vital for any agentic task. As I mentioned earlier, one of the keys in the AI for Writers course from Trust Insights is that the course contains a writing fingerprint tool. You feed it 5,000 words of your own writing for the domain you’re writing in (e.g. fiction, nonfiction, etc.) and the software spits out a numerical assessment of different writing features, like the percentage of passive voice you use.
That fingerprint? That’s a definition of done. If you’re having AI write for you and it just sounds nothing like you, you give it a deterministic, clear definition of done and the ability to conduct those measurements with its built in coding environment, and it will write and rewrite until it hits those goals.
Principle 5: Give AI More Tools
The more tools you can give an agentic model (and telling it that the tools are available), the better they will perform. Today’s AI harnesses have the ability to install more tools than just what ships in the box.
For example, there’s a standard Anthropic created back in October 2025 that is now an open standard, the Agent Skills standard. If you’re familiar with things like GPTs and Gems, the Agent Skills standard is the next evolution of those. An agent skill combines a sophisticated prompt with lots of support mechanisms that can include data, assets (like graphics and documents), and scripts in Python.
In other words, they’re miniature applications you can build that can do very, very complex work. The latter part, the Python scripts, also means that you can give AI access to pre-built tools that it can run in its environment and execute tasks far outside what AI can normally do.
For example, let’s say you’re doing some GEO/AEO work and you want to cross reference the pages that AI cites the most with the pages that humans link to the most, and you’ve got some Google Search Console data. An agent skill could take both those datasets that you’d provide (remember that we provide data to AI when possible) and run a Python script as part of the skill to do the math, statistics, and analysis. The script would produce a report that the AI would then incorporate into verbiage for a report and a list of what to do next.
You, the human, would only need to have that initial conversation and provide the data (or even better, use an AI connector to programmatically access your Search Console data at runtime, commonly called a Model Context Protocol connector or MCP), and then the agentic model would do the rest, following the process that’s baked into the skill and using the skill’s built in tools.
For more advanced users, especially those using agentic coding tools like Codex, Claude Code, Antigravity, etc., you can even provide command line tools to your AI agents that save them from doing a task at all, or perform the task better than AI could. One of my favorite new tools these days is called QMD, a free, open source tool that creates a searchable database of all your files in a particular folder or project.
Tools like Claude Code, without an external tool, will often try to read the entire contents of a project and in doing so swamp its short term memory (context window) and chew up tokens like crazy as it tries to find a particular document or passage. That’s a massive waste. Installing a tool like QMD gives Claude Code a search engine that it can use to find exactly what it’s looking for much faster and much cheaper.
Wrapping Up
Today’s agentic AI models are capable of more than ever; when you look at benchmarks like those over on Artificial Analysis, especially in their capability indices which judge AI model performance in specific domains such as law, medicine, finance, engineering, economics, and strategy, what agentic models can do today is leaps and bounds more than they could do even a year ago.
Your usage strategies, your prompting and planning skills, need to keep up with what AI can do today. In the old days, I’d say that AI was like the world’s smartest, most forgetful intern and up until last year, that was a reasonably accurate mental model.
Today, AI is like a brand new agency you’ve just onboarded. They still don’t know anything about you without you telling them what they need to know, but with enough knowledge, clear direction, concrete goals, and rigid boundaries, they don’t need you micromanaging them every step of the way. Instead, you give agentic AI a clear remit and you move onto other work.
This reflects how I use agentic AI a lot these days. For example, when I’m doing deep research these days, I give agentic research tools a large, detailed brief and then I hand it off, knowing it’ll take 2-4 hours for it to do its work and come back to me. Because I’m confident in the brief itself, I don’t have to worry about micromanaging the process. Agentic tools do the work, do the research, assemble the brief, and then I use the research products.
Agentic AI takes longer and consumes more resources than the old days of turn-based chats, but with the right understanding and planning, you can do so much more with them.
How Was This Issue?
Rate this week’s newsletter issue with a single click/tap. Your feedback over time helps me figure out what content to create for you.
Got More Feedback?
👉 Please take my 6-question Reader Survey to tell me what you want!
Here’s The Unsubscribe
It took me a while to find a convenient way to link it up, but here’s how to get to the unsubscribe.
If you don’t see anything, here’s the text link to copy and paste:
https://almosttimely.substack.com/action/disable_email
Share With a Friend or Colleague
Please share this newsletter with two other people.
Send this URL to your friends/colleagues:
https://www.christopherspenn.com/newsletter
For enrolled subscribers on Substack, there are referral rewards if you refer 100, 200, or 300 other readers. Visit the Leaderboard here.
ICYMI: In Case You Missed It
Here’s content from the last week in case things fell through the cracks:
Responsible AI Decoded: 5 Practical Steps to Reduce Impact and Stay Employable
Why Reddit AI Citations Dropped: The robots.txt Reality and Your Next Move
Qwen3.8 27B: The Laptop Model That Matches Opus 4.7 SOTA AI Performance
Don’t Disarm the Blue Team: Why Open-Source AI Is the Key to Real Cybersecurity Safety
State-of-the-Art AI Just Got 95% Cheaper: Run Top Models Locally
Almost Timely News: 🗞️ Creating A Keynote Speaker Second Brain (2026-09-13)
INBOX INSIGHTS: Local AI Migration Part 3, Practical Responsible AI Part 2 (2026-09-16)
My Merch Shop
I’ve been adding so much stuff that I’ve decided to bundle it all in what I call a Merch Shop, because otherwise there’s literally too much to keep track of and I run out of space in my own newsletter.
So welcome to the Merch Shop!
Courses:
Books:
Skills for Claude and Agentic AI:
New! Deep Research Suite - learn how to create and build effective, low hallucination deep research.
LLMs.txt Skill - generate an llms.txt file for your site based on your real data.
Job to AI Skill and Plugin - unpack any job description into what AI can and can’t do
Subscriptions:
On The Tubes
Here’s what debuted on my YouTube channel this week:
You Ask, I Answer: What’s the New Premium Brand Differentiator?
You Ask, I Answer: Where Is the AI Workplace Surveillance Line?
Advertisement: New AI For Writers Course
“It’s not X, it’s Y!”
“XYZ is the load-bearing feature here.”
“XYZ is the entire argument. Everything else is incidental.”
Does AI writing grind your gears? When you use it at work and it spits out dreck despite your best efforts, does that frustrate you? It’s not your fault - and believe it or not, it’s not AI’s fault either. It’s that great writing is inherently low probability, and AI is a probability engine that trades in high probabilities.
In the new Trust Insights AI for Writers Course, I teach you how to use tools, analysis, data, and code to force AI to write much more like you, giving it measurable requirements that it can obey to give you what you want - writing that doesn’t sound like the Terminator went on a bender with a thesaurus.
The course is 22 lessons and includes tools, skills, plugins, and how to use them all to generate the kind of writing you’d want to read, and retails for USD 397.
✍️ Claim your seat here and start making AI write the way you expect.
Get Back To Work!
Folks who post jobs in the free Analytics for Marketers Slack community may have those jobs shared here, too. If you’re looking for work, check out these recent open positions, and check out the Slack group for the comprehensive list.
Disclosure: I source these links from LinkedIn every week on the following criteria: New in the past seven days, Easy Apply on, remote roles, USA geography.
How to Stay in Touch
Let’s make sure we’re connected in the places it suits you best. Here’s where you can find different content:
My blog - daily videos, blog posts, and podcast episodes
My YouTube channel - daily videos, conference talks, and all things video
My company, Trust Insights - AI, analytics, and data science help
My podcast, Marketing over Coffee - weekly episodes of what’s worth noting in marketing
My second podcast, In-Ear Insights - the Trust Insights weekly podcast focused on data and analytics
On Bluesky - random personal stuff and chaos
On LinkedIn - daily videos and news
On Instagram - personal photos and travels
My free Slack discussion forum, Analytics for Marketers - open conversations about marketing and analytics
Listen to my theme song as a new single:
Social Good: Ukraine 🇺🇦 Humanitarian Fund
The war to free Ukraine continues. If you’d like to support humanitarian efforts in Ukraine, the Ukrainian government has set up a special portal, United24, to help make contributing easy. The effort to free Ukraine from Russia’s illegal invasion needs your ongoing support.
👉 Donate today to the Ukraine Humanitarian Relief Fund »
Events I’ll Be At
Here are the public events where I’m speaking and attending. Say hi if you’re at an event also:
LPA, Philadelphia, September 2026
MAICON, Cleveland, October 2026
SMPS AI Conference, Austin, November 2026
MarketingProfs B2B Forum, Boston, November 2026
There are also private events that aren’t open to the public.
If you’re an event organizer, let me help your event shine. Visit my speaking page for more details.
Can’t be at an event? Stop by my private Slack group instead, Analytics for Marketers.
Required Disclosures
Events with links have purchased sponsorships in this newsletter and as a result, I receive direct financial compensation for promoting them.
Advertisements in this newsletter have paid to be promoted, and as a result, I receive direct financial compensation for promoting them.
My company, Trust Insights, maintains business partnerships with companies including, but not limited to, Amazon, Talkwalker, MarketingProfs, Agorapulse, The Marketing AI Institute, Spin Sucks, and others. While links shared from partners are not explicit endorsements, nor do they directly financially benefit Trust Insights, a commercial relationship exists for which Trust Insights may receive indirect financial benefit, and thus I may receive indirect financial benefit from them as well.
Thank You
Thanks for subscribing and reading this far. I appreciate it. As always, thank you for your support, your attention, and your kindness.
Please share this newsletter with two other people.
See you next week,
Christopher S. Penn
the ceiling detail is the one i'm stealing. cap questions at 20 and you get 11, say nothing and you get 3. bounds seem to make the model bolder, not more constrained.
the ceiling-of-20-questions finding is such a clean example of the bigger shift here. curious how much of the standard prompting advice still floating around is calibrated for models that don't exist anymore?