What should you learn after you have mastered basic prompting?

After basic prompting, the next skills to learn are how AI models actually work under the hood: what a token is, the practical difference between open and closed models, why a context window runs out mid-task and how to protect against a model you rely on changing or disappearing. These sit underneath every prompt you will ever write.

Why Prompting Skill Stops Being The Limit

Once a prompt reliably gets a good result, prompting stops being the thing holding anyone back. The next bottleneck sits one level down, in why a model behaves the way it does rather than in how a request is worded. Someone who only knows how to phrase a good prompt can fix a bad result by trial and error. Someone who understands the mechanics underneath can usually see the actual cause before they even try again.

A phrasing trick stops helping the moment the model changes, because it was built around one model’s particular habits. Knowing why a model behaves the way it does keeps paying off, because that understanding carries over to every model that comes after it.

Tokens And Why They Quietly Run Out

A token is the small chunk of text an AI model actually reads and writes in, usually a few characters or part of a word rather than a whole one. Every model has a fixed budget of tokens it can hold in mind at once, called its context window, covering the conversation so far, any documents pasted in and the reply still being generated.

Once that budget runs out, older parts of a conversation start getting dropped or summarised, quietly and without much warning. A long document pasted in at the start of a session, followed by a long back-and-forth, is one of the most common ways a model appears to “forget” an instruction it was given only minutes earlier. Knowing that this is a token limit, not a lapse in the model’s attention, changes how the work gets structured: shorter sessions, restated instructions and documents trimmed to what is actually needed.

Open Models Versus Closed Models

A closed model is owned and run entirely by one company, accessed only through its own app or API, with its exact workings kept private. An open model has its underlying weights published, so it can be run, inspected and modified by anyone with the technical means to do so.

The practical difference is control. A closed model can be faster to start with and often produces stronger results out of the box, but it can also change behaviour, raise its price or be discontinued entirely without much notice, because none of that is in the user’s hands. An open model trades some of that convenience for the ability to keep running a specific version indefinitely, on infrastructure the business actually controls.

Model Availability Risk And How To Protect Against It

Model availability risk is the exposure a business takes on when a single closed model becomes load-bearing for a real process, without a plan for what happens if that model changes or disappears. It is a real and growing risk, not a hypothetical one. OpenAI’s own published deprecation records show the pattern clearly: it gave developers roughly three months’ notice before retiring the o1-preview model in July 2025, then around six months’ notice before retiring DALL-E 2 and DALL-E 3 in May 2026. A further batch of GPT-5 and o3 model snapshots is already scheduled for retirement in December 2026, on the same roughly six months’ notice. The notice period moves. The retirement does not stop coming.

Protecting against it does not mean avoiding closed models altogether. It means treating any AI-dependent process the same way a business would treat a single supplier: documenting exactly what the process needs from a model, testing an alternative occasionally and never building something so specific to one model’s quirks that switching becomes a rebuild rather than a swap.

What This Means For A Team, Not Just One Person

Model availability risk rarely sits with one person, because the AI-dependent process it threatens is usually shared: a workflow one team member built that others now rely on without knowing which model sits underneath it or what happens if that model changes. If only the person who built it understands the tokens and the model choice behind it, the whole team inherits a risk none of them can see coming.

A team that shares this understanding treats a model change as a planned swap rather than an emergency rebuild, because someone already tested the alternative and documented what the process actually needs. That difference, tested versus untested, is usually invisible until the day a model is retired and the team without a plan loses days it did not need to lose.

What To Do Next

Take one AI-dependent task already running regularly and ask three questions about it: how many tokens does a typical run actually use, would it survive on an open model if the current one changed overnight and does the person running it understand why it works, not only that it does. Any question with an uncertain answer is exactly where to spend the next round of learning.

If you’re watching all of this happen and wondering how to make AI work for your business, book a call with me today – I train and consult businesses and teams to use AI with confidence, skill and intention.

This is one of several questions answered as part of the full pillar article, What Separates Skilled AI Users From Confident Ones.

Reference: OpenAI’s published API deprecation records, developers.openai.com/api/docs/deprecations, including the o1-preview retirement (announced April 2025, retired July 2025), the DALL-E 2 and DALL-E 3 retirement (announced November 2025, retired May 2026) and the scheduled GPT-5 and o3 snapshot retirement (announced June 2026, due December 2026). This material also forms Session 3 of Guy Galloway’s AI Training Programme, covering tokens, open and closed models and model availability risk, detailed at craftingyourintelligence.com