Four Of These Nine Skills Need Something You Do Not Have Yet
“9 AI Skills to Master in 2026.” A clean nine-box grid: prompt engineering, workflow automation, AI agents, RAG, fine-tuning, multimodal, video generation, tool stacking, LLM evaluation. Three named tools under each. The caption asks the obvious question: “Which one are you mastering first?”
For five of them that question makes sense. For the other four it does not, and the reason has nothing to do with difficulty.
The source card — “9 AI Skills to Master in 2026”, BANI Academy. Nine skills, three tools each, with this footer: “Tool names and features change often. Check each tool’s official site before you use it. Skills take practice, so results vary.”
Four of the nine are not things you learn
They are things you do to something you already have. Read the card’s own descriptions:
- RAG — “Connect LLMs to your own data such as PDFs, documents, or databases.” Without the documents, there is nothing to connect. You can follow a tutorial that indexes somebody else’s sample PDFs, and at the end of it you will not have built anything you use.
- Fine-tuning and custom GPTs — “Train or adapt existing models for a specific task or brand voice.” A brand voice is a dataset: a few hundred examples of writing that already sounds like you. If you do not have that body of work, there is nothing to fine-tune on.
- LLM evaluation and management — “Track and improve your AI’s performance by measuring accuracy, cost, and response quality.” This one is explicit. It measures a system. No system, nothing to measure.
- Workflow automation — “Link multiple apps together… scheduling, data entry, or content delivery.” This needs apps that already hold your work, and a repetitive task that already annoys you. Automating a process you do not yet run is a demo.
None of those four is hard in the way a language is hard. They are gated. You can read about them all week and be no further forward, because the missing ingredient is not knowledge, it is a thing you own.
Diagram: what each skill needs before you can start
| Skill | What you must already have | Start today? | What “learning” it looks like without that |
|---|---|---|---|
| 1. Prompt engineering | Nothing. A question you care about | Yes | There is no without. This is the floor |
| 6. Multimodal AI | Nothing. An image or a recording | Yes | Fine — the skill is mostly knowing what the tools accept |
| 7. AI video generation | An idea and a script | Yes | Fine, though the script is the hard part |
| 3. AI agents | A multi-step process you already do by hand | Partly | Building an agent for a workflow you have never run manually |
| 8. Tool stacking | Apps that already hold your work | Partly | Connecting two empty accounts together |
| 2. Workflow automation | A repetitive task that already annoys you | No | A Zap that fires on nothing |
| 4. RAG | Your own documents, and a reason to query them | No | Indexing a sample PDF from a tutorial |
| 5. Fine-tuning | Hundreds of examples of the voice you want | No | Fine-tuning on a public dataset you do not care about |
| 9. LLM evaluation | A system in production producing output | No | Reading dashboards with no traffic in them |
So the order is not one to nine
If four skills are gated on an asset, the sequence writes itself, and it is not the order on the card.
Make something first. One small thing you actually use: a weekly report you write by hand, a set of notes you keep re-reading, a video you publish. That artefact is the key that unlocks the other four. The moment you have a pile of your own documents, RAG stops being a tutorial. The moment you have published forty pieces in your own voice, fine-tuning has a dataset. The moment something runs on a schedule, evaluation has a number to watch.
Which inverts the usual advice. The four “advanced” skills are not harder. They arrive later because their inputs arrive later. Nobody is held back from RAG by the difficulty of RAG.
We checked the tools, because the card says to
Credit first: the footer is honest — “Tool names and features change often. Check each tool’s official site before you use it.” So we did, on a sample, and two things came back worth knowing.
AutoGen. Named as one of three tools for building AI agents. Its own repository README opens with a caution box:
“AutoGen is now in maintenance mode. It will not receive new features or enhancements and is community managed going forward. New users should start with Microsoft Agent Framework.”
The status badge next to the project’s title reads maintenance mode and links to its successor. The repository is not archived and still has 61,000 stars, so it is alive — but on a card headed 2026, a tool whose own front page tells new users to start somewhere else is the one to know about before you spend a weekend on it.
GPT-4. Listed under Multimodal AI. That model is several generations back, and the card is titled “to Master in 2026”. The skill is right; the example is a few years stale.
This is the same exercise we ran on a card full of GitHub numbers a few weeks ago, where checking went the other way and the card turned out to be under-stating. Checking is cheap, and it is the instruction this card gives you itself.
And two boxes are one skill
Compare them side by side. Box 2, AI Workflow Automation: “Link multiple apps together using automation tools so AI handles repetitive tasks.” Box 8, AI Tool Stacking: “Combine different AI and productivity apps into one connected system that automates workflows end to end.”
That is the same job described twice — and Zapier is listed in both boxes. So the grid is nine squares and eight skills, which matters only because of the question underneath it: if you are choosing one to start with, two of your nine options are the same choice.
All nine, with what we found
| Skill | Tools named | Note |
|---|---|---|
| 1. Prompt engineering | ChatGPT, Claude, Gemini | The one genuinely open to everybody on day one |
| 2. AI workflow automation | Zapier, Make, n8n | Same job as box 8, and Zapier appears in both |
| 3. AI agents | Crew AI, LangGraph, AutoGen | AutoGen is in maintenance mode per its own README |
| 4. RAG | LangChain, Vectara, LlamaIndex | Needs your documents before any of these mean anything |
| 5. Fine-tuning & custom GPTs | OpenAI GPT Builder, Hugging Face, Cohere | Two very different jobs in one box: a custom GPT is configuration, fine-tuning is training |
| 6. Multimodal AI | GPT-4, Gemini, Grok | GPT-4 on a 2026 card |
| 7. AI video generation | Runway, OpusClip, Pika | OpusClip is a clipper rather than a generator — a different job from the other two |
| 8. AI tool stacking | Notion, ClickUp, Zapier | See box 2 |
| 9. LLM evaluation & management | Helicone, PromptLayer, TruLens | The only box that measures anything, and the last one anybody picks |
The order that actually works
Same nine skills, rearranged by what each one needs rather than by how advanced it sounds.
| Stage | What you do | Which skills it unlocks | How you know you are done |
|---|---|---|---|
| 1. Ask well | Prompt engineering, on real work, for a fortnight | — | You stop re-explaining your context every conversation |
| 2. Make something | Publish, write, record — produce the artefact. Video and multimodal belong here | Nothing yet, and that is the point | You have a folder of your own output |
| 3. Notice the repetition | Watch which step you do every week by hand | Automation and tool stacking — now there is something to automate | You can name the task in one sentence |
| 4. Point a model at your pile | Your documents, your archive, your transcripts | RAG, and later fine-tuning once the pile is big enough | It answers from your material, not the internet |
| 5. Hand off a multi-step job | Only a process you have run manually and can describe | Agents | You can write down the steps before you automate them |
| 6. Measure it | Accuracy, cost, quality, over time | Evaluation — which needed all of the above to exist | A change makes a number move |
Stage five deserves one note, since agents are the glamorous box. A process you cannot describe in writing is a process you cannot hand to an agent, and the describing is most of the work. We made the same point about agent platforms: the platform is rarely the thing in the way.
Four honest caveats
- The asset column is our judgement, not the card’s. You could reasonably argue that RAG over a public dataset teaches you RAG, and you would be partly right — you learn the mechanics. Our claim is narrower: you do not end up with a thing you use, which is what “mastering” implies.
- We verified one tool properly. AutoGen, against its own repository README, read this week. The GPT-4 observation is a reading of the card against the calendar. The others we did not test, and several may be excellent.
- “Maintenance mode” is not dead. AutoGen still gets fixes, still has 61,000 stars, and a community fork exists. If you already run it, nothing broke. The point is only what its front page tells somebody arriving new in 2026.
- The card’s footer is better than most. It says tool names change, says to check the official site, and says skills take practice and results vary. Everything above is us taking that instruction seriously, not contradicting it.
The one-line version
Nine boxes, eight distinct skills, and four of them gated on something you have to own before the skill is available to you at all. Nobody is stuck on RAG because RAG is hard. They are stuck because they have no documents worth querying.
So the honest answer to “which one are you mastering first?” is: the first one, and then make something — and the four advanced boxes open on their own, in order, as the thing you made grows. The ninth box, evaluation, is last for the best possible reason: it is the only one that needs all the others to exist first.
Before you pick a skill, check a folder. Open the place where your own work lives — drafts, transcripts, notes, past videos — and count what is in it. That number decides which four of these nine are available to you this month, and no amount of study changes it.
