This course only pays off if it changes what you do next. Not what you admire, not what you bookmark, not what you tell yourself you understand. What you do.
So the capstone is simple on purpose: write a six-month learning map that forces selection. You are not building a grand theory of your AI future. You are choosing one compounding skill, one evaluation method, one source system, and one shipping cadence that can actually survive your working life.
Keep learning in the step-change open beside you as you do this. Its role is to remind you that the point is not maximum coverage. The point is sane adaptation in a volatile environment.
Step one is to choose one compounding skill for the next six months. Not three. One.
If you are a PM who keeps getting dragged into fuzzy AI planning conversations, prompt-as-spec is usually the best choice because it sharpens product thinking and gives you a review language the team can share. Start with prompt design as product design.
If you are already writing prompts or prototyping workflows but cannot tell what good looks like, choose evaluation. Start with eval before launch.
If your team has already crossed that threshold and is trying to make models act safely inside systems, choose tool-use schema design. Start with tool use, function calling, and agents.
If you are tempted to choose "agents" because it sounds advanced, slow down and ask whether you have truly earned the prior layer. Usually the right capstone choice is the deepest missing prerequisite, not the most exciting headline.
Lever: one skill concentrates attention and produces visible compounding.
Risk: you may worry that focusing means missing the rest of the frontier.
Rollback: keep a small watchlist for adjacent topics, but do not let the watchlist become the curriculum.
Step two is to write the twelve-example self-eval for your chosen skill. Use the structure from the prior lesson. Make the questions concrete enough that a wrong answer would be obvious to another smart person. If your chosen skill is evaluation, your self-eval should include questions about test-set quality, failure mode coverage, rubric design, and what counts as a meaningful regression. If your chosen skill is prompt-as-spec, your self-eval should include questions about boundary conditions, failure behavior, reviewability, and why each instruction line exists.
This self-eval is your baseline. Run it now. Then run it again in six weeks and in twelve weeks. If the questions stop feeling hard, sharpen them.
Step three is to choose two trusted sources and three ignore-list sources.
Your trusted pair should cover different functions. One should help you detect real change. The other should help you interpret it. This is where the operator-plus-explainer combination is strong. You can also replace one with a case-based source if cases help you think more clearly than commentary.
Your ignore list matters just as much. Name three source types you will deliberately downgrade for six months. Examples: recap feeds with no primary evidence, obituary accounts that declare whole categories dead, or frontier-posters who rarely discuss cost, failure, or rollback. The point of the ignore list is not moral superiority. It is cognitive hygiene.
Lever: source discipline protects the learning plan from panic contamination.
Risk: you may become overconfident in a narrow reading diet.
Rollback: review the source list every quarter and add one dissenting but evidence-rich source if your thinking starts to feel too comfortable.
Step four is to schedule a three-week ship-loop around the chosen skill.
Pick one artifact that matters enough to get honest feedback. A founder might choose an AI-assisted support workflow. A PM might choose a prompt review and rewrite for a live internal assistant. A designer might choose an interface pattern for human review around AI output. An engineer might choose a small tool-using assistant for a repetitive internal task. The artifact should be real enough to matter and small enough to finish.
Then put dates on it. Week one is setup and first draft. Week two is real use. Week three is one revision based on one serious piece of feedback. Put the review meeting on the calendar now. If it is not scheduled, it is not a plan.
Step five is to define your monthly and quarterly checks.
Monthly, ask: did I actually spend time on the chosen skill, and did the last ship-loop make my understanding more precise?
Quarterly, ask: is this still the highest-leverage skill, or has reality exposed a deeper missing layer?
That second question matters. The point of a learning map is commitment, not stubbornness. If you discover that your chosen skill depends on a more basic gap, update the map. That is not failure. That is intelligence.
You now have the whole capstone deliverable:
- One compounding skill for six months.
- One twelve-question self-eval.
- Two trusted sources.
- Three ignore-list source types.
- One three-week ship-loop scheduled with a real artifact and review date.
- One monthly check and one quarterly check.
That is enough. More planning would mostly be decoration.
People who feel permanently behind on AI often believe the cure is a bigger syllabus. Usually the cure is a narrower one. The market wants you scattered because scattered people are easy to alarm and easy to sell to. A good learning map is an act of refusal. It says: I will not learn everything. I will learn the few things that improve my judgment, my output, and my sequence.
This is how you stop churning. Not by caring less, but by deciding better.
Rules from this lesson
- Choose one compounding skill for the next six months; focus beats frontier FOMO almost every time.
- Write and rerun a twelve-question self-eval so progress is measured, not imagined.
- Protect the plan with source discipline: two trusted sources and three ignored source types.
- Put the three-week ship-loop on the calendar; unscheduled learning plans are fiction.