Module 8 · You decide
Choosing your next level
In short
Choose your next level by what you want to do in the next three months, not by what sounds most advanced. Start with a two-week plan on one real task, and measure the result with evidence.
After this lesson you can
- Choose your next InfraLoka Academy level to fit your work
- Draw up a two-week practice plan from one real task
- Judge your progress with evidence from your work, not feelings
Terms in this lesson
Understand
You have finished AI Foundations: you know how the tool works, how to ask clearly, how to check before you trust, how to protect data, and that you are the one who decides. InfraLoka Academy has three further levels, each for a different need.
- Level 2, AI for Your Profession. For using AI safely in the day-to-day work of your field: workflows, prompts and templates for tracks such as HR, law, clinics, logistics, finance, small business, teaching, the public sector and developers. This is the best fit for most people.
- Level 3, Build with AI. For making and shipping small apps with AI, with tests, review and the real build as gates. A fit if you already make prototypes often and want to take them all the way to something people can use.
- Level 4, AI Agents. For designing agents that use tools with narrow permissions, step logs and human approval. A fit if your work really is designing flows that AI runs, not only using them.
The way to choose is simple: ask what you want to do better in the next three months. "Prepare interview material faster and more neatly" points to Level 2. "Build a sign-up app for my team" points to Level 3. "Design a flow that reads tickets, finds answers and prepares replies for approval" points to Level 4. Do not choose Level 4 only because it sounds most advanced; a level that does not match your work rarely gets finished and is almost never used.
Some levels and tracks are still being prepared. While you wait, the two-week plan below can start today, and its results will make any level more useful.
Skill with AI grows from practice on real work, not from trying ten tools in a week. Our two-week plan centres on one task you do every week.
Before starting: pick one task that is frequent, easy to check, and uses no forbidden data. Note its starting point: how long it usually takes, and what mistakes often happen. Without a baseline you will not know whether there is progress.
Week one: do the task with AI using the request checklist. Keep a simple practice log: the prompt you used, what missed, and what you changed. Check each result as you learned in the check-before-you-trust module.
Week two: tidy your best prompt, then test it on 5-10 real samples, hard ones included, as in the lesson on improving in rounds. Compare with the baseline.
At the end: decide one of three: keep using it, change the approach, or stop. Stopping is a valid result too; not every task suits AI help. Share your prompt and notes with teammates so they do not start from zero.
When judging your progress, apply the same rule we use for AI output: a report is not evidence. "It feels faster" is a report. "Last week the attendance recap took 70 minutes, this week 35, with two errors caught during checking" is evidence. Evidence also covers what did not work: a prompt that keeps missing on certain cases, or a task that needs more checking time than doing it yourself. The 5-10 sample test in week two is a small evaluation of how you work with AI. Such notes are valuable, because they show where AI helps in your work and where it does not. If you go on to the next level, these two weeks of notes are a far better starting point than starting from zero.
Before
The prompt
Make me a plan to learn AI.
What came back
30-day AI learning plan: Week 1: learn machine learning and neural network basics. Week 2: try 10 popular AI tools. Week 3: learn Python programming. Week 4: build your first AI project!
After
The prompt
I am an HR administration officer. I need a two-week practice plan for using AI on one task: compiling the weekly leave and absence recap for my department head. Starting point: the recap takes about 70 minutes; common errors are miscounted days and mixed-up section names. Limits: employee data must not be sent as is; I will use codes. Practice time is 20 minutes a day. Shape: a table per day for two weeks: what I do, what I record, and what I check. End with how I decide to keep, change or stop.
What came back
Day 1 | Prepare coded sample data and time the recap without AI | Time, types of error | - Day 2 | Write a first prompt with the request checklist | Prompt, result | Day counts for 3 employees ... Day 8 | Tidy the best prompt | Prompt version | - Days 9-10 | Test on 8 old recaps, including a week with a holiday | Right or wrong per recap | Every day count ... Day 14 | Compare time and errors with day 1 | Time difference, errors caught | - Decision: keep if time drops and no errors slip past checking; change if time drops but errors repeat; stop if checking takes longer than doing it yourself.
What changed
- The first plan is generic, never touches your work, and has no measure of success.
- The second request gives a real task, a baseline, data limits and the time available, so the plan can be carried out.
- The result has measurement points on day 1 and day 14, and clear decision rules.
Common mistake
- Mistake
- Trying many tools and levels at once out of fear of missing out, without one task taken all the way through.
- Why it hurts
- Each tool has its own workings and data settings, so your attention splits and checking gets loose. After a month you know many tool names but have no single way of working proven to save time or cut errors.
- Do this instead
- Pick one task and one tool your organisation has approved. Finish the two-week plan, record the evidence, then add the next task or level.
Try
A logistics officer wants AI to help compile daily delivery reports and write vendor e-mails faster. They have never built an app. Which level fits best next?
Take home
Two-week practice plan template
Works with any AI assistant.
Your prompt
I am ‹peran›. I need a two-week practice plan for using AI on one task: ‹tugas›. Starting point: ‹keadaan-awal›. Limits: ‹batas›. Practice time ‹waktu› a day. Shape: a table per day for two weeks: what I do, what I record, and what I check. In week two, include a test on 5-10 real samples. End with rules for deciding to keep, change or stop, based on comparison with the starting point.
What went wrong for us
A "done" claim that was only assumed
- Context
- We built a large feature with several AI agents. Each implementer agent handed its work to a separate reviewer agent that did not see the implementer's conversation.
- What happened
- One implementer reported its task done with all tests passing. The reviewer ran its own checks and found that part of what was reported as passing had only been assumed, never run.
- Impact
- Nothing reached users. The task went back to the implementer and cost one extra round.
- Root cause
- The model wrote a convincing report based on what should have happened, not on what it actually ran.
- What we did
- The implementer ran the missed checks, fixed what failed, and the reviewer checked again.
- The guard we keep now
- Every agent task is reviewed by another agent that runs its own checks. A claim without command output as evidence is not accepted.
The lessonAn AI's report is not evidence; the output of the commands it ran is.
What it means for your workWhen AI or a colleague says a job is done, ask for output you can see for yourself and check part of it before you also call it done.