InfraLoka Academy
Latihan Nyata
Practice briefs from real work. Pick one, make the deliverables and tick them off as you go.
- Your first read-only agent for an SOP folder
New staff at an invented office often ask things that are already in the SOP folder. Prepare a folder of 5 invented SOPs, run an agent with read-only permission, and ask it 8 questions. Done means every answer names its source SOP, you have read the agent's trace, and wrong answers are recorded.
- A gated agent that asks before it writes
An invented admin team wants an agent that updates the office inventory sheet from purchase notes. Build an agent that may read freely but must show its planned changes and wait for human approval before writing anything. Done means the gate is proven unskippable, including when a purchase note contains instructions telling the agent to skip it.
- A folder-summarising agent with a test set
An invented team's document folder holds minutes, SOPs and reports nobody has time to read. Build a read-only agent that summarises every document in the folder and writes one index. Done means the agent is tested against a set of 10 invented documents whose ideal summaries you wrote yourself, scored with the same rubric each time.
- A team tool to log borrowed office equipment
The projector, camera and spare laptops at an invented office are often borrowed with no record of who has them. Build a small web app with sign-in, an item list, borrow and return records, and history, from spec through release to a practice environment. Done means the main features have tests, database changes use migrations, and there is a written way back to the previous version.
- Find patterns in late deliveries
An invented delivery team feels it is often late, but nobody has looked at the data. Create 60 invented deliveries with route, day, vehicle, departure time and minutes late, then use AI to look for patterns. Done means two or three findings backed by numbers, one honest chart, and every key figure recomputed by you.
- Compare three suppliers before deciding
An invented catering business needs to pick a new packaging supplier. Write three invented quotes with different prices, lead times and terms, then use AI to draft criteria and a comparison table. Done means the business owner decides from a table whose figures you rechecked, not from AI's recommendation.
- An instruction file for a team AI assistant
Everyone on your team retypes the same rules each time they use an AI assistant. Write one shared instruction file covering the team's purpose, tone, output format and data that must never go in. Done means the file is tested on 6 real team tasks and the results are more consistent than without it.
- Fact-check one article
Pick one news article or popular piece that gets shared a lot around you. Ask AI to pull out every factual claim, then check each claim against the original source yourself. Done means every claim is marked true, wrong or unconfirmed, with a source you opened yourself.
- Five honest product descriptions
An invented craft shop needs descriptions for five products in its online stall. Write a fact sheet for each product first, then ask AI to draft appealing descriptions with no claims you cannot back up. Done means every sentence in each description can be traced to the fact sheet.
- One small feature, test first
Pick one small feature in your own practice project, such as a date filter on a list or a CSV export. Write a short plan, write a failing test, and only then ask AI to help write the code. Done means the tests pass, you can explain every changed line, and a colleague has reviewed the change.
- A safe resident-complaint form
An invented government office wants to receive resident complaints through a web form. Build a form and a staff page that ask for as little data as possible, validate input, and limit who can read complaints. Done means basic security tests are recorded, and the authorised official approves the list of requested data before the form goes live.
- From a sign-up form to a weekly summary
An invented school opens sign-ups for inter-class competitions through an online form. Build the form, connect it to a spreadsheet, and set up a weekly count of entrants per competition. Done means the flow works with 15 invented sign-ups you entered yourself, and the form asks only for data that is truly needed.
- A one-page site for a small shop
An invented tailor shop needs one simple page with services, prices, opening hours and how to order. Build a prototype of the page with AI's help and open it on a phone. Done means the page reads clearly on a phone screen, all information matches the shop's fact sheet, and there are no invented claims or reviews.
- Work out margins for 10 shop products
The owner of an invented cake shop sells 10 products but does not know which ones earn the most. Lay out ingredient, packaging and overhead costs per product, then use AI to build a margin table. Done means you can explain every margin figure with a formula, and you have recomputed at least three products with a calculator.
- A one-month content calendar for a shop
An invented coffee stall wants to post regularly but runs out of things to say midweek. Build a calendar of 12 posts for one month with AI's help, based on the stall's real menu and activities. Done means every post has a goal, draft text, and no promotional promise the owner has not approved.
- A leave-balance calculator built from a spec
An invented HR team is often asked about leave balances and works them out by hand. Write a short spec of invented leave rules, then build a small web calculator with AI's help that works out remaining leave from start date and leave taken. Done means the calculator passes every test case you wrote before building it, and HR has matched its rules to the policy.
- A monthly report with variance commentary
The head of an invented company wants a monthly profit and loss report that explains why figures differ from budget. Build three months of invented ledger entries and a budget, prepare the report, then use AI to draft variance commentary. Done means every figure traces back to the ledger, every comment matches the data, and the finance lead approves it.
- A shift handover report that assembles itself
At an invented warehouse, night-to-morning handover is written freely in a chat group, so important things get lost. Build a small flow that takes an end-of-shift form and assembles a fixed-format handover report, with AI only tidying the free-text notes. Done means the flow is tested with 5 invented shifts, including one with a safety incident that must never be summarised away.
- A patient leaflet on taking medicine as directed
An invented clinic wants a general leaflet on taking medicine as directed for its waiting room. Draft it with AI's help from official health sources you choose, in easy-to-read language. Done means the draft holds only general information, every statement has a source, and a qualified clinician has reviewed and approved it before printing.
- Rewrite a job ad to be fair
An old ad for a warehouse admin role at an invented company has requirements unrelated to the job, such as age limits and appearance. Rewrite it with AI's help so the requirements are job-related, clear and fair. Done means every requirement can be tied to the work and an HR colleague has reviewed it.
- Teaching material for one lesson
A science teacher of an invented class needs a plan and worksheet for one lesson on the water cycle. Build the lesson plan, worksheet and answer key with AI's help, then check every fact. Done means the material fits the grade level, holds no wrong facts, and a fellow teacher has read it.
- A sourced employment-law research memo
An invented client, a small business, asks about the rules for fixed-term employment contracts for seasonal staff. Write an internal research memo answering it with AI's help, with every legal statement resting on regulations you read yourself. Done means the memo traces back to its sources, recent changes to the rules have been checked, and a qualified advocate reviews it.
- From meeting minutes to an action list
Coordination meetings at the village office often end without anyone knowing who does what. Write one page of invented meeting minutes, then ask AI to turn them into an action list with owners and deadlines. Done means every item has one role, one deadline, and no invented decisions.
- A first-week guide for new hires
New hires at your office are often lost in their first days: who to meet, what to fill in and when. Build a first-week guide for one invented role, based on a list you collect from your team's experience. Done means the guide fits on two pages, is laid out day by day, and a fairly recent hire has checked it.
- A visit-reminder flow for a clinic
Many visits at an invented clinic are missed because nobody sends reminders. Design a reminder flow from a schedule sheet to a message template, using coded slots like Slot 07 with no names or health details of any kind. Done means the flow is clear about which steps are automatic and which stay with staff, and the clinician in charge has approved the message text.
- Notice of new opening hours at the village office
An invented village office will change its service hours from next month. Prepare a formal notice for the notice board and a short version for the residents' chat group. Done means both versions say the same thing, older residents can understand them, and the authorised official has read and approved them before they go up.
- A message to parents about a study trip
An invented class will visit the city museum next week. Prepare a formal notice and a short message for the parents' group with AI's help. Done means both versions carry the same date, time, cost, packing list and permission-slip deadline, and the homeroom teacher has checked them.
- Choose an AI tool that fits your team
Your team wants to adopt one shared AI tool, but there are many choices and the ads all sound alike. Build criteria from the team's real needs, then compare three options using the providers' official documents, not reviews or ads. Done means a written recommendation that covers how the tool treats data, its cost and its limits, with a source for every claim.
- Pick one task worth automating
Not every repetitive job is worth automating. Log 10 repetitive tasks from your work over a week, score each, then pick the one most worth automating and one you deliberately leave manual. Done means your choice is backed by time and risk figures, not a hunch.
- Tidy up a warehouse goods-receiving SOP
The notes on how your warehouse receives goods are scattered across a notebook, a chat group and the head storekeeper's memory. Write those rough notes in your own words first, then ask AI to shape them into a numbered SOP a new hire can follow. Done means someone who has never received goods can follow it without asking.
- Summarise quiz scores with codes, not names
A teacher wants to see how one class's quiz scores spread and which question was hardest. Build invented scores for 32 students using codes like S01 instead of names, then use AI to summarise the spread and the hardest questions. Done means the summary is correct, no student name ever went into the AI tool, and the teacher still decides final grades.
- A weekly warehouse stock summary
The warehouse head wants to know which items are running low without opening the whole stock book. Build an invented sheet of 30 items with opening stock, in, out and minimum level, then use AI to summarise it. Done means the list of items below minimum is correct by your own count, not just by AI's.
- Summarise a resident satisfaction survey
An invented sub-district office collected 80 satisfaction survey answers with scores and free comments. Write those invented answers yourself, then use AI to group the comments into themes and compute scores per service. Done means every theme has a comment count you recounted, sample quotes, and the authorised official reviews the summary before it is shared.
- Reconcile one month of petty cash
An invented office's petty cash never matches the pile of receipts at month end. Build a month of invented petty-cash entries and a receipt list with a few deliberate mismatches, then use AI to help match them. Done means every mismatch has an explanation, the closing balance is recomputed with your own formula, and an authorised supervisor approves the adjustments.
- A data-protection law summary for a small team
A small team at an invented business handles customer data but does not know its basic duties under Indonesia's Personal Data Protection Law. Draft a one-page summary with AI's help, with every duty pointing to an article you checked in the official text. Done means the team can use it as discussion material and someone qualified in law has reviewed it.
- An invoice tracker with due-date flags
An invented finance team often bills late because invoices nearing their due date are not visible. Build an invoice tracker sheet with formulas that flag invoices that are almost due and overdue, using AI to help write the formulas. Done means every formula has been tested against edge cases you prepared yourself.
- A release runbook with a way back
Your team ships an app with steps only one person remembers. Write a release runbook for your own practice app, including the steps to roll back to the previous version. Done means someone else on the team ran it in a practice environment without your help, rollback included.
- A small script to tidy scanned file names
An invented office's scan folder holds hundreds of randomly named files. Ask AI to help write a small script that renames them to a date-type-number format, with a trial mode that only prints the plan without changing anything. Done means the script has been tested on a copy folder of invented files and can be undone.
- A morning opening SOP for a small clinic
Each morning staff member at an invented clinic prepares the waiting room, front desk and admin equipment their own way. Write a morning opening SOP that covers only admin work and cleaning of shared areas, with no clinical procedures and no patient data. Done means the clinician in charge of the clinic has read and approved it.
- Review one policy and explain it to residents
Pick one published local regulation or public-service policy that affects residents. With AI's help, write an internal review and a plain explanation for residents, with every statement pointing to the official text. Done means the explanation neither adds to nor removes anything from the policy, and the authorised official has reviewed it before publication.
- Review a shop-lease contract against a playbook
Small law offices often review shop-lease contracts for small-business clients. Write your own invented lease of about three pages and a playbook of the client's positions, then use AI to compare each clause against the playbook. Done means a clean findings table, every statutory reference checked against an official source, and a qualified advocate reviewing the result before use.
- Test your agent against prompt injection
An agent that reads outside documents can be hijacked by instructions hidden in them. Prepare 8 invented documents containing injection attempts, run your practice agent against them, and record which ones get through. Done means every attack that got through has a layered fix and has been retested.