Use AI on a recurring job while keeping its sources, checks and decisions visible. Start AI for actual work.
Two free courses. No coding required.
Both courses start with a real, low-risk job. Pick the one that matches the problem in front of you.
Use AI on a recurring job while keeping its sources, checks and decisions visible. Start AI for actual work.
Turn scattered reading and meeting notes into a brief you can stand behind. Choose what AI may read, check what it writes, and leave a handover for the next person. Start Notes that do work.
A quick question, a job you repeat, and a task that needs company files are different problems. The course shows you when chat is enough, when to save a method as a skill, and when a connection needs tighter permissions and review.
For a one-off question, use chat and check the answer yourself.
If you repeat the same job, save its context, steps, examples and checks as a skill.
If the job needs files or systems, connect only the approved source and keep the permissions narrow.
If a mistake would matter, have someone else review the result before it is used.
The capstone brings the work into five files. Together they show what you asked AI to do, what it used, how you checked it, what you decided, and what happens next.
Working brief (working-brief.md). Who the work is for, what it must do, what it must not do, and when it is finished.
Source contract (source-contract.md). The sources AI may use, the things it must not guess, and where each claim came from.
Checks and evidence (checks-and-evidence.md). Checks that would catch a missing, unsupported or unsafe result, plus evidence that you ran them.
Decision record (decision-record.md). What you accepted, changed or rejected, and why.
Handover (handover.md). The result, the evidence behind it, the open questions, and the next action.
The final result stays with the sources, checks and decisions behind it.
Each lesson leaves you with a working record that the next lesson can use. By the capstone, those records support one fresh run of the whole job.
Start with real work. Pick a recurring, low-risk job and decide what a useful result would need to do.
Build context that survives the session. Write down the rules you keep having to explain, and say when they go stale.
Choose the mechanism before the prompt. Use policy for a known choice, a workflow for a known sequence, and an agent only where the job needs bounded judgement.
Choose the work surface. Decide whether the job belongs in chat or needs a longer delegated run that you can inspect.
Build checks that can prove you wrong. Write checks from the brief, including one that would catch a plausible wrong result.
Make consequential work inspectable. Stop when important context is missing. Keep enough evidence for someone to replay the decision.
Close delegated work properly. Do not take the tool's word for "done". Verify the result where it was meant to land and write down what happens if automation stops halfway.
Package repeated work. Turn the part that repeats into a skill. Add a connection only when the job needs current data or an external action.
Research and write with evidence. Work out what the evidence supports before drafting. Then edit the argument as carefully as the sentences.
Produce work and act safely. Check the document, spreadsheet or meeting brief itself. Do not send, post or change a record without explicit authority.
Keep the understanding and improve the setup. Keep the useful miss. Change one part of the setup, rerun the job, and hand over the result.
A month-end variance note, contract review table, or weekly operations pack can work if you have approved source material and time to review the result.
It happens more than once, has a known audience and source material, and can be reviewed before anyone acts on it.
It is a live crisis, uses data your organisation has not approved, or ends in an irreversible or high-consequence decision.
Choose the job and material you will use before you begin.
