Start with the work in front of you. Find a pattern, make something small, and check whether AI actually helped.
The aim might be better quality, stronger judgment, a new capability, or time saved. The right starting point depends on your task and what you can check.
12 activity patterns. All are garden drafts, authored and not yet human-tested.
check
Check claims before they travel
Compare each consequential statement with its source and look for missing caveats, changed numbers or invented certainty. A second AI opinion can suggest checks; a person decides whether the evidence supports the claim.
Make the trade-offs between a few realistic choices explicit. Agree the criteria first, include doing nothing and leave the final choice with the person responsible.
Choose the detail, examples and vocabulary a reader needs to act. Preserve the meaning and uncertainty while changing how the explanation is presented.
Locate evidence for a particular question, check where it came from and record what it cannot establish. A plausible result is a lead until its source has been opened and checked.
Try the same bounded workflow on changed inputs, record where it failed and adjust its instructions and checks. Keep an accountable person and a fallback before considering automation.
Work out what a table can tell you by checking its units, missing values and denominators before calculating. Describe the observed pattern without inventing a cause.
Separate observations, possible explanations and unanswered questions before choosing what to try. Keep contradictory examples and the limits of a small sample in view.
Turn an agreed goal into a short sequence of actions, with an owner, a finish condition and unresolved dependencies. Do not invent commitments to fill gaps.
Turn scattered notes into a short account of what changed, what is blocked and what needs a decision. Keep a source beside each claim and leave missing information visible.
Reduce a source set to the points that matter for one question. Keep the trail back to the source and distinguish missing evidence from evidence that something is absent.
Choose a small, reversible trial that could change your mind. State the expected signal before trying it, inspect failures and decide what the result warrants.
For a beginning-to-end example, try a weekly status update, a presentation story, or a discovery brief. Each uses synthetic inputs and includes a check and a failure test. They are authored garden drafts, not human-tested practices.