Automation & agents.

Choose one supported app event, narrow the matching rule, and check a private report before relying on an event-triggered task.

Use a fictional source to show how a document can contain instructions that the reader never authorised.

Use AI to draft a one-page ownership note, leaving undecided responsibilities visible for the people who must agree them.

Turn a short process description into a handoff map that separates work, waiting and decisions.

Use ChatGPT to turn one reporting job into a remit with sources, allowed actions, review and an end date.

Define a few inbox categories, test them on past emails and give AI the label and draft switches before anyone discusses send or delete.

Ask ChatGPT, Claude or Copilot for specific accessibility barriers tied to the page's code, check them with the keyboard, and fix one at a time.

Write the valid and invalid cases before AI builds the form, run them yourself and repair the rule that lets the wrong value through.

Give AI the rules of a process, check a state table first, then click through a one-file prototype whose save and approve buttons say they're simulated.

Ask AI to build a one-file searchable page from a procedure or glossary, then check that every rule, exception and definition kept its exact wording.

Keep Codex read-only, ask for a plan that names every file it will read, create, change or delete, and resolve vague instructions before anything is written.

Create a proposed classification in an internal review queue, then check errors, notifications and safe recovery before activation.

Choose Copilot Chat, Cowork or an agent for a supplier job, say where it must stop, and check what was written and what was done.

Understand the words connector and MCP, then turn a supplier's answers into a short access-review brief.

Tell a local preview, a GitHub change and a deployed website apart, and record evidence for the stage the work really reached.

Count, test and write down where one recurring task belongs, with the condition that would change the decision and a date to look again.

Specify and build a small local calculator, check its known answers and invalid inputs, and test the page in a browser before wider use.

Connect one small practice folder to Claude Cowork, check what it can read, and review one new output file without changing the sources.

Write a short access note for a Claude connector — what it reads, what it can write and who approves — before an owner enables it for everyone.

Use AI to review a routing table, then build mutually exclusive Make routes with a fallback and visible failed-delivery handling.

Set up one read-only recurring task with visible source gaps, a checked first result and a way to notice a missing run.

Write a stop list for one agent or AI workflow, attach a real control to each line in Zapier, Claude or ChatGPT, and check the settings that quietly undo it.

Check the content, audience, access level and connected behaviour before sharing, then check what the recipient can actually open and do.

Find where Codex shows its skills, plugins and MCP connections, test access with one harmless read and keep answers to what Codex could really reach.

Review the files a Codex task changed, protect the source rules and undo an out-of-scope edit before accepting the result.

Define a small CSV summary, have Codex build a reusable helper and compare its output with counts, missing values and duplicates you know.

Brief a small packing calculator in Gemini Canvas, test it with quantities you can check, and use the same discipline to edit a draft paragraph by paragraph.

See what Notion Agent can edit, have it update a project page from meeting notes, and check and undo anything it changed that you didn't ask for.

Decide whether your job needs Claude Code or a simpler Claude option, then make one small change in a company code project and hand a checked diff to its owner.

Save the instructions for a task that works, test them with five fresh cases and, if your account supports it, package them as a skill.