Level 2 · Power user

Your team sees a bad document or a wrong figure

21books at this level
T01-L0201

Check AI Output Before Your Team Relies on It

Once colleagues act on your AI-assisted work, a wrong number or a dropped condition travels. Sort output by where it goes, run a five-step check on anything that leaves the team, and attach a short check note so a colleague can trace every claim without asking you.

Outcome: A corrected team deliverable (a brief, a summary, an FAQ or a caption) with a check note: source title, version and date, each material claim with its quote and location or marked unknown, and any arithmetic written out. It passes when a colleague traces three claims you didn't pick to the source without asking you.

Power user15 minLast edited: Sep 23, 2026
T02-L0202

Set Up Projects a Colleague Can Pick Up

Stop pasting the same background into every chat. Give each job its own project with instructions, source files and a dated decisions file, keep memory from carrying one job into another, and set it up so a colleague can continue your work in a fresh chat while you're away.

Outcome: A project for one real team job, with instructions, source files and a dated decisions file whose master copies sit in a shared folder, and a colleague's continuation in a fresh chat. It passes when the continuation follows the current decision, not an old one, and the colleague needed nothing from you beyond the project and its folder.

Power user15 minLast edited: Sep 23, 2026
T03-L0203

Build a Context Packet Your Team Can Rerun

Most bad answers come from missing or wrong sources, not from weak wording. Keep the prompt fixed, change only the sources, and see which document the answer really needed. Then package the question, the labelled sources and their exact wording so a colleague gets the same answer without your chat history.

Outcome: A context packet for one real team question: the fixed prompt, the question, the labelled sources with versions and the exact quotes the answer rests on, plus a before and after run showing which source made the difference. It passes when a colleague reruns the packet in a fresh chat and gets the same answer with the same quotes.

Power user15 minLast edited: Sep 23, 2026
T04-L0204

Curate a Document Set Your Team Can Ask

A shared document assistant answers from whatever is in it, including last year's policy. Write a rule for what goes in, keep a one-line register per document, test the set with three questions, share it at the right level, and name who keeps it current, so colleagues get today's answer with a citation they can open.

Outcome: A shared document set with its written rule and source register, answering three test questions (a current rule, an exception, and one the set can't answer) with citations that open to the right passage, plus one answer a colleague has used in real work. It passes when every document in the set appears in the register as current and a colleague can find the person who keeps it current.

Power user15 minLast edited: Sep 23, 2026
T05-L0205

Put AI in the Middle of a Workflow, and Check Every Answer

Add one model step to an n8n workflow that sorts incoming papers or emails for your team. Ask for fixed fields, check each answer in the workflow before anything moves on, send the doubtful ones to a colleague, and keep anything that leaves the team behind a human Approve button.

Outcome: A published n8n workflow with a model step, a check step, a shared review table and, if anything leaves the team, an approval step; plus a one-page note naming the reviewer and deputy, the monthly cost estimate and the spend limit. It passes when a colleague feeds in one normal item and one broken answer and the broken one lands in the review table, not in the team's list.

Power user13 minLast edited: Oct 18, 2026
T06-L0206

Build a Small Shared Tool From a Prompt, and Test It With a Colleague

Turn a shared spreadsheet your team keeps breaking into a small web app built from prompts in Base44. Describe the screens and data first, add one change at a time, restrict who can read and edit, and let a colleague prove it works in their own browser before anyone relies on it.

Outcome: A live Base44 app used by at least one colleague, plus a half-page note: the screens and fields, the access and permission settings as set, the colleague's test with one allowed and one refused action, the CSV export opened in a spreadsheet, and the account holder and deputy. It passes when a signed-out visitor sees no data and the colleague's test matches the note.

Power user15 minLast edited: Sep 23, 2026
T07-L0207

Hand a Task to an Agent, and Check What It Did

Give Claude Cowork a real preparation task on a copy of your team's files: a comparison table from a pile of protocols or supplier quotes. Write a delegation card with the outcome, the limits and the stop point, watch what the agent actually does, and have a colleague check the result before the team uses it.

Outcome: The finished table or brief, plus a half-page delegation note: the card as given, the folder used, what the agent did and where you stepped in, your spot checks and the colleague's, and who signed it off. It passes when every value in the result names its source file, and the colleague's blind check of five values finds them all in the sources.

Power user15 minLast edited: Oct 12, 2026
T08-L0208

Get a Script Your Colleague Can Rerun, Without Being a Developer

Turn a monthly spreadsheet chore into a small Python script with OpenCode Desktop, then make it something a colleague can run without you: a known-answer test worked out by hand, clear refusals for bad input, a one-page README, and a shared Git repository your team can find.

Outcome: A shared Git repository with the script, a README, a known-answer test file and its expected output; it passes when a colleague, on their own computer and using only the README, gets exactly the expected output from the test file, and gets a clear refusal with a line number from one broken row.

Power user21 minLast edited: Oct 18, 2026
T09-L0209

Keep One Look Across Your Team's Images

Turn one good image into a style kit your whole team can use: a one-page style sheet, three references you have the rights to, a prompt frame and a template. Then test it the only way that counts: a colleague makes a matching image from the kit without asking you.

Outcome: A shared style-kit folder (style sheet, three references with their sources, prompt frame, template, label line) plus five finished images from it, one made by a colleague; it passes when a third person can't pick out the colleague's image and finds no reference without a recorded source.

Power user17 minLast edited: Sep 23, 2026
T10-L0210

Clean a Messy Export Once, So Your Team Can Refresh It

Turn a monthly export with a broken header, mixed units and placeholders into a clean table a colleague can refresh without you. Let an assistant find the problems, agree the rules with the person who owns the data, and let Excel's Power Query apply them the same way every month, with checks that show every row.

Outcome: A shared workbook that cleans the export with recorded steps, plus its rules sheet and check block, refreshed on a second month's file by a colleague; it passes when rows in equal clean rows plus flagged rows plus removed duplicates, and the colleague explains one flagged row from the workbook alone.

Power user17 minLast edited: Sep 23, 2026
T11-L0211

A Local Model Your Tools Can Call

Turn the model in LM Studio into a local server that other programs can use. Connect Open WebUI or n8n, installed through Heidelberg AI Desk, with three settings, check the results with a fixed set of examples, and choose between local, EU-hosted and frontier models task by task.

Outcome: One tool installed through Heidelberg AI Desk (Open WebUI or n8n) that gets its answers from LM Studio's local server, plus a check sheet of five examples with known answers; it passes when all five come out right with the network switched off, and your three lines name the address, the model and the task.

Power user15 minLast edited: Oct 12, 2026
T12-L0212

Clean the Copy, Label the Result, Spot the Trap

Before shared data goes into an assistant, make a working copy with only what the task needs and check names, dates, free text and hidden parts of the file. Let a second reader try to recognise someone. Say what AI did in a note that stays with the text, and stop when a document gives orders.

Outcome: A cleaned working copy for one real team task, with a method note and an AI-use note. It passes when a second reader who hasn't seen the original can't single anyone out, one result the task needs (a count, a grouping) comes out the same from the copy as from the original, and the AI-use note is still attached in the document where the result ended up.

Power user15 minLast edited: Sep 23, 2026
T13-L0213

Show a Colleague AI on Their Own Task, in 30 Minutes

Run a 30-minute live demo on a colleague's own task instead of a slide deck. Pick the task with them, use material that's safe to show, check the result in front of them, hand over the keyboard, and let them decide to try, decline or wait, with the decision written down where the team can find it.

Outcome: A one-page demo record from a session with a real colleague: their task, the material and tool, the timed draft and check, the change they requested, the objection raised, and their own choice; it passes when the session took 30 minutes or less, the result was checked in front of them, and the choice is in their words.

Power user14 minLast edited: Sep 23, 2026
T14-L0214

Brief Your Team From the Shared Drive, Through a Connector

Connect Claude to your team's SharePoint and OneDrive through the Microsoft 365 connector, and turn the current files into a brief a colleague relies on. Scope the question to one site, make every claim name its file and date, check that the recipient can open each source, and switch the connector off when you're done.

Outcome: A brief built through the connector, plus a short note: the site or folder you scoped it to, the list of cited files with their dates, which of them the recipient can open, what you corrected, and how you switched the connector off. It passes when the recipient opens three cited files at random and finds each claim, and no cited file is one they can't open.

Power user14 minLast edited: Sep 23, 2026
T15-L0215

Record an AI Step So a Co-Author Can Rerun It

Make one AI-assisted research step repeatable by someone else. Pin the model, keep the input, prompt and raw output together, decide before rerunning what counts as the same answer, and report the differences honestly when a co-author, reviewer or auditor asks you to run it again.

Outcome: A shared folder for one AI-assisted step (input with checksum, prompt version, run record, raw outputs of at least two runs, comparison) plus a co-author's rerun from the folder alone; it passes when the co-author's rerun is compared on the rule you set beforehand and the Methods sentence states the agreement and what couldn't be fixed.

Power user16 minLast edited: Sep 23, 2026
T16-L0216

Choose Tools Your Team Can Use, and Afford

Choose an AI tool for a team of three to eight and know what it will cost in a busy month, not just a quiet one. Read an allowance, estimate one task from the vendor's prices, pick seats or pay-per-use, set a cap and an alert, and test that colleagues can actually pick up each other's work.

Outcome: A one-page tool decision for the team: the workload, a monthly cost estimate for a normal and a busy month from the vendor's dated prices, the cap and alert that are actually set, a named admin and deputy, and the trial result; it passes when a colleague can recompute the estimate from the page and the vendor's price page.

Power user16 minLast edited: Sep 23, 2026
T17-L0217

Frame a Problem Your Team Agrees On

Turn a vague request like "we need an AI tool for X" into a one-page problem statement your team signs off: who has the problem, what's hard, what a useful result looks like and how you'll know. Includes a short stakeholder conversation and an assistant that challenges your framing.

Outcome: A one-page problem statement for a real request in your team, with who, what's hard, a useful result, a test and a list of what's out of scope, which passes when two colleagues have read it and signed it with no open objection.

Power user15 minLast edited: Sep 23, 2026
T18-L0218

Predict a Protein Structure and Read Its Confidence Honestly

Run a structure prediction on the free AlphaFold Server, read pLDDT, PAE, pTM and ipTM for what they do and don't say, cross-check with a BLAST search against the PDB, and write a confidence note that a labmate can check against the same files before anyone designs a construct from it.

Outcome: One AlphaFold Server prediction with its downloaded files and a written confidence note (confident regions, relative positions, interface scores, one cross-check, what it can't tell you); it passes when a labmate opens the same files and agrees with every claim in the note.

Power user16 minLast edited: Sep 23, 2026
T19-L0219

Work on a Shared Gitea Repository From Your Computer With OpenCode

Connect your computer to Gitea with an SSH key, copy a repository down, commit, push and pull with OpenCode, give a colleague access, and stop OpenCode from settling a conflict on its own.

Outcome: A shared repository with a commit pushed from your computer under your name, a colleague's Gitea commit pulled to your computer, a .gitignore for raw data, and one same-line conflict stopped by OpenCode and decided by the two of you; every message says why.

Power user16 minLast edited: Oct 18, 2026
T20-L0220

Choose a Harness by Testing It on Five of Your Own Tasks

Run DeepSeek Harness next to OpenCode on the same free model, test both on five tasks from your own work with answers kept away from the agent, and write a recommendation your group can check.

Outcome: A ten-row table (five tasks x two harnesses: known answer, minutes, three habits) and a recommendation of at most five lines naming the pick, the deciding score, the habit that mattered, what the test did not cover and when to test again.

Power user12 minLast edited: Oct 18, 2026
T21-L0221

Turn a Task You Repeat Into a Skill a Colleague Can Run

Do a task once in OpenWork, then save the way you did it as a skill: a short text file the agent follows the next time someone asks. Read the file, test it on new material, and hand it to a colleague who runs it without you.

Outcome: The skill folder with its SKILL.md, plus a one-page test sheet: your run on the first set of files, your run on a second set, one request that correctly did not use the skill, and a colleague's run with the model they used and where they got stuck. It passes when the colleague produced the result without your help and three rows they checked match the sources.

Power user14 minLast edited: Oct 12, 2026