Only you
Learn the four-quadrant method: Structure it, Understand it, Choose tools, Build it. Take one real task from your week to a first working result, and learn what AI assistants can and can't do along the way.
Outcome: One working first result for a task from your own week, with five lines on the problem, what exploring changed, your choice, the result and your next step.
Put the material, the reader, an example and the exclusions into one chat request, then improve the answer with follow-ups that change one thing at a time, so you can say which change fixed it and reuse it.
Outcome: A before/after pair from your own notes, with the one follow-up that fixed it, which passes when a colleague can see the improvement and finds nothing in the final version that isn't in your notes.
Write a prompt with a role, a task, constraints and an example, test it on five real inputs, improve it one change at a time, and tell an unclear instruction from missing information, so a weekly task gives the same clean result every time.
Outcome: One saved prompt improved across three revisions on the same five inputs, with a one-line note per revision, which passes when a colleague can see what each change did and every missing field reads "not stated" instead of a guess.
Load a set of your own documents into Gemini Notebook, or a local tool, ask questions they can answer, and follow every citation back to its passage, labelling each claim supported, unsupported or not found before you use it.
Outcome: Three answers from your own documents with every claim labelled and every supported claim traced to file, page and a short quote, which passes when a colleague can find each passage within a minute and nothing marked supported says more than its passage.
Build one n8n workflow that turns a form entry into a logged row with a timestamp. Test it before you switch it on, break it once on purpose so you know what failure looks like, and find the off switch before you need it.
Outcome: A screenshot of your workflow's Executions tab showing one successful and one deliberately failed run, plus one line saying where its off switch is; it passes when the failed run left no row in your table.
Describe a small calculator or helper by its input, action and output, have OpenCode Desktop build it as one HTML file, check it against answers you worked out by hand, then improve it once because using it showed you what was missing.
Outcome: Your index.html plus three test cases, each with the input, your hand-calculated answer and the tool's answer; it passes when all three match, an impossible input is refused, and one improvement came from real use.
Step by step, with a screenshot for every step: install OpenCode Desktop, pick a free model, have it explain a small page, change one line, see exactly what changed and take it back.
Outcome: A before-and-after table of five searches for your changed page, a screenshot of the Review panel showing the one changed line, and one sentence saying how you undid the change; it passes when all five results match, only the intended line changed, and Undo brought the old behaviour back.
Make one finished image or 60-second narrated clip for your own work. Let the generator supply the look, keep exact words and structure in your own editor, check the rights for your use, and label what AI made before anyone sees it.
Outcome: One finished image or clip with its AI label, its alt text or captions, and the prompt that produced it; it passes when a colleague finds no wrong word, no missing label and no input you lack the rights to use.
Upload a table to a chat assistant, get an answer without writing a formula, and check it: ask which rows it counted, recount one group in your spreadsheet, and catch the three silent failures of dropped rows, numbers read as text and filters you never asked for.
Outcome: Three answers from your own table, each with the assistant's row accounting and your own recount of one group; it passes when every source row is accounted for and each recount matches or explains the difference.
Install LM Studio, download a small open model that fits your laptop's memory, and get an answer with the network switched off. Learn what running locally gains you, what it costs, and which tasks a small model handles well enough.
Outcome: A screenshot of LM Studio answering your own prompt while your laptop shows no network connection, plus three lines (model and download size, your installed RAM, one task you'll use it for); it passes if the answer was generated after you disconnected.
Sort anything you want to paste into an AI tool as green, yellow or red in seconds, and turn a red request into a green one by anonymising, asking about the form instead of the content, or using an approved tool. For labs and small companies.
Outcome: Your last five pastes into an AI tool, each classified green, yellow or red with a one-line reason, and every red one rewritten; it passes when a colleague agrees with each colour and no rewrite still contains red material.
Use an AI assistant for literature triage, language and code without misrepresenting your work. Keep a one-line-per-use log as you go, check what your journal or funder requires, and write a disclosure statement specific enough that a reviewer knows what the tool did and what you did.
Outcome: A disclosure statement for one real piece of your work, checked against the target journal's or funder's AI policy, with the use log behind it; it passes when a co-author can match every sentence of the statement to a log entry and to the policy.
Choose a tool for one task of your own: write the must-haves, knock out candidates on paper, then try one challenger against the way you do it today on the same input. Compare finished results, including repair time, cost, data handling and how you get your work out.
Outcome: A one-page choice note for a task from your own week: must-haves, the candidates you knocked out and why, a trial of one challenger against your current method on the same input, and a one-sentence decision; it passes when a colleague could repeat your trial from the note and would reach the same choice.
Run Cellpose, a free pretrained cell-segmentation model, on a public microscopy image on your own laptop. Check its outlines by eye, compare its count with your own count of one region, and learn the signals that tell you not to trust a count before it goes into your notebook.
Outcome: A segmented image plus a manual-vs-model count on one region, with your counting rule and one sentence on where the model failed; it passes when a labmate can recount the same region with your rule and get your manual number.
Put a file into a private Gitea repository, check each change in red and green before you save it, read its history, open old versions and take one change back with Revert, all in the browser.
Outcome: A private Gitea repository whose main file has at least four commits in its History, including one revert; every message says why, and you can show the red-and-green view of one change and the earlier version within a minute.
What an AI harness is, seen through three real runs of one bug fix: same model, same task, three harnesses, different steps and costs. Then add a run of your own in OpenCode.
Outcome: A four-row comparison table (three given runs plus your own OpenCode run) with result, minutes and step counts, one sentence naming the biggest difference, and three lines for your own task: the task, a check the agent cannot see, and two measures.
Install OpenWork, a free desktop app in which an AI agent works on the files in one folder. Give it a practice folder, switch it so that every change waits for your click, hand it one task, and check the table it wrote against the files.
Outcome: The overview table the agent wrote for your practice folder, plus a five-line note: the model you used and where your files went, files counted against rows, three rows checked against their sources, one thing the agent got wrong or marked as not stated, and which requests you allowed. It passes when the row count equals the file count and the source files are unchanged.