Where I test ideas before I trust them.
Everything here is honest about what it is — including the parts that are scripted, unfinished, or possibly wrong. That's the point of a lab.
Pick a question people actually ask me:
The answer appears here. Calmly, of course.
Prompts I actually reuse, free to steal.
Each one has survived months of real work. Copy, adapt, improve — and if you make one better, tell me.
The honest reviewer
Get critique instead of complimentsReview the following [document/code/plan] as a demanding but fair senior colleague. List the three weakest points, why each matters, and one concrete fix per point. Do not mention anything positive until the problems are addressed.
The translator
Expert text → plain languageRewrite the following for a smart reader with no background in this field. Keep every fact. Replace jargon with everyday words, and add one short concrete example where the original is abstract. Target reading level: a curious 15-year-old.
The decision clarifier
Untangle a stuck decisionI'm deciding between the options below. Ask me the five questions whose answers would most change the decision — one at a time, waiting for my answer before the next. Then summarise what my answers reveal I actually want.
The differentiator
One lesson, three levelsTake this teaching material and produce three versions: one for a struggling learner (more scaffolding, smaller steps), one standard, one for a fast learner (extension questions, edge cases). Keep the learning objective identical in all three.
The meeting compressor
Notes → decisions & actionsFrom these raw meeting notes, extract only: decisions made (with who decided), actions (with owner and deadline), and open questions (with who should answer). Discard everything else. If an action has no owner, flag it in bold.
The assumption hunter
Stress-test a planHere is a plan. List every assumption it silently makes, ranked by how badly the plan fails if the assumption is false. For the top three, suggest the cheapest possible test that would verify or kill the assumption within one week.
Current questions on the workbench.
Self-writing status reports
RunningAn agent that watches a project's actual activity — commits, documents, calendar — and drafts the weekly status update with citations to its sources. Humans edit, never start from blank.
“Can a report that writes itself stay trustworthy?”
Trade-language AI tutor
PrototypeAn AI tutor that explains digital skills in the learner's own vocational vocabulary — the welder learns spreadsheets through weld logs, the cleaner through site checklists.
“Does teaching in trade language double completion rates?”
The silent assistant
PrototypeAn AI helper with a hard budget: it may interrupt its user at most once per day. Forced scarcity makes it choose the single most valuable thing to say.
“Is one perfect interruption worth more than fifty good ones?”
Document → explanation pipeline
IdeaFeed in any bureaucratic document (EU funding call, regulation, contract) and get back what it actually requires of you, as a checklist with deadlines.
“Can AI make bureaucracy feel briefly humane?”
Want to pilot one of these in your organisation, or argue that one is a terrible idea? Both messages are welcome. You can also read the thinking behind the lab in the library.