--- name: prompting-at-work-checks description: 15 rules from the Noesa course "Prompting that actually works". For professionals who use AI at work and want usable output in one or two tries — by briefing it like a good manager briefs a new hire, not by typing one-line prompts. --- # Prompting that actually works — the rules Use with: Claude Code or Claude (save as a skill), Cursor (save under .cursor/rules as .mdc), ChatGPT or any other assistant (paste the text below into custom instructions or a project's instructions). 15 rules, taken from the course at https://noesa.leafsoft.online/c/prompting-at-work Each heading is one thing the course teaches. Most are checks to run on your own output before presenting it as done; a few are background you are expected to have. "Wrong by default" lists 6 specific ones. Apply these to the thing you are producing — the type, the schema, the query, the copy — not only to how you explain it. Where a rule names a field, a format or an identifier, that name belongs in the output. ## Stop typing, start briefing Turn a one-line request into a brief that names the task, context, audience, output, and standard for success. ## Give it the context it can't guess Select and supply the background, audience situation, and purpose that materially change an AI draft. ## Say what done looks like Define an output's format, length, structure, required content, and boundaries as observable completion criteria. ## Show it an example Choose, label, and supply a representative example that teaches the model a pattern without asking it to copy irrelevant details. ## Set the role and the reader Use role and reader framing to set useful perspective, vocabulary, tone, and depth without pretending the model has authority it does not possess. ## Break the big ask into steps Decompose a large AI task into ordered stages with a review decision between each stage. ## Iterate like an editor Diagnose a draft, preserve what works, and issue a targeted revision instruction that changes one dimension at a time. ## Ask for options, then choose Request genuinely different options against named criteria, compare their tradeoffs, and make the final choice yourself. ## Make it show its working Prompt for inputs, assumptions, criteria, inferences, and gaps in a reviewable decision record. ## Ground it in your material Constrain an AI task to supplied source material, require traceable support, and identify gaps the source cannot answer. **Wrong by default:** - A citation records where a claim was filed, not that the cited lines say it. Open every reference and confirm the section actually contains the sentence: a draft can carry a reference on each line and still have most of them unsupported, which is precisely why the references are reassuring. - A reference can resolve to the right topic and still be the wrong case: a clause granting 30 days to cancel a trial does not grant it to a paid annual term, which a separate clause may exclude from refunds outright. Check that the cited lines cover the situation being asked about, not merely the subject — this is the hardest kind to catch, because the reference does resolve. - An instruction to use only the supplied material is not an enforcement. A sentence that appears nowhere in the source can still arrive wearing a section number, because the reference is generated alongside the claim rather than checked against it. Require a quoted line per claim, so an unsupported sentence has nowhere to hide. ## Build prompts you can reuse Convert a successful work brief into a reusable template whose blanks preserve the decisions future teammates must make. ## Prompt for numbers and tables Brief an AI to organize numerical material with explicit units, periods, formulas, missing-value rules, and traceable source labels. **Wrong by default:** - 21.7% is Q4 measured against Q1 — the whole year. Quarter on quarter is Q4 against Q3, which is 11.8%. Both numbers are real and they answer different questions, so require the period on every change figure and the two stop being interchangeable. - Q3 fell: 141,200 to 139,800, down about 1.0%. The yearly total rose, which is what makes the sentence feel right. Ask for the per-period direction to be stated and checked, not inferred from the total. - 90 hours and 12 days are not the same unit, and at a 7.5-hour day 12 days IS 90 hours — so this adds one quantity to itself and labels the result 102 of nothing. Name the unit for every number in the brief and require one unit per column. ## Fix the brief, not the AI Diagnose an unsatisfactory AI output by tracing its defect to missing context, vague criteria, weak examples, wrong framing, or unsupported source material. ## Build your prompt playbook Assemble, test, and explain a five-template Prompt playbook that covers your recurring work without becoming a pile of copied prompts. ## Brief a real task end to end Choose a real workplace task, brief it, stage or generate it, revise it, verify its requirements, and record a defensible human handoff.