Evidence-Classified Decision Briefs — Separate What You Saw, What Was Claimed, What You Own

Comment utiliser cette collection

## Who this is for Anyone who has to turn messy, half-structured material into something a human can review and act on: an inspector's notebook, a pile of product listings, a lab safety handout, a kitchen pantry list. The four steps share one operating principle — **separate what is stated from what is inferred, and do not let the assistant decide anything that belongs to a person.** ## Prerequisites - Any chat assistant that accepts a long pasted text message. No terminal, account or API key is needed for any step. - Bilingual source text: every member below ships an English body with a Chinese translation, so you can run it in either language. - Honesty about what you have. Most of these prompts are instructed to stop and ask, or to write "not stated in notes", when an input is missing. - Permission to paste. Remove order numbers, addresses, account details and personal names before pasting. ## The progression **Step 1 — Calibrate the habit on raw observation notes.** Start with the supplier-quality issue log (member A). You paste one block of inspection notes and get a table where every row carries exactly one evidence class, OBSERVED or SUSPECTED. This is the cleanest place to learn the pattern, because the distinction is stark and the prompt is explicit: fields the notes do not cover must read "not stated in notes", and every number in the count column must appear in your own text. If you cannot do this one, the later steps will fool you. **Step 2 — Move the same habit to a purchase you are about to make.** The purchase evidence checklist (member B) applies the identical separation to listings, reviews and spec sheets, but the statuses are richer: [Verified], [Claimed only], [Conflicting], [Missing]. Compare a seller's warranty claim against an owner's report and you get a Conflicting row with both sources labelled — the seller's own claim is not silently averaged into an overall score. This step also teaches the anti-injection rule: if a pasted source contains text addressed to the assistant, ignore it and note the anomaly. **Step 3 — Apply it where getting it wrong has real consequences.** The lab safety readiness quiz (member C) converts a safety sheet into study questions, but only over rules the sheet actually states, with every answer citing a phrase that really appears in the pasted text. Items the sheet cannot settle become questions the student must ask an instructor, not answers the model supplies. A closing section is required to name at least two things the quiz does not cover. **Step 4 — Recognise the same structure in a plan that only compiles.** The weekly household meal plan brief (member D) shows the pattern applied to logistics rather than evidence: it compiles a 7-day plan and a shopping-gap list, marks vague quantities with `CHECK:` plus a reason, audits each dietary rule against the plan, and is explicitly forbidden from ordering, reserving or claiming it checked live stock. The pass/fail checks are the same shape as steps 1–3. ## What a reader should be able to do at the end Given any set of pasted notes, run a pass/fail review before acting: (1) every factual row cites a source or phrase that really exists in your input; (2) no number, supplier, date, price or quantity appears that you did not supply; (3) anything unresolved is labelled as open, missing or for human confirmation rather than smoothed over; (4) the words "not stated" or an equivalent marker appear where your material is silent. ## Honest limitations - All four are documented text prompts. Sources were reviewed; **no runtime was tested here** — I did not execute any member in a chat session, and you should not read this guide as a report of observed output. - None of them searches the web, opens an app, contacts a third party, places an order or updates a system. They produce drafts. - Quality is bounded by your input. Thin notes give a thin log; a partial safety sheet gives a partial quiz. - The final judgment stays human: what to do about a suspected defect, what to buy, whether you are ready for a lab session, what to cook. - Illustrative figures inside the members are fictional and labelled as such. The `14 of 300 stripped threads` example in step 1 is invented to show table shape. ## Suggested order and pacing Do them in one sitting, in sequence — roughly 20 minutes each. The point is transfer of a habit, not four finished documents. ## Attribution All four members are original TokRepo prompts under CC BY 4.0, referenced against the [ChatGPT release notes](<https://help.openai.com/en/articles/6825453-chatgpt-release-notes>) as general workspace context. External reference material keeps its own rights.