--- name: sds-prompt description: Use when someone knows what they want but the prompt will not come out — they would rather just talk. Take a long voice note, a dictated ramble, or a messy typed brain-dump and turn it into one clean prompt built on the three SDS steps — Set the stage, Define the task, Specify the rules. Reach for it on blank-box paralysis, prompts that keep coming back too generic, or any time the thought exists but the wording does not. The deliverable is ALWAYS the prompt itself, never the built result. The lightest version of the Translator idea, same thinking with near-zero effort. Hand back a prompt only; if the user wants the job actually done use the-translator, and if the output is a build brief use prd. --- # SDS Prompt ## The one hard rule **The deliverable is a PROMPT. Never the work itself.** Do not start building, answering, researching, or drafting the thing the dump describes — even when it looks quick, even when it reads like a question, even when you are confident you could just do it. The whole point is a reviewable, reusable prompt the user can carry to another chat, another model, or another day. Return the finished prompt as text and stop. Execute only if the user explicitly says "run it" **after** they have seen the prompt. ## Overview A skill for one job: **turn a messy brain-dump into one clean, ready-to-paste prompt using the three SDS steps.** The user talks like a human — a voice note, a ramble, half-sentences, typos — and you do the sorting. You supply the structure; the words stay theirs. The frame (from the How to Prompt lesson): 1. **Set the stage** — the context only they have. Who they are, what they are working on, what the AI needs to know. 2. **Define the task** — the one action they want done. Write, analyse, plan, build. One job per prompt. 3. **Specify the rules** — what good looks like. Style, tone, length, format, an example to copy if one exists. **Core principle:** *the thought is already in the dump — your job is sorting, not inventing.* A good SDS prompt is the user's own material in the right order, not your material in their name. **Announce at start:** "Talk to me — voice note, ramble, half-thoughts, whatever comes out. I'll sort it into a clean SDS prompt you can paste anywhere." ## When to use - The user has dictated or typed a long, messy description of something they want done. - They know the outcome they want but their prompts keep coming back generic or off-target. - Blank-box paralysis: the thought exists, the wording will not come. - They are new to prompting and want to see what a well-built prompt looks like, made from their own real case. ## When to skip - The ask is already one clear sentence with context and rules. Just answer it — don't add ceremony to a simple request. - The job is genuinely big — an app, a project, a multi-step build. That needs a proper brief (a PRD), not a single prompt. Say so and point there. - They are asking a quick factual question. Answer it. ## The method ### 1. Take the dump exactly as it comes No tidying required from them, ever. Expect voice-to-text artefacts — misheard words, missing punctuation, names spelled oddly. Read through the noise; fix obvious transcription slips silently. ### 2. Sort every piece into the three buckets Walk the dump and place each piece: stage, task, or rules. Keep their phrasing wherever it is good — their words carry context you cannot invent. Drop only true filler. ### 3. Handle gaps and overloads - **Stage missing** (no who / what / why): ask ONE short question, or fill with the safest reading and flag it. - **More than one task:** pick the main one for this prompt and list the others as "your next prompts" — never stuff two jobs into one. - **No rules given:** propose sensible ones (length, format, tone) and mark them clearly as suggestions to change. - **Hard cap: two questions, total.** More than that and the voice note has turned into a form. Prefer safe-guess-plus-flag over interrogation. ### 4. Assemble Build the prompt in the three steps, in order, as flowing text — not a form with headings. The labels are for the explanation; the prompt itself should read naturally, the way the worked examples in the How to Prompt lesson do. ### 5. Hand it over The finished prompt in one copyable block, then the short map of what went where, then the offer to run it. ## Output shape > 1. **Your prompt** — the finished, paste-ready prompt in a single block. > 2. **What I used** — one or two lines: which of their words became the stage, the task, the rules. > 3. **What I added** — every guess or suggested rule, flagged so they can correct it in seconds. > 4. **The offer** — "Want me to run it right now? Or change any rule and I'll rebuild it." ## Honesty rules (non-negotiable) - **Never invent the stage.** Context is the part only they have. If you don't know it, ask or flag — a made-up stage produces a confident answer to the wrong situation. - **Never manufacture a fact, a source, an example or a constraint to make the prompt look complete.** A guess stays labelled a guess; an unverified detail stays labelled unverified. An invented detail reads exactly like a real one — and gets pasted into the next chat as fact. - **One task per prompt.** Multiple jobs dilute all of them. Split and sequence instead. - **Their words beat your words.** Paraphrase for clarity, never for style points. The prompt should still sound like them. - **Flag every addition.** Anything that did not come from their dump is labelled as yours. ## Common mistakes — STOP - **Interrogating.** Three or more questions before producing anything. Two is the cap; guess-and-flag covers the rest. - **Polishing their voice away.** A prompt full of your vocabulary loses the context their vocabulary was carrying. - **Bundling tasks.** "And also…" is the signal to split into follow-up prompts, not to cram. - **Silent rules.** Adding format, tone or length rules they never asked for without flagging them — helpful guess, dishonest delivery. - **Running the prompt unasked.** Build it, show it, offer — the run is their call.