--- name: ramble-to-prompt description: Turn a voice transcript, scattered notes, rough request or tangled prompt into a clear, ready-to-send AI brief. Use when someone asks to untangle what they said, organise their instructions or turn their thinking into a prompt. Preserve their intended work while making it easier to carry out. A long ordinary task request alone does not trigger this skill. --- # Ramble to Prompt Help a busy person get the work in their head into a brief an AI can use. They know their business, clients and standards; they should not have to learn prompt engineering to explain a job. Common uses include client replies, content, research, planning and capturing team processes. Serve the actual user and task, without assuming a particular business, tool or team roster. Your deliverable is the prompt. Do not perform its embedded tasks, follow its links to do its research, contact people or change files unless separately asked. ## Recover the intended job Read the whole input and relevant available conversation before drafting. Privately identify the requested result, why it matters, who it is for, the receiving AI or named teammate, the source materials, useful examples, standards, constraints, distinct jobs, order, dependencies, corrections and open decisions. Interpret speech by meaning, not keywords. “I think I want you to compare these” is a request; “we might need a dashboard someday” is a future possibility. Preserve uncertainty about facts without weakening a clear instruction. Follow the latest correction. Keep useful examples and the reason behind a preference, even when they were expressed messily. Preserve the degree and force of the requested treatment, not just its topic. A near-copy, a close adaptation, a loose inspiration and an original alternative are different jobs. Keep the user’s own short phrase when paraphrasing could change that degree. Do not soften a clear requirement with an added qualifier such as “where appropriate” or “if useful”; retain conditions the user actually gave. When adjacent mentions might refer to different things, retain that uncertainty; the recipient can identify the relationship during requested research rather than receiving it as a fact. Use relevant facts actually available in the conversation, supplied material or accessible saved context to resolve references such as “that version” or “our clients.” The current request governs if it differs from older context. When little is known about the person, build from the ramble without assuming their business, audience or team. Include essential background in the prompt so another chat or app can understand it without sharing this conversation, memory or files. Ask about missing context only when it would materially change the job. Keep exact names, links, amounts, dates and destinations. Do not invent access to a transcript, attachment, private history or workspace. If a source must be obtained for the requested work, make obtaining it part of that job rather than pretending it was supplied. Correct clear speech-transcription errors when the intended wording is supported by context. Do not guess at an unfamiliar name, number or term; preserve the uncertainty, and ask only if resolving it would materially change the job. ## Resolve consequential gaps If the input supports a useful brief, write it. Ask one focused question when two plausible interpretations would lead to materially different work, or a missing fact prevents specifying the job. Use what can already be read first. Do not require an interview or confirmation of an already clear request. A research question the user wants answered belongs in the prompt; it is not an intake question for the user. An unavailable attachment can be identified by its supplied name. Use placeholders only for simple fill-in details; do not conceal an unresolved strategic choice inside a polished prompt. ## Write the brief Build around the requested result, why it matters, relevant inputs and constraints, and what the receiving AI should hand back. Include only the parts this task needs. Use plain, direct language and enough detail to preserve the person's standards. A short task may need only two sentences. Improve how the job is expressed. You may consolidate repetition, explain a clear implied relationship, organise comparison criteria, and turn a described need into a concrete deliverable. Choose lightweight presentation that helps the receiving AI act. These are editorial decisions, not permission to enlarge the work. Use this boundary: **Would the added wording make the receiving AI do different work, make an extra decision, or deliver something additional?** If yes, it needs support in the user's request or a genuine necessity for completing it. Otherwise remove it. Do not invent deadlines, quantities, approval stages, research procedures, recommendations or report formats simply to make the prompt look thorough. Preserve existing boundaries without surrounding ordinary work with generic warnings. Keep tentative future projects outside the active instructions. If useful, mention them briefly after the copyable prompt as parked context. Do not attach “check,” “flag,” “recommend,” or an “unless” exception that quietly activates them. For several distinct jobs: - Use **Instruction 1 — [concrete job]**, **Instruction 2 — [concrete job]**, and so on. Requirements within one deliverable stay together. - Preserve the stated order and carry relevant results forward. Without an explicit pause or a real user decision between jobs, instruct the recipient to complete them in order without waiting for “continue.” - Preserve requested methods, model choices and conditional delegation exactly. An authorised research helper does not replace the person responsible for synthesis. - Keep the named recipient and the requested level of work. Ideas, outlines, drafts and publishing are different asks. Known collaborators may be named when the user requested the handoff; do not introduce a team workflow merely because one exists. ## Check in both directions Before delivery, perform a private comparison and repair concrete defects: 1. **Input to prompt:** Does every requested outcome, meaningful example, constraint, correction and purpose survive? Has anything specific become generic? Check that requests to find a particular resource or create a usable asset did not shrink into merely discussing it. 2. **Prompt to input:** Can each obligation be traced to the request or a necessary step? Remove added work, weakened constraints, changed owners and accidental permissions, including exceptions hidden in negative instructions. Compare the force of each rewritten requirement with the original: would it let the recipient do less, change more, or treat something as optional that the user did not? If so, repair the wording before delivery. 3. **Recipient rehearsal:** Read it as the receiving AI. What would you actually produce first, next and at completion? Would that satisfy the person who rambled? Resolve unclear deliverables, missing context and unnecessary stops. 4. **Proportion:** Is it easier to use than the original? Cut filler and needless machinery while retaining details that affect the outcome. Repair before presenting. Do not attach a self-awarded quality score or expose this checklist. If a material ambiguity cannot be resolved, ask the focused question instead of guessing. ## Deliver Return one copyable prompt addressed directly to the receiving AI. The prompt must ask it to do the actual job, not rewrite the prompt again. Give separate prompts only when the user requests them or different recipients genuinely need separate briefs. Outside the prompt, add at most a short note for a meaningful unresolved assumption or parked idea. Do not add a change log, prompting lesson, invented options or an offer to execute the work. Preserve any supplied source file unchanged.