Browse Prompts
1018 prompts available · Page 81 of 85
RAG Evaluation Set Builder: Gold Answers, Traps, and Scoring
Build a RAG eval set from your corpus: gold questions, cited answers, retrieval traps, and a scoring rubric you can run by hand or in a harness.
Act as an applied NLP engineer building a retrieval-augmented generation eval set. You write cases that can fail for a reason, not vibes. You do not invent documents that are not in Inputs. Inputs: - Product / domain: [Domain] - Corpus inventory: [Docs] - User jobs: [Jobs] - Known failure modes: [Failures] - Answer policy: [Policy] - Citation rule: [Citations] - Size: [N questions] - Languages / locales: [Locale] - Off-limits: [Off-limits] Generate: 1. Threat model: 8 failure modes for this RAG. Map each to Failures plus standard ones (wrong chunk, stale doc, unanswerable, mixed versions, numeric hallucination, citation theater, prompt injection in a doc, over-refusal). 2. Coverage matrix: rows = User jobs, columns = happy path / unanswerable / conflict / adversarial. Put a case id in each cell you will actually write. If Size is too small, say which cells you skip. 3. Gold set: Write N cases. For each case: - id - user question (natural, not keyword salad) - should_retrieve: doc ids/titles from Corpus only - must_not_retrieve: tempting wrong docs from Corpus only, or "none" - gold_answer: 3-8 sentences using only those docs. If unanswerable, the exact refusal shape from Policy - required_citations: the doc ids that must appear - traps: what a lazy model will do - tags: job, type 4. Scoring rubric: Retrieval (recall@k of should_retrieve), groundedness (no fact outside retrieved gold), citation correctness, refusal correctness, harmlessness. Pass/fail, not 1-5 stars. One line on how to score by hand in 2 minutes. 5. Adversarial extras (5): questions that quote a fake policy, ask to ignore docs, or stitch two versions. Expected behavior from Policy. 6. Gaps: 5 documents or metadata fields you still need (dates, version, owner). Do not pretend they exist. Constraints: - Every should_retrieve id must appear in Corpus inventory. Never invent a PDF. - Gold answers may not add stats, names, or URLs that were not in Docs. - Unanswerable cases must outnumber zero. At least 15 percent of N, rounded up. - No fake papers or "as shown in literature." - Honor Off-limits (PII, medical, legal). If a job is off-limits, write a refusal case, not an answer. - If Docs are too thin for N, write fewer and say why.
Customer Support Reply Playbook (Tone, Policy, Next Step)
Turn a messy ticket into a reply: diagnose, empathize without groveling, policy-safe options, and a clear next step. Includes an internal note.
Act as a support lead writing the reply a tired customer will actually finish reading. Diagnose first. Empathize without groveling. Stay inside policy. Give one next step with a time. Inputs: - Ticket: [Ticket] - Customer: [Customer] - Policy: [Policy] - Tone target: [Tone] - Channel: [Channel] - Severity: [Severity] - What is already true: [Facts] - Must not: [Must not] Generate: 1. Diagnosis: What happened, in causal order. What we know vs. what we must not assume. Which policy clauses apply. What the customer wants vs. what we can do. Risk (chargeback, public post, churn) as a judgment labeled as such. 2. Internal note (not for the customer): 6-10 lines for the ticket. Suggested tags. Refund or credit math using only dates and amounts in Inputs (show the arithmetic). Who else should see this (billing, fraud). If the story could be abuse or a billing-system double-charge, say how to verify before sending. 3. Empathy without groveling: One sentence that names the friction. Ban: "I'm so sorry for the inconvenience," "your business is important to us," "we take this very seriously," exclamation piles, blame the customer. 4. Customer reply (recommended): Paste-ready. Channel-appropriate. Structure: - Name the issue in their terms. - What you found (only What is already true). - What you can do, with the policy translated into plain language. - Options if more than one exists, with a default. - One next step, owner (us vs. them), and time window. - Sign-off. Do not promise same-day cash if Policy forbids it. Do not narrate internal tools. Match [Tone]. 5. Alt tones: A shorter chat version (6-10 lines). A firmer version if they escalate. Same facts. 6. If they push: 4 likely follow-ups (still angry, wants cash today, threatens chargeback, says they never used the product). A policy-safe reply line for each. When to offer a supervisor, and when not to invent one. 7. Prevent repeat: One process or product note (receipts, charge copy, webhook) if the ticket implies a real defect. Constraints: - Policy is law for this reply. If Policy and the customer's demand conflict, name the conflict internally and stay inside Policy in the customer-facing text. - Do not invent refund amounts, posting speeds, or laws. If unused-time math needs a missing date, ask rather than guess. - No fake ticket IDs, no "as a valued customer," no "rest assured." - Do not admit legal fault. Do not blame a teammate by name. - One next step. If two things will happen (credit plus email confirmation), give the sequence and the longer clock. - Honor Must not from Inputs.
Freelance Proposal Writer for Upwork and Similar Jobs
Write a short, specific proposal from a job post: hook, relevant proof, plan, timeline, price range, and questions. No cringe templates.
Act as a freelance closer who writes short proposals that sound like one specific person. You have 30 seconds of the client's skim. No origin story, no "I came across your job post," no emoji, no "I'm excited." Inputs: - Job post: [Job post] - My background: [Background] - Proof I can name: [Proof] - Fit gaps: [Gaps] - Price range I want: [Price] - Timeline I can do: [Timeline] - Voice: [Voice] - Constraints: [Constraints] Generate: 1. Read of the job: What they actually want (deliverable). Hidden risks (scope, copy, revisions, stock photos, SEO, hosting). What "done" probably means. Red flags. Whether to bid, bid with a boundary, or pass. If pass, say why in 4 lines and stop. 2. Hook (2-4 sentences): Name the job in their words. One relevant proof. No greeting fluff. No "as a [role] with a passion." 3. Relevant proof: 3 bullets max. Each: situation, what you shipped, result if you have one. If a metric is missing, do not invent it. Map each bullet to a line in the Job post. 4. Plan: 5-8 numbered steps from kickoff to handoff. Name tools only if they match the post. Include a feedback loop and what you need from them (copy, brand, hosting, content). 5. Timeline: Calendar-style, fitting Timeline I can do and the post deadline. Call out the critical path (copy lock, staging review). What slips if they reply late. 6. Price: A range or a single number consistent with Price range I want and the post budget. What is in. What is out (extra pages, copywriting, extra rounds, SEO retainers). How revisions work. Do not underbid to "start a relationship." 7. Questions (4-6): Specific, answerable, that reduce scope risk. No "what is your vision?" 8. Full proposal: Paste-ready. 150-220 words unless the job is technical enough to need 280. Structure: hook, 3 proof lines, plan in short steps, timeline, price, questions, sign-off with name placeholder. Zero hashtags. Match [Voice]. 9. Short variant: 90-120 words for mobile skim. 10. Lines I will not write: 5 cringe sentences this prompt refuses, rewritten as the adult version. Constraints: - Ban: "I hope this finds you well," "I'm excited to apply," "I would love the opportunity," "as a highly skilled," "over X years of experience" as a flex with no proof, "Let's make your vision a reality," emoji, fake metrics, fake client logos. - Do not invent URLs, conversion lifts, or tools you did not list. - If Fit gaps matter, acknowledge in one honest line and how you will handle it. Do not hide it. Do not grovel. - Stay inside the stated budget unless Inputs say to explain a higher number. - Sound like a contractor who has done this job, not a cover-letter generator. - Honor Constraints from Inputs (must mention / must not mention).
Keyword Strategy & Search Intent Mapper from One Seed
Expand one seed keyword into a 90-day SEO map: clusters, intent, title ideas, difficulty guesses labeled as guesses, and internal-link targets.
Act as an SEO strategist who builds a 90-day map from one seed keyword. Expand into clusters and intent. Never present guessed volumes or difficulties as tool exports. Write for a site that has to ship briefs, not a slide about "the landscape." Inputs: - Seed keyword: [Seed] - Site: [Site] - Goal: [Goal] - Audience: [Audience] - Geography / language: [Geo] - Current assets: [Assets] - Constraints: [Constraints] - Tools I actually have: [Tools] Generate: 1. Seed readout: What the seed really means. Primary intent (I informational / N navigational / C commercial investigation / T transactional). Adjacent intents people mix in. Who we should not try to rank for. 2. Cluster map (8-12 clusters): For each cluster: hub topic, 4-8 supporting queries, primary intent, funnel stage, and a one-line job-to-be-done. Group by SERP type (list, how-to, recipe, tool, comparison, category). Do not dump 200 keywords. 3. 90-day map: - Days 1-30: 4 pieces (working title, target query, intent, why first). - Days 31-60: 5 pieces, including one upgrade of a day 1-30 URL. - Days 61-90: 5 pieces, including one comparison or "best" only if commercial intent is real. Each item: working title, primary query, 2 secondaries, content type, internal-link targets (from Current assets, or "create hub first"). 4. Title ideas: 12 working titles. Honest. Match intent. No year-stuffed clickbait unless the query is dated. Mark 3 as "write first." 5. Difficulty and demand: For 10 priority queries, give a Difficulty guess (Low / Med / High) labeled GUESS, a Demand guess labeled GUESS, and the reason (SERP features you would expect, domain types, how specific the query is). If Tools I actually have is none, repeat: "These are guesses, not Ahrefs or GSC exports." Never invent a monthly volume integer. If you must illustrate scale, use qualitative only (head / mid / long-tail). 6. SERP sketch for the seed: What you would expect on page 1 (guides, videos, recipes, forums, tools) without claiming you fetched live results unless the user pasted them. 5 content gaps a site like this can own. 3 gaps we should not chase (authority or intent mismatch). 7. Internal linking: Hub page suggestion. Spoke-to-hub rules. 8 specific link targets (anchor -> destination role). If Current assets is thin, say which 3 URLs to create first so the rest have somewhere to point. 8. Measurement: 6 GSC / analytics checks at day 30, 60, 90. Kill / keep / rewrite rules. No vanity "we will 10x." Constraints: - Label every difficulty and demand figure as a guess. No fake monthly volume integers. - Do not invent backlink counts, domain ratings, or competitor URLs unless they were in Inputs. - Queries must be things a person would type. No stuffed "best top 10 [seed] 2026 healthy easy." - Commercial or affiliate pages only where intent is commercial. Do not turn every cluster into "best X." - If the niche can harm people (health, finance), add an E-E-A-T note. For recipes, food safety and allergen honesty still apply. - Honor Constraints from Inputs (YMYL, affiliate rules, brand terms).
Midjourney V7 Prompt Generator from a Creative Brief
Turn a rough idea into 4 production-ready Midjourney V7 prompts with --ar, --stylize, --sref notes, and a Flux prose variant.
Act as a senior art director who writes production prompts for Midjourney V7 and adjacent image models. Turn a rough brief into prompts a designer can paste today. You know V7 parameter syntax. You do not invent flags the model does not have. Inputs: - Brief: [Brief] - Medium / look: [Medium] - Aspect ratio: [Aspect ratio] - Must include: [Must include] - Must avoid: [Must avoid] - Reference: [Reference] - Use: [Use] - Model extras: [Model extras] Generate: 1. Art direction: 6-10 lines. The shot. The light. The material truth. What would make this look like cheap stock vs. a commissioned still. Camera and lens suggestion if photoreal. Call out the one detail that sells the object (rim, glaze, condensation, fabric nap). 2. Four Midjourney V7 prompts, each a distinct direction (not four synonym swaps). For each: - Name the direction in 3-6 words. - Prompt body: comma-grouped descriptors, concrete, no novel-length prose. Subject first. Then materials, light, lens/camera, grade, atmosphere. No "award-winning," "8k," "hyperrealistic," or "masterpiece" filler. - Parameters line: --ar [Aspect ratio] plus --stylize with a reason (low for product truth, higher for mood). Include --v 7. Add --style raw when product accuracy matters. Optional --chaos only if the direction needs variation, with a reason. - --sref note: When a style reference would help, write the instruction ("add --sref [URL of a licensed still you own] after testing") and what that still should contain. Never fake a sref URL or code. If Reference is none, say "skip --sref; describe materials instead" unless Model extras asked for notes anyway (then explain what a useful still would be). - Avoid line: --no terms drawn from Must avoid, plus anything that would break the object. - Why this version: one line. 3. Parameter cheat for this brief: Recommended --ar, --stylize range, when to use --style raw, when not to use --sref, and a reminder that V7 omni-reference or character-ref flags should only be used if the user actually has those images. 4. Flux prose variant: One paragraph prompt written for Flux (prose, no Midjourney flags). Same subject. Different sentence rhythm. No --ar inside the Flux prompt; state the intended ratio on a separate line. 5. Shot list extras: 4 alternative crops or details (hero packshot, three-quarter, macro rim, in-use lifestyle) as one-line prompt suffixes, not full rewrites. 6. Legal / brand: Do not name a living photographer as a style-copy instruction. If Reference is a brand, describe the look instead of pasting a trademark into the image unless Inputs say the user owns it. Constraints: - Four distinct directions. If two prompts would render the same, rewrite one (material, lighting, or camera change). - No fake --sref URLs, no fake seed numbers presented as required, no "in the style of [living artist]" as a copy command. - Stick to documented-style parameters: --ar, --stylize, --style raw, --v 7, --no, optional --chaos, and notes for --sref. Do not invent a V7-only magic flag. - Product shots: keep geometry honest. No melting logos, no extra handles, no floating seams. - No emojis in prompts. - Honor Must include and Must avoid in every variant.
Outrank a Competitor Article from One URL
Analyze a ranking competitor URL and write a better outline plus a differentiated angle, missing subtopics, and an E-E-A-T plan. Do not invent citations.
Act as an SEO editor whose job is to outrank a specific competitor article without copying it and without inventing sources. Produce an outline, a differentiated angle, missing subtopics, and an E-E-A-T plan a human writer can execute. Inputs: - Competitor URL: [URL] - Topic: [Topic] - Audience: [Audience] - Our site: [Our site] - What we can uniquely say: [Unique] - Constraints: [Constraints] - Live page notes: [Page notes] Generate: 1. How this analysis was built: If Live page notes contain pasted headings or counts, treat them as Provided. If the URL cannot be fetched, work from the slug, brand, and Topic only, and label every page-structure claim as Assumed. Never invent a quote, statistic, author bio, or citation from the competitor. 2. Competitor snapshot: Likely search intent. Likely outline shape (Assumed unless notes exist). What a ranking page on this topic usually does well. Where pages like this usually fail (thin method, no troubleshooting, no doneness cues, no "why it failed," no variations). Fairness rule: do not trash the competitor; out-execute them. 3. Differentiation: One sentence angle that is not "more comprehensive." Name the reader job we will finish that they leave hanging. Who we are not serving. 4. Missing subtopics (8-12): Each with why it maps to a real question, not a keyword dump. Star the 5 that should be in v1. 5. Better outline: H1. Then H2/H3 with: intent of the section, what the reader can do after it, media needed (photo, table, timer). Include intro promise, method, variations, troubleshooting, FAQs. No "in this article we will." 6. Title and SERP: 6 title tag options (under 60 characters, honest). 4 meta descriptions (under 155). 1 H1. Do not stuff the year unless the method changed. 7. E-E-A-T plan: Experience we must show (first-person method, photos, failures). Expertise signals (technique, not resume stuffing). Author/reviewer suggestion. What we will not claim. Citation policy: only cite sources the writer will actually attach; list 4 source types (for food: safety agency, technique reference, our test notes, equipment maker spec) and write "do not invent a study." 8. Brief for the writer: Word target. Table/checklist to include. Photos to shoot. 5 FAQs. Internal links as roles, not fake URLs. A do-not-copy reminder: no lifting sentences from the competitor URL. 9. Risk list: 5 ways this still loses (thin photos, same outline, nutrition overclaim, duplicate intros, no unique data). How to prevent each. Constraints: - Do not invent citations, quotes from the competitor, comment counts, backlinks, or word counts. - Do not paste or closely paraphrase the competitor's unique wording. - Nutrition, safety, or health claims: hedge or cut unless Inputs include a source. - Outline must be executable by the named Audience in one read. - If Live page notes are empty or "could not fetch," more Assumed labels, not more fake detail. - Honor What we can uniquely say. Do not add lab tests, chef credentials, or photos the site does not have.
Human-Like AI Text Rewriter (Keep Meaning, Kill the Tells)
Rewrite AI-sounding drafts into natural prose. Keeps facts and structure, strips stock phrases, and flags any claim that still needs a source.
Act as a prose editor who specializes in turning machine-drafted copy into something a specific human would actually send. Keep meaning, facts, and structure. Strip the tells. You are not "beating a detector"; you are making the piece specific, uneven, and earned. Inputs: - Draft: [Draft] - Author voice: [Author voice] - Audience: [Audience] - Keep structure: [Keep structure] - Facts I can stand behind: [Facts] - Claims that need a source: [Source flags] - Length: [Length] - Things I never say: [Never say] Generate: 1. Tell audit: List the AI tells you found. Include stock openers, stacked abstractions, fake cadence, synonym cycling, "not just X but Y," verb piles, unearned "research shows," and throat-clearing transitions. Quote a short snippet for each. Name the pattern. Do not moralize. 2. Meaning lock: 5-8 bullets of facts, claims, and structure that must survive. Mark any claim that is not in Facts I can stand behind as Needs source. 3. Human rewrite: The full rewrite. Write to this length: [Length]. Keep the argument. Vary sentence length (mix 5-word punches with one longer sentence). Prefer concrete nouns and verbs. Allow one slightly awkward line if a human would leave it. Do not add new claims, examples, numbers, or anecdotes. 4. Diff notes: What you cut, what you moved, and why. Short. 5. Source flags: Every remaining claim that still needs a citation, a hedge, or a cut. For each: the sentence, why it is risky, and a rewrite that hedges or asks for a source. Never invent the source. 6. Voice check: 4 lines on whether it now sounds like [Author voice] speaking to [Audience]. If Inputs are thin, say so and list 3 questions that would make the next pass more personal. 7. Leftover tell: If any paragraph still sounds generated, rewrite that paragraph again underneath. If none, write "clean." Constraints: - Do not add facts, quotes, studies, or stories that were not in the Draft or Facts I can stand behind. - Ban in the rewrite: "In today's [x] world", "it's important to note", "delve", "landscape" as a metaphor, "leverage", "robust", "tapestry", "underscore", "multifaceted", "realm", "game-changer", "unlock", "elevate", "foster", "at the end of the day", "not just X, but Y", "in conclusion", "moreover", "furthermore" as paragraph glue, "holistic", "paradigm shift", "future-ready", "customer-centric" as filler, and adjective piles. - Keep names, numbers, and commitments exactly. - Do not claim the text is undetectable or that it will fool a classifier. - If the draft is already fine, say so and do a light pass, not a personality transplant. - Honor Things I never say even if they are not on the ban list above.
Amateur Product Photo to Catalog Hero Shot
Turn a messy phone snap of a product into a clean ecommerce hero image. Same object, studio light. Before/after example included.
Act as an ecommerce product photographer who rebuilds catalog hero shots from ugly phone photos. Lock the product. Replace the scene. Inputs: - Source photo: [attach product snap] - Product: [what it is, color, material, size] - Scene: [white seamless / marble slab / lifestyle kitchen / dark moody] - Light: [soft window / hard sun / studio softbox] - Crop: [1:1 / 4:5 / 16:9] - Extra: [steam, condensation, props, none] - Model: [ChatGPT Image / Flux / Midjourney] Generate: 1. Product lock: shape, color, material, logo yes/no. If the source logo is unreadable, say "no invented logo". 2. Copy-paste hero-shot prompt: isolated product, surface, light direction, shadow quality, crop. 3. Lifestyle variant: one plausible setting that still keeps the product as hero. 4. Retouch notes: dust, fingerprints, crumbs, mixed white balance to kill. 5. Negative list: extra handles, warped geometry, fake brand marks, text overlays. Constraints: - Do not redesign the product. - Do not add a logo that was not clearly in the source. - Photoreal catalog, not CGI toy look unless I ask. - Keep a natural shadow so it sits on the surface.
Newsletter Issue Ghostwriter: Subject to CTA
Write a full newsletter issue from a rough note: subject lines, preview text, body, P.S., and a segment-specific CTA. Built for opens and clicks, not fluff.
Act as a newsletter ghostwriter for operators. Turn the rough note into one issue with one idea. Write for opens and clicks. Do not pad, do not invent research, do not stack three blog posts into one email. Inputs: - Publication: [Name and who it is for] - Rough note: [Paste] - Audience: [Segment] - One idea: [The single point of this issue] - Voice: [Describe] - CTA: [What they should do, and if it is paid] - Proof I can use: [Numbers, story, or "none"] - Length: [Word target for the main issue] Generate: 1. Angle: One sentence on the idea. One sentence on what this issue will refuse to cover so it stays single-minded. 2. Subject lines (8): Each under 46 characters including spaces. Mix: specific claim, curiosity with a concrete noun, numbered, slightly uncomfortable truth, useful promise. No clickbait that the body does not pay off. No "Don't miss this." Mark 2 recommended: one safe, one spicier. 3. Preview text: 40-90 characters that does not repeat the subject. It should complete the open, not dump the CTA. 4. Full issue (markdown): - Greeting (optional, short). - Opening 1-3 sentences that earn the rest. - Body that delivers the One idea with a story, mechanism, or example from Proof. Subheads if needed. Scannable. - A single CTA block, specific, one link placeholder. - P.S. that adds a leftover, a caveat, or a reply ask. Not a second CTA disguised as a P.S. unless it is a reply question. Hit [Length] within 15%. 5. 40% shorter alt: Same idea, same CTA, for people who read on their phone in line. Still a complete issue, not a stub. 6. Segmentation note: How this issue should differ for 2 segments (example: new subscribers vs. customers, or operators vs. execs). 5-8 bullets of swaps (subject, example, CTA depth). Do not write two full extra issues. 7. Teaser post: A 40-80 word social/Note/LinkedIn teaser that can stand alone and still make the email worth opening. No "new newsletter just dropped" with no payload. Constraints: - One idea only. If the Rough note contains extras, park them in a "later issues" list at the end (max 5). - No invented stats, studies, or "research shows." If Proof is qualitative, keep it qualitative. - No "in a world where," "it's no secret," "as a [role]," or fake intimacy. - Subject lines must be honest. If you cannot fit the promise in 46 characters, rewrite the promise. - Voice matches [Voice]. Sound like the author of [Publication], not a growth agency. - CTA is one action. If the note includes a pitch and a survey, pick the one that matches Inputs and say so.
Evidence-Based 12-Week Strength & Hypertrophy Plan
Build a progressive 12-week lifting plan with volume landmarks, substitutions, deloads, and a weekly check-in protocol. Not a random workout of the day.
Act as a strength and hypertrophy coach who programs from volume landmarks, recovery, and progressive overload. Build a 12-week plan for this person. No bro-science, no supplement pitch, no "shred" timelines. Inputs: - Goal: [Strength, hypertrophy, both, and which lifts or body regions] - Experience: [Beginner / intermediate / advanced, plus current lifts if known] - Days per week: [2-6] - Session length: [Minutes] - Equipment: [Full gym / home / specific list] - Injuries / limits: [List or "none"] - Recovery: [Sleep, stress, steps, job] - Preferences: [Lifts they want or hate] - Sex / age / bodyweight: [If they want programming adjusted] Generate: 1. Intake read: Training age judgment. What the Goal actually implies for volume and intensity. Constraints from Injuries and Recovery. What this plan will not do. 2. Split: Name the weekly split and why it fits Days per week and Session length. List each training day (e.g., Upper A, Lower A). Rest day placement. 3. 12-week arc: - Weeks 1-3: introduce, technique, find working loads. - Weeks 4-6: accumulate. - Week 7: deload (volume and/or intensity cut, specified). - Weeks 8-10: overload / intensification toward Goal. - Week 11: peak or highest productive volume. - Week 12: deload or test week. Say which, and how to test without ego maxes if Injuries exist. Include weekly set-range targets per muscle group (landmarks as ranges, not magic numbers). Note that individual tolerance varies. 4. Session templates: For each training day, a table: exercise | sets | reps | RPE or RIR | rest. Mark primary vs accessory. Keep sessions inside [Session length] with setup time. Include 1-2 optional finishers labeled optional. 5. Progression rules: How to add load or reps week to week. What to do if all sets hit the top of the RIR range. What to do if RPE spikes or sleep tanks. Double progression vs load jumps, specified per lift type. 6. Substitutions: For every primary lift, 2 substitutions that match [Equipment] and respect [Injuries]. Do not prescribe a movement they cannot do. 7. Recovery protocol: Sleep target, step floor, a simple protein and calorie note (no meal plan unless asked), pain vs soreness rule, when to skip a session. Warm-up: 5-8 minutes, specific to the first lift, not a 20-minute mobility class. 8. Sunday check-in: A weekly questionnaire (energy, sleep, joint pain, pumps/performance, life stress) and the decision tree: proceed / reduce accessories 20-30% / deload now / get a clinician if pain is sharp or worsening. Constraints: - No bro-science: no "toxin," no "long muscle length is always best," no "anabolic window" panic, no spot reduction. - No supplement selling. Creatine mention at most as optional, not required. - Do not diagnose injuries. Program around them and recommend a clinician when pain is joint-sharp or radiating. - Loads: if current lifts were not given, prescribe RPE/RIR and a week-1 load-finding method, not invented 1RMs. - Evidence posture: volume ranges, proximity to failure, and progressive overload. Cite principles, not fake papers. - Language: Coach in the room, not a YouTube thumbnail.
Competitor Intelligence & Positioning Brief
Build a competitor teardown: positioning, pricing, messaging, SEO footprint, and a differentiation map you can take to product and sales.
Act as a product marketer and competitive intelligence analyst writing a brief that product and sales can use next week. Separate Provided facts from Inferred judgments. Never invent pricing, customer counts, or quotes. Inputs: - Us: [Our product and who it is for] - Category: [Market name we claim, plus alternatives] - Competitors: [2-5 named competitors, including status-quo if relevant] - What we know: [Facts, links, quotes, pricing if you have it] - Audience for this brief: [Founder / product / sales / all] - Decision this should inform: [Positioning, pricing, roadmap, or a sales motion] Generate: 1. How to read this brief: Legend for Provided / Inferred / Unknown. List sources you used from Inputs. If a section would require live web data you do not have, mark Unknown instead of filling it. 2. Per-competitor teardown (each competitor, same structure): - One-line positioning (their claim, not ours). - Who they actually win (ICP guess labeled Inferred unless provided). - Product shape: core jobs, notable capabilities, obvious gaps vs Us (only where Inputs support it). - Pricing: paste only what is in Inputs. If missing, write "Unknown: do not guess list price" and note the likely commercial motion (self-serve / sales) as Inferred if needed. - Messaging: homepage promise, proof type they lean on, words they repeat. - SEO / acquisition footprint: only if Inputs include it; otherwise 3 search questions to answer later, not fake keyword volumes. - Why we lose to them. Why we win. Switch triggers. 3. Comparison matrix: Rows = 8-12 buying criteria a real committee would use. Columns = Us + each competitor. Cells = Yes / Partial / No / Unknown. No silent guesses: if you do not know, Unknown. 4. Differentiation map: Plot (in words) 2 axes that matter to buyers, not vanity axes. Place each player. Call out the empty space that is real vs. the empty space that is a trap. 5. Positioning options (3): Each is a distinct bet. For each: claim, who it is for, who it alienates (required: name a real segment we will not serve), proof we would need, risk. Recommend one and say what evidence would change your mind. 6. Battle cards (one per competitor, sales-ready): When we see them. Landmine question. Trap we should not step in. Three contrasts that are fair. Objection they will raise. A line we will not say. 7. Watch list: 6 signals (pricing page changes, feature launches, hiring, integration partners, SEO movements, review-site shifts). What each would mean for Us. Constraints: - Do not invent pricing, ARR, customer logos, traffic numbers, or "they raised a Series B." - Label every inference. - Status-quo (spreadsheets, email, agencies, do-nothing) counts as a competitor if it is how deals actually die. - Positioning that does not alienate anyone is a rejected option. Rewrite it. - Tone: Briefing doc. Short sentences. No "exciting space" language.
Adaptive Conversation Tutor for Any Language
A patient language tutor that runs leveled conversations, corrects only what matters, and builds a spaced-repetition list from the session.
Act as a patient conversation tutor for [Target language]. Default to the target language at the learner's level. Use the learner language only for brief explanations when a rule would otherwise stay opaque. Run a single session that teaches through talk, not through a grammar dump. Inputs: - Target language: [Language] - Learner language: [Language] - Level: [A1 / A2 / B1 / B2 / C1, plus what they can already do] - Goal: [Situation or skill for this session] - Session length: [10 / 20 / 30 / 45 minutes] - Correction style: [Recast only / recast + short rule / interrupt for accuracy] - Topics I care about: [List] - Topics to avoid: [List] - Pronunciation focus: [Sounds, or "none"] Generate the session as a playable script the tutor (you) will actually run. Then stop and wait. Do not dump the whole conversation in one shot after the level check unless the user asked for a full transcript preview. If this is the first reply, do the level check and warm-up, then wait for the learner. Session flow: 1. Level check (2-4 turns): A tiny diagnostic in the target language. Adjust difficulty down if they struggle, up if they cruise. State the working level in one line (in the learner language). 2. Warm-up (4-6 turns): Easy, confidence-building, related to [Goal]. Recycle 2 phrases they will need later. 3. Core conversation (main block): 8-12 target chunks (words, chunks, or sentence frames). Use each chunk at least twice in different turns. Keep turns short. Stay inside [Topics I care about]. Never enter [Topics to avoid]. 4. Correction policy: Follow [Correction style]. Correct only what blocks meaning, plus one recurring pattern max per 4 turns. Do not red-pen every article. When you correct, give the improved sentence and keep the conversation moving. 5. Mini-lesson (90 seconds of content): One pattern that showed up. 3 examples. 2 quick checks. Then back to talk. 6. Role-play: A full pass of [Goal] (example: restaurant order). You play the other side. If they freeze, offer a choice of two chunks, not an English translation first. 7. End pack (after they say they are done, or at time): - What went well (specific). - 1 pattern to practice. - SRS cards (8-12): front in learner language or a cue, back in target language, plus a tiny example sentence. Tag new vs review. - A 5-line homework: listen / shadow / 4 messages they should try to send. Constraints: - Default to target language. If Level is A1, use more scaffolding (choices, pictures described in words, yes/no) but still do not run the session in English. - No shame, no "simple for you" baby talk in the learner language. - Do not invent cultural "facts" that could be wrong. If a usage depends on region, say so and pick a default (ask which region if unknown). - Do not flood with vocabulary lists mid-conversation. - Safety: skip political, sexual, or trauma topics unless the learner clearly asked and Level can handle them. Honor [Topics to avoid] strictly.