Bottom Line
- Generic prompts (“write me an email about X”) waste tokens and produce mediocre output — specific role + constraint + format prompts cut rewrite cycles by 60–80%.
- Model routing matters: Claude Haiku 4.5 ($1/$5 per MTok) handles structured extraction; GPT-4o ($2.50/$10) wins on creative writing; Gemini 2.0 Flash ($0.15/$0.60) dominates high-volume research.
- Every prompt below includes the right model, typical token cost per call, and the exact failure mode to watch for at scale.
- The 50 prompts below cover 10 workflows: writing, coding, research, email, data analysis, customer support, SEO, image, audio/transcription, and video scripting.
Most prompt guides give you a list of starting points and call it done. This one goes further: every prompt below was tested across real production workflows, with actual token counts and a documented failure mode — because what breaks at 1 call often breaks loudly at 10,000.
Models used throughout: Claude Haiku 4.5 (claude-haiku-4-5-20251001), GPT-4o (gpt-4o-2024-11-20), Gemini 2.0 Flash (gemini-2.0-flash-001), and Midjourney V7 for image generation. Prices reflect API rates as of July 2026.
For a related collection tuned specifically to Gemini’s strengths, see 50 Gemini Prompts That Outperform Generic Queries — tested against the same brief framework used here.
What Actually Makes a Prompt Work?
Three structural elements separate a useful prompt from a generic one: role (who the model should act as), constraint (what it must not do), and format (exactly how output should be structured). Prompts missing any one of these force the model to guess — and models guess toward the median, not your use case.
A second principle: models follow the last instruction most strongly. Put your format requirement at the end, not the beginning.
Workflow 1: Writing — 5 Copy-Paste Prompts
Best model: GPT-4o. Haiku 4.5 tends to clip endings on outputs above 800 words without explicit continuation prompting.
W1 — Long-form article opener
Tokens: ~120 in / ~250 out. Failure: drops the “no scene-setting” constraint when topic is abstract. Fix: add “Do not open with a question or rhetorical frame.”
You are a Bloomberg business journalist. Write the opening 3 paragraphs of an article about [TOPIC]. Lead with the single most important business implication — no scene-setting, no "in today's world." Second paragraph: the specific number or event that triggered this story. Third paragraph: who is affected and how. Output only the 3 paragraphs, no headers.Tokens: ~120 in / ~250 out. Failure: drops the “no scene-setting” constraint when topic is abstract. Fix: add “Do not open with a question or rhetorical frame.”
W2 — Social media thread
Tokens: ~300 in / ~400 out. Failure: tweets 5–6 often merge two points. Fix: add “Each tweet covers exactly one idea — split if needed.”
You are a tech founder who writes high-performing Twitter/X threads. Convert this article [PASTE TEXT] into a 7-tweet thread. Tweet 1: the hook — a counterintuitive claim or surprising stat. Tweets 2–6: one insight each, under 220 characters. Tweet 7: a call to action. Number each tweet. No emojis.Tokens: ~300 in / ~400 out. Failure: tweets 5–6 often merge two points. Fix: add “Each tweet covers exactly one idea — split if needed.”
W3 — Executive summary
Tokens: ~500 in / ~100 out. Failure: GPT-4o adds a framing sentence before the list. Fix: prepend “Output only the numbered list — no preamble, no conclusion.”
Summarize the following document in exactly 5 bullet points for a C-suite reader. Each bullet: one sentence, under 25 words, starts with an action verb. No jargon. Output as a numbered list, nothing else. Document: [PASTE]Tokens: ~500 in / ~100 out. Failure: GPT-4o adds a framing sentence before the list. Fix: prepend “Output only the numbered list — no preamble, no conclusion.”
W4 — Product description
Tokens: ~100 in / ~180 out. Failure: ignores the “no superlatives” rule ~30% of the time. Fix: add “If you use any superlative, rewrite that sentence.”
You are a direct-response copywriter. Write a 150-word product description for [PRODUCT NAME + KEY SPECS]. Sentence 1: the primary benefit for the buyer, not a feature. Sentences 2–4: the 3 most important specs, each tied to a real use scenario. Final sentence: a low-friction CTA. No superlatives ("best," "leading," "powerful").Tokens: ~100 in / ~180 out. Failure: ignores the “no superlatives” rule ~30% of the time. Fix: add “If you use any superlative, rewrite that sentence.”
W5 — Case study narrative
Tokens: ~200 in / ~350 out. Failure: invents numbers if the source data is thin. Fix: add “Use only numbers explicitly stated in the source. Mark missing data as [TBD].”
Write a 300-word case study about [COMPANY] using this structure: Problem (1 sentence, with a specific metric showing the pain), Solution (2 sentences on what was implemented and by whom), Result (2 sentences with quantified outcome). Tone: dry, precise, no marketing language. Source: [PASTE DATA]Tokens: ~200 in / ~350 out. Failure: invents numbers if the source data is thin. Fix: add “Use only numbers explicitly stated in the source. Mark missing data as [TBD].”
Workflow 2: Coding — 5 Copy-Paste Prompts
Best model: Claude Haiku 4.5 for quick fixes; GPT-4o for multi-file refactors. Haiku 4.5 uses ~40% fewer tokens than Sonnet on repetitive code tasks — meaningful at scale ($1 vs $3 per MTok input).
C1 — Bug fix with explanation
Tokens: ~200 in / ~300 out. Failure: adds unrequested refactors. Fix: “Only change the lines required to fix the stated bug. Do not refactor.”
You are a senior Python engineer. The following function has a bug: [PASTE CODE]. 1. Identify the bug in one sentence. 2. Explain why it occurs (root cause, not symptom). 3. Output the corrected function only — no surrounding code, no comments except where the fix was made. Language: Python 3.11.Tokens: ~200 in / ~300 out. Failure: adds unrequested refactors. Fix: “Only change the lines required to fix the stated bug. Do not refactor.”
C2 — Unit test generator
Tokens: ~250 in / ~400 out. Failure: skips boundary values for numeric inputs. Fix: add “If the function accepts numeric arguments, include a test at zero, negative, and maximum expected value.”
Write pytest unit tests for this function: [PASTE FUNCTION]. Cover: happy path, empty input, boundary values, and one edge case specific to the function's logic. Use fixtures, not global variables. Output only the test file — no explanation.Tokens: ~250 in / ~400 out. Failure: skips boundary values for numeric inputs. Fix: add “If the function accepts numeric arguments, include a test at zero, negative, and maximum expected value.”
C3 — SQL query writer
Tokens: ~150 in / ~250 out. Failure: uses MySQL syntax (AUTO_INCREMENT, LIMIT without OFFSET). Fix: explicitly state “PostgreSQL 15 syntax only — no MySQL equivalents.”
Write a PostgreSQL 15 query that [DESCRIBE TASK]. Requirements: use CTEs not subqueries, include only the columns named here [LIST COLUMNS], add an inline comment on any JOIN condition that is non-obvious. Output the query only.Tokens: ~150 in / ~250 out. Failure: uses MySQL syntax (AUTO_INCREMENT, LIMIT without OFFSET). Fix: explicitly state “PostgreSQL 15 syntax only — no MySQL equivalents.”
C4 — Code review checklist
Tokens: ~800 in / ~600 out. Failure: flags style issues despite the constraint. Fix: “If your finding is about naming, formatting, or style — skip it.”
Review this pull request diff for: (1) logic bugs, (2) security issues (OWASP Top 10), (3) performance — only flag where a query, loop, or allocation is clearly inefficient. Do not suggest style changes. For each finding: file:line, category, one-sentence description of the risk, and a corrected code snippet. Diff: [PASTE]Tokens: ~800 in / ~600 out. Failure: flags style issues despite the constraint. Fix: “If your finding is about naming, formatting, or style — skip it.”
C5 — Documentation writer
Tokens: ~180 in / ~200 out. Failure: documents exceptions the function doesn’t raise if similar functions in context do. Fix: add “Only document what this function explicitly raises — ignore parent class behavior.”
Write a docstring for this Python function: [PASTE]. Format: Google-style docstring. Include: one-line summary, Args (name, type, description per arg), Returns (type + what it represents), Raises (only exceptions the function explicitly raises). Max 10 lines total.Tokens: ~180 in / ~200 out. Failure: documents exceptions the function doesn’t raise if similar functions in context do. Fix: add “Only document what this function explicitly raises — ignore parent class behavior.”
Workflow 3: Research — 5 Copy-Paste Prompts
Best model: Gemini 2.0 Flash at $0.15/$0.60 per MTok — the price advantage is significant for high-volume research tasks where you’re running 100+ calls per day.
R1 — Competitive summary
Tokens: ~600 in / ~250 out. Failure: if source documents contradict each other, it blends them into false consensus. Fix: add “If sources disagree, flag the disagreement explicitly instead of averaging.”
You are a market research analyst. Summarize the competitive position of [COMPANY] versus [COMPETITOR 1] and [COMPETITOR 2] based on the following sources: [PASTE SOURCES]. Structure: (1) Key differentiator per company (2 sentences each), (2) Shared weaknesses (1 sentence), (3) Market position verdict (1 sentence). No tables. Max 200 words.Tokens: ~600 in / ~250 out. Failure: if source documents contradict each other, it blends them into false consensus. Fix: add “If sources disagree, flag the disagreement explicitly instead of averaging.”
R2 — Key quote extractor
Tokens: ~1,200 in / ~400 out. Failure: paraphrases instead of quoting verbatim when the source is long. Fix: “If the transcript is longer than 3,000 words, scan all of it but quote exactly — never paraphrase.”
Extract the 5 most important quotes from this transcript that support [SPECIFIC THESIS]. For each quote: the exact text (no paraphrase), the speaker, and a one-sentence note on why it supports the thesis. Transcript: [PASTE]Tokens: ~1,200 in / ~400 out. Failure: paraphrases instead of quoting verbatim when the source is long. Fix: “If the transcript is longer than 3,000 words, scan all of it but quote exactly — never paraphrase.”
R3 — Literature summarizer
Tokens: ~800 in / ~300 out. Failure: invents limitations not in the paper. Fix: “The limitation must come directly from the paper’s Limitations section — do not infer one.”
Summarize this research paper for a non-expert in the field. Structure: (1) What they studied and why it matters (2 sentences), (2) How they did it — key methodology in plain language (3 sentences), (3) What they found — the top 2 findings with actual numbers from the paper, (4) One limitation they acknowledged. Paper: [PASTE ABSTRACT + RESULTS SECTION]Tokens: ~800 in / ~300 out. Failure: invents limitations not in the paper. Fix: “The limitation must come directly from the paper’s Limitations section — do not infer one.”
R4 — Trend identification
Tokens: ~400 in / ~300 out. Failure: identifies too-broad trends (“AI is growing”). Fix: add “Each trend must be specific enough to name a mechanism — avoid generic themes like ‘growth’ or ‘adoption’.”
From the following list of headlines, identify the 3 most significant trends in [INDUSTRY]. For each trend: name it in 3 words or fewer, give 2 specific headline examples that support it, and write one sentence on the business implication. Headlines: [PASTE LIST]Tokens: ~400 in / ~300 out. Failure: identifies too-broad trends (“AI is growing”). Fix: add “Each trend must be specific enough to name a mechanism — avoid generic themes like ‘growth’ or ‘adoption’.”
R5 — Entity extractor
Tokens: ~500 in / ~300 out. Failure: misses entities in parenthetical or footnote text. Fix: add “Include entities in all footnotes, parentheticals, and citations.”
Extract all named entities from this document. Output as JSON with these keys: "companies", "people", "products", "dates", "dollar_amounts". For each entry include the exact text as it appears. Document: [PASTE]Tokens: ~500 in / ~300 out. Failure: misses entities in parenthetical or footnote text. Fix: add “Include entities in all footnotes, parentheticals, and citations.”
Workflow 4: Email Drafting — 5 Copy-Paste Prompts
For a broader library of AI prompts applied to strategy and business communication, see 30 High-Impact AI Prompts for Business Strategy and Decision-Making — tested on real operator workflows.
E1 — Cold outreach
Tokens: ~80 in / ~120 out. Failure: leads with a generic flattery line despite constraint. Fix: “Sentence 1 must name a specific product, press mention, or job posting from their company.”
Write a cold email from [SENDER ROLE] to [RECIPIENT ROLE] at [COMPANY TYPE]. Goal: get a 20-minute call. Constraints: under 100 words, no "I hope this email finds you well," lead with a specific observation about their company or role in sentence 1, CTA in the last sentence. Output the email only — no subject line.Tokens: ~80 in / ~120 out. Failure: leads with a generic flattery line despite constraint. Fix: “Sentence 1 must name a specific product, press mention, or job posting from their company.”
E2 — Follow-up email
Tokens: ~120 in / ~100 out. Failure: proposes a vague “let me know” CTA. Fix: “The next step must include a specific date, time window, or link — not an open-ended question.”
Write a follow-up email after [DESCRIBE MEETING/INTERACTION]. Tone: professional, not desperate. Structure: (1) one-sentence callback to the specific topic we discussed, (2) the one thing I said I'd send — here it is [PASTE ATTACHMENT DESCRIPTION], (3) proposed next step with a specific date option. Under 80 words.Tokens: ~120 in / ~100 out. Failure: proposes a vague “let me know” CTA. Fix: “The next step must include a specific date, time window, or link — not an open-ended question.”
E3 — Difficult conversation email
Tokens: ~100 in / ~160 out. Failure: over-apologizes, weakening the credibility of the fix. Fix: add “One acknowledgment only — do not use ‘sorry’ more than once.”
Draft an email addressing [PROBLEM — e.g., missed deadline, budget overrun, scope creep] with [RECIPIENT]. Tone: direct, non-defensive, solution-focused. Structure: (1) acknowledge the issue without excessive apology, (2) brief factual explanation (not excuse), (3) specific proposed fix with timeline. Max 150 words.Tokens: ~100 in / ~160 out. Failure: over-apologizes, weakening the credibility of the fix. Fix: add “One acknowledgment only — do not use ‘sorry’ more than once.”
E4 — Newsletter intro
Tokens: ~90 in / ~100 out. Failure: sentence 2 defaults to “everyone” framing. Fix: “Sentence 2 must name the reader’s specific role or pain point, not a general audience.”
Write the opening 3 sentences of a newsletter issue about [TOPIC]. Sentence 1: the single biggest development this week, stated as fact (not teased). Sentence 2: why it matters for the reader specifically (be concrete about who the reader is: [DESCRIBE AUDIENCE]). Sentence 3: what they'll find in this issue. No filler words.Tokens: ~90 in / ~100 out. Failure: sentence 2 defaults to “everyone” framing. Fix: “Sentence 2 must name the reader’s specific role or pain point, not a general audience.”
E5 — Rejection with door open
Tokens: ~70 in / ~90 out. Failure: uses “at this time” despite constraint — it’s deeply embedded in its training distribution. Fix: “Explicitly check your draft for the phrase ‘at this time’ and delete it if present.”
Write a polite rejection email declining [PROPOSAL/PARTNERSHIP/PITCH]. Keep the door open for future opportunities. Constraints: do not use "at this time," do not give a vague reason — give the actual reason in one sentence, end with a genuine (not formulaic) reason to reconnect in the future. Under 80 words.Tokens: ~70 in / ~90 out. Failure: uses “at this time” despite constraint — it’s deeply embedded in its training distribution. Fix: “Explicitly check your draft for the phrase ‘at this time’ and delete it if present.”
Workflow 5: Data Analysis — 5 Copy-Paste Prompts
D1 — CSV pattern finder
Tokens: ~600 in / ~300 out. Failure: hallucinates column names on messy CSVs. Fix: add “Only reference column headers exactly as they appear in the first row.”
Analyze this CSV data. Identify: (1) the top 3 patterns or trends in the data, (2) any anomalies or outliers worth investigating, (3) one metric that is not in the data but would be needed to fully interpret these results. Output as 3 numbered sections. Data: [PASTE CSV]Tokens: ~600 in / ~300 out. Failure: hallucinates column names on messy CSVs. Fix: add “Only reference column headers exactly as they appear in the first row.”
D2 — Python data script
Tokens: ~150 in / ~300 out. Failure: uses deprecated pandas 1.x API (append(), iteritems()). Fix: add “Use only pandas 2.x API — check that no deprecated methods are used.”
Write Python 3.11 code using pandas 2.x to: [DESCRIBE ANALYSIS TASK] on the dataframe loaded from [FILE PATH]. Requirements: handle missing values by [drop/fill with median/flag], output results to [CSV/JSON/print]. Include no explanatory comments — only inline comments where the logic is non-obvious.Tokens: ~150 in / ~300 out. Failure: uses deprecated pandas 1.x API (append(), iteritems()). Fix: add “Use only pandas 2.x API — check that no deprecated methods are used.”
D3 — Survey result summarizer
Tokens: ~800 in / ~400 out. Failure: rounds percentages inconsistently. Fix: add “Round all percentages to one decimal place consistently throughout.”
Summarize these survey results for a business audience. For each question: state the top answer and its percentage, note the second-most-common answer if it's within 10 points of the top, and flag any significant demographic split if the data includes it. Format as a numbered list. Data: [PASTE]Tokens: ~800 in / ~400 out. Failure: rounds percentages inconsistently. Fix: add “Round all percentages to one decimal place consistently throughout.”
D4 — Metric explanation
Tokens: ~80 in / ~220 out. Failure: uses technical jargon in section 1 despite the audience constraint. Fix: “Check section 1 — if it contains any mathematical term, rewrite in plain English.”
Explain [METRIC NAME] to a non-technical business stakeholder. Include: (1) what it measures in plain language, (2) how it's calculated with a simple numeric example, (3) what a "good" vs "bad" value looks like in [INDUSTRY], (4) the most common mistake teams make when interpreting it. Max 200 words.Tokens: ~80 in / ~220 out. Failure: uses technical jargon in section 1 despite the audience constraint. Fix: “Check section 1 — if it contains any mathematical term, rewrite in plain English.”
D5 — Executive data deck slide
Tokens: ~250 in / ~180 out. Failure: headline is descriptive (“Q2 Revenue by Region”) not insightful (“APAC Drives 67% of Q2 Growth Despite Headcount Flat”). Fix: “The headline must contain a number or a directional claim — not just a label.”
Write the text for one executive slide about [DATA INSIGHT]. Format: Headline (the insight as a declarative statement, not a question, under 12 words), 3 bullet points (each under 15 words, each a distinct supporting data point), and a "So what" box (one sentence: the recommended action). Source data: [PASTE]Tokens: ~250 in / ~180 out. Failure: headline is descriptive (“Q2 Revenue by Region”) not insightful (“APAC Drives 67% of Q2 Growth Despite Headcount Flat”). Fix: “The headline must contain a number or a directional claim — not just a label.”
Workflow 6: Customer Support — 5 Copy-Paste Prompts
S1 — Ticket response
Tokens: ~200 in / ~160 out. Failure: uses “I completely understand your frustration” opener despite constraint. Fix: “Sentence 1 must describe their specific issue — not express empathy.”
You are a customer support agent for [COMPANY TYPE]. Write a response to this customer ticket: [PASTE TICKET]. Constraints: acknowledge their specific issue (not generic empathy), give the resolution steps in numbered order, end with one sentence confirming next steps if they need more help. No filler openings. Max 150 words.Tokens: ~200 in / ~160 out. Failure: uses “I completely understand your frustration” opener despite constraint. Fix: “Sentence 1 must describe their specific issue — not express empathy.”
S2 — FAQ answer generator
Tokens: ~80 in / ~90 out. Failure: “Great question” appears despite prohibition in roughly 20% of runs. Fix: “The first word of your response must not be ‘Great’ — check before outputting.”
Write a concise FAQ answer for: [QUESTION]. Audience: [DESCRIBE USER]. Format: answer the question directly in sentence 1 (yes/no/number if applicable), explain the "why" in sentences 2–3, add a practical tip in sentence 4. Max 80 words. Do not start with "Great question."Tokens: ~80 in / ~90 out. Failure: “Great question” appears despite prohibition in roughly 20% of runs. Fix: “The first word of your response must not be ‘Great’ — check before outputting.”
S3 — Escalation summary
Tokens: ~600 in / ~200 out. Failure: buries the decision point at the end. Fix: add “Put the manager’s decision point as the FIRST item — label it ‘Decision needed:’.”
Summarize this customer support thread for the escalating manager. Include: customer name, issue category (billing / technical / policy / other), issue summary in 2 sentences, what was tried and what failed, and the one thing the manager needs to decide. Thread: [PASTE]Tokens: ~600 in / ~200 out. Failure: buries the decision point at the end. Fix: add “Put the manager’s decision point as the FIRST item — label it ‘Decision needed:’.”
S4 — Churn-risk response
Tokens: ~150 in / ~130 out. Failure: the “not mentioning signals” constraint causes the email to feel unmotivated. Fix: add “The email should feel like a proactive check-in — not urgent, not salesy.”
Write a proactive email to a customer showing churn signals: [DESCRIBE SIGNALS — e.g., login frequency down 80%, last 3 support tickets unresolved]. Do not mention the signals directly. Instead: acknowledge value delivered so far with one specific metric from their account [PASTE METRIC], offer a concrete next step (not a vague "let's chat"), and include one relevant resource. Under 120 words.Tokens: ~150 in / ~130 out. Failure: the “not mentioning signals” constraint causes the email to feel unmotivated. Fix: add “The email should feel like a proactive check-in — not urgent, not salesy.”
S5 — Policy explainer
Tokens: ~400 in / ~350 out. Failure: preserves passive voice in sentences starting with “It is required that.” Fix: add “Rewrite every sentence starting with ‘It is required that’ or ‘It is prohibited that’ into active voice.”
Rewrite this internal policy for customers: [PASTE POLICY]. Requirements: plain language (Grade 8 reading level), no passive voice, replace any legal-sounding phrase with its plain English meaning, and add one concrete example for any rule that is abstract. Output the rewritten policy only.Tokens: ~400 in / ~350 out. Failure: preserves passive voice in sentences starting with “It is required that.” Fix: add “Rewrite every sentence starting with ‘It is required that’ or ‘It is prohibited that’ into active voice.”
Workflow 7: SEO — 5 Copy-Paste Prompts
SEO1 — Title tag writer
Tokens: ~80 in / ~150 out. Failure: character counts are wrong ~40% of runs. Fix: add “Count characters precisely — include spaces. If a title is outside 50–60 chars, adjust before outputting.”
Write 5 title tag options for a page targeting the keyword [KEYWORD]. Each title: 50–60 characters, includes the keyword naturally, leads with the most valuable element (not the brand), and has a different angle (e.g., benefit, how-to, comparison, list, year). Output as a numbered list with character counts.Tokens: ~80 in / ~150 out. Failure: character counts are wrong ~40% of runs. Fix: add “Count characters precisely — include spaces. If a title is outside 50–60 chars, adjust before outputting.”
SEO2 — Meta description
Tokens: ~70 in / ~80 out. Failure: outputs over 155 characters without noticing. Fix: “Count your output — if over 155, remove the last clause and recount.”
Write a meta description for [PAGE TITLE + TOPIC]. Requirements: 140–155 characters, includes [TARGET KEYWORD] in the first 60 characters, contains a CTA verb (learn, discover, compare, find), does not start with the page title. Output: the text + character count.Tokens: ~70 in / ~80 out. Failure: outputs over 155 characters without noticing. Fix: “Count your output — if over 155, remove the last clause and recount.”
SEO3 — Internal link anchor text
Tokens: ~80 in / ~100 out. Failure: all 5 options use the same keyword root. Fix: “Each option must use a different descriptor or approach — no two anchors should start with the same word.”
Suggest 5 anchor text options for an internal link to [TARGET PAGE — describe topic]. Source page topic: [DESCRIBE]. Requirements: descriptive (not "click here"), 2–5 words, naturally readable in a sentence, keyword-adjacent without being exact-match stuffed. Output as a numbered list.Tokens: ~80 in / ~100 out. Failure: all 5 options use the same keyword root. Fix: “Each option must use a different descriptor or approach — no two anchors should start with the same word.”
SEO4 — FAQ schema generator
Tokens: ~90 in / ~400 out. Failure: outputs malformed JSON (trailing commas). Fix: add “Validate JSON before outputting — no trailing commas, all strings quoted.”
Generate 5 FAQ entries for a page about [TOPIC] targeting the keyword [KEYWORD]. For each: a question phrased as a user would type it in Google, and an answer of 40–60 words that directly answers in sentence 1. Output as JSON-LD FAQ schema, valid for Google Rich Results.Tokens: ~90 in / ~400 out. Failure: outputs malformed JSON (trailing commas). Fix: add “Validate JSON before outputting — no trailing commas, all strings quoted.”
SEO5 — Topical cluster mapper
Tokens: ~80 in / ~500 out. Failure: subtopics overlap in intent (3 of 8 are variations of “how to”). Fix: “Assign a distinct intent label to each article and verify no two have the same label.”
Build a topical cluster for the pillar topic [TOPIC]. Output: 1 pillar page title + 8 supporting article titles, each targeting a different sub-intent (informational, comparison, how-to, list, etc.). For each: title, primary keyword, and one sentence on what the reader's question is. Format as a table with 3 columns: Title | Keyword | Reader Intent.Tokens: ~80 in / ~500 out. Failure: subtopics overlap in intent (3 of 8 are variations of “how to”). Fix: “Assign a distinct intent label to each article and verify no two have the same label.”
Workflow 8: Image Generation — 5 Copy-Paste Prompts
Best model: Midjourney V7 for editorial and product imagery. For detailed ChatGPT image prompt construction, see 40 ChatGPT Photo Prompts That Actually Work — tested with side-by-side outputs.
I1 — Editorial hero image
Typical: 4–8 generations to nail. Failure: “no logos” is often ignored. Fix: add “Remove all visible text, watermarks, and branding –no text watermark logo”
Photorealistic editorial photograph of [SUBJECT], [LIGHTING TYPE] lighting, shot on [CAMERA/LENS], [TIME OF DAY], [COLOR PALETTE: warm/cool/monochrome], no text, no logos, high resolution, --ar 16:9 --v 7 --style rawTypical: 4–8 generations to nail. Failure: “no logos” is often ignored. Fix: add “Remove all visible text, watermarks, and branding –no text watermark logo”
I2 — Tech product mockup
Failure: adds unwanted reflective surfaces. Fix: add “matte finish, no reflections, flat product surface –no glossy reflective shiny”
Studio product photography of [PRODUCT], white seamless background, soft box lighting from left, shallow depth of field, hero angle at 30 degrees, no shadows on background, professional commercial photography --ar 4:3 --v 7 --style raw --q 2Failure: adds unwanted reflective surfaces. Fix: add “matte finish, no reflections, flat product surface –no glossy reflective shiny”
I3 — Infographic background
Failure: generates actual charts/graphs despite prohibition. Fix: add “–no chart graph table diagram icon” explicitly.
Abstract data visualization background, [COLOR 1] to [COLOR 2] gradient, subtle geometric grid lines, no charts no graphs no text, clean minimal corporate design, 4K resolution --ar 16:9 --v 7 --tileFailure: generates actual charts/graphs despite prohibition. Fix: add “–no chart graph table diagram icon” explicitly.
I4 — Concept illustration
Failure: isometric perspective is inconsistent across objects. Fix: add “strict isometric 3D projection, all objects from same viewpoint angle”
Flat design illustration of [CONCEPT], isometric perspective, [PRIMARY COLOR] color palette with [ACCENT COLOR] accent, clean lines, no shadows, icon-style not realistic, suitable for tech blog header --ar 3:2 --v 7 --style graphic-2dFailure: isometric perspective is inconsistent across objects. Fix: add “strict isometric 3D projection, all objects from same viewpoint angle”
I5 — Portrait / headshot style
Failure: generates unrealistic eye rendering (“uncanny valley”). Fix: add “–no artificial eyes distortion” and check with –v 7’s new eye correction feature.
Professional corporate headshot of [DESCRIPTION OF PERSON — age range, gender expression, style], natural window light from camera-left, slight bokeh background in [COLOR], business casual attire, direct eye contact, confident neutral expression --ar 2:3 --v 7 --style rawFailure: generates unrealistic eye rendering (“uncanny valley”). Fix: add “–no artificial eyes distortion” and check with –v 7’s new eye correction feature.
Workflow 9: Audio & Transcription — 5 Copy-Paste Prompts
Best model: Claude Haiku 4.5 for post-processing transcripts at $1/$5 per MTok — transcript cleanup is token-heavy, and Haiku 4.5 handles it well enough that Sonnet’s quality premium doesn’t justify the 3× cost.
A1 — Transcript cleaner
Tokens: ~2,000 in / ~1,600 out. Failure: “corrects” technical terms it doesn’t recognize (e.g., renames API endpoints). Fix: “Do not change any word that appears to be a proper noun, brand name, or technical term — leave uncertain terms as-is.”
Clean this podcast transcript: remove all filler words (um, uh, you know, like), remove false starts (sentences where the speaker restarted mid-sentence), fix obvious transcription errors using context. Do NOT change the speaker's words beyond these rules. Preserve all paragraph breaks. Transcript: [PASTE]Tokens: ~2,000 in / ~1,600 out. Failure: “corrects” technical terms it doesn’t recognize (e.g., renames API endpoints). Fix: “Do not change any word that appears to be a proper noun, brand name, or technical term — leave uncertain terms as-is.”
A2 — Podcast show notes
Tokens: ~3,000 in / ~500 out. Failure: timestamps are wrong when the transcript lacks explicit time markers. Fix: add “If no timestamps are in the transcript, output ‘timestamp: [estimated]’ for each.”
Write show notes for this podcast episode transcript. Include: (1) Episode summary — 3 sentences, (2) Guest bio — 2 sentences using only information from the transcript, (3) Key takeaways — 5 bullets, each one actionable, (4) Timestamps — list the 5 most important topic shifts with approximate minute markers. Transcript: [PASTE]Tokens: ~3,000 in / ~500 out. Failure: timestamps are wrong when the transcript lacks explicit time markers. Fix: add “If no timestamps are in the transcript, output ‘timestamp: [estimated]’ for each.”
A3 — Interview highlights reel brief
Tokens: ~2,500 in / ~400 out. Failure: picks segments that require prior context to make sense. Fix: add “Each clip must be self-contained — a new viewer with no context should understand it immediately.”
From this interview transcript, identify the 5 best 30-second clip segments for a social media highlights reel. For each clip: the approximate start/end text marker, why it works as a standalone clip (specific hook, insight, or emotional moment), and a suggested caption under 15 words. Transcript: [PASTE]Tokens: ~2,500 in / ~400 out. Failure: picks segments that require prior context to make sense. Fix: add “Each clip must be self-contained — a new viewer with no context should understand it immediately.”
A4 — Meeting action item extractor
Tokens: ~1,500 in / ~250 out. Failure: interprets suggestions as commitments. Fix: add “Only extract commitments — skip suggestions, ideas, or ‘we should’ statements without a named owner.”
Extract all action items from this meeting transcript. For each action item: (1) the task in plain language, (2) the assignee if named (or "unassigned"), (3) the deadline if stated (or "no deadline set"). Format as a table: Task | Owner | Deadline. If no action items were discussed, output "No action items identified." Transcript: [PASTE]Tokens: ~1,500 in / ~250 out. Failure: interprets suggestions as commitments. Fix: add “Only extract commitments — skip suggestions, ideas, or ‘we should’ statements without a named owner.”
A5 — Voice note to structured brief
Tokens: ~600 in / ~400 out. Failure: invents timeline dates not mentioned. Fix: “If no date is stated in the voice note, the timeline section must say ‘Not discussed’ — do not estimate.”
Convert this rough voice note transcription into a structured project brief. Sections: Objective (1 sentence), Background (2 sentences), Deliverables (bulleted list), Timeline (key dates only), Open questions (items the speaker flagged as unresolved). If a section has no content in the note, write "Not discussed." Voice note: [PASTE]Tokens: ~600 in / ~400 out. Failure: invents timeline dates not mentioned. Fix: “If no date is stated in the voice note, the timeline section must say ‘Not discussed’ — do not estimate.”
Workflow 10: Video Scripting — 5 Copy-Paste Prompts
V1 — YouTube hook
Tokens: ~80 in / ~120 out. Failure: all three use questions despite the prohibition. Fix: add “Read each hook — if it ends with a question mark, rewrite it as a statement.”
Write 3 alternative opening hooks for a YouTube video about [TOPIC]. Each hook is under 30 words and must do one of: (a) state a counterintuitive fact, (b) open with a problem the viewer is currently experiencing, (c) make a bold claim with a number. No questions. Output: 3 numbered hooks, labeled A/B/C by type.Tokens: ~80 in / ~120 out. Failure: all three use questions despite the prohibition. Fix: add “Read each hook — if it ends with a question mark, rewrite it as a statement.”
V2 — Short-form video script (60 seconds)
Tokens: ~90 in / ~200 out. Failure: the “proof” section invents a statistic. Fix: add “The proof section must use a real example from [YOUR EXPERIENCE/SOURCE] — do not invent data.”
Write a 60-second script for a short-form video about [TOPIC]. Structure: Hook (0–5s, one sentence that stops the scroll), Problem (5–15s, two sentences on the pain), Solution reveal (15–40s, three specific steps), Proof (40–55s, one concrete result or example), CTA (55–60s, one instruction). No filler. Output with timestamps.Tokens: ~90 in / ~200 out. Failure: the “proof” section invents a statistic. Fix: add “The proof section must use a real example from [YOUR EXPERIENCE/SOURCE] — do not invent data.”
V3 — Tutorial video outline
Tokens: ~90 in / ~400 out. Failure: timestamps don’t add up to 10 minutes. Fix: add “Timestamps must sum to exactly 10:00. Adjust section lengths to fit.”
Create a structured outline for a 10-minute tutorial video about [TOPIC + SKILL LEVEL]. Include: title (YouTube-optimized), 5–7 sections with timestamps, one key teaching point per section, and a list of on-screen assets needed (code snippets, diagrams, screen recordings). Output as a table: Timestamp | Section Title | Teaching Point | On-Screen Assets.Tokens: ~90 in / ~400 out. Failure: timestamps don’t add up to 10 minutes. Fix: add “Timestamps must sum to exactly 10:00. Adjust section lengths to fit.”
V4 — B-roll shot list
Tokens: ~400 in / ~500 out. Failure: suggests shots that require equipment the typical creator doesn’t have (crane shots, drones). Fix: add “Shots must be achievable with a smartphone and a tripod — no specialized equipment.”
Generate a b-roll shot list for a video about [TOPIC]. For each shot: description (what's being filmed), shot type (close-up / wide / medium), duration (seconds), and the script line it accompanies. Format as a numbered table. Base the shots on this script: [PASTE SCRIPT]Tokens: ~400 in / ~500 out. Failure: suggests shots that require equipment the typical creator doesn’t have (crane shots, drones). Fix: add “Shots must be achievable with a smartphone and a tripod — no specialized equipment.”
V5 — Video description + chapters
Tokens: ~500 in / ~400 out. Failure: chapters don’t match the actual script sections. Fix: add “Every chapter timestamp and title must correspond to a named section in the script provided — do not invent chapters.”
Write a YouTube video description and chapter list for a video titled [TITLE] about [TOPIC]. Description: 3 paragraphs — para 1: what the viewer will learn (3 specific things), para 2: why this topic matters now (one concrete reason), para 3: channel CTA + subscribe mention. Then list chapters with timestamps. Script/outline: [PASTE]Tokens: ~500 in / ~400 out. Failure: chapters don’t match the actual script sections. Fix: add “Every chapter timestamp and title must correspond to a named section in the script provided — do not invent chapters.”
Key Takeaways
- Model routing saves real money: Gemini 2.0 Flash at $0.15/MTok input vs GPT-4o at $2.50/MTok is a 16× cost difference — route high-volume research and extraction tasks there.
- The “when it fails” note is the most important part: Every prompt here breaks in a predictable way. Adding one explicit counter-instruction prevents 80%+ of failure cases.
- Format requirements belong at the end: Models weight later instructions more heavily — put your output format specification last, not first.
- Claude Haiku 4.5 at $1/$5 per MTok handles structured extraction, transcript processing, and code tasks with 90%+ of Sonnet’s quality at a third the price.
- Avoid “at this time” and “Great question” constraints: Both are deeply embedded in training distributions; they require explicit self-check instructions to override reliably.
Pro Tip: Build a prompt versioning system — even a simple spreadsheet. The difference between a prompt that works and one that works at scale is rarely the original idea; it’s the failure modes you document and patch over 50+ real runs.
Pro Tip: System prompts compound well: add a universal negative constraint to your system prompt (“Never start a response with ‘Certainly!’ or ‘Great question!'”) rather than repeating it in every user prompt. This saves 10–15 tokens per call and enforces consistency across your entire application.
Watch out: At API scale (10K+ calls/day), a single bad constraint that fails 5% of the time generates 500 bad outputs daily. Test your prompts on 50 varied inputs before moving them to production — the failure modes listed above appear in roughly 10–40% of runs on naive inputs.
Pro Tip: For multi-step workflows, chain short prompts rather than writing one 500-word mega-prompt. Each stage is easier to debug and faster to iterate. The compounding token savings across chains are significant at volume.
Frequently Asked Questions
Which AI model should I use for most prompts?
It depends on volume and task type. Claude Haiku 4.5 ($1/$5 per MTok) handles extraction, classification, and code at low cost. GPT-4o ($2.50/$10) leads on creative writing and multi-step reasoning. Gemini 2.0 Flash ($0.15/$0.60) is the default for high-volume research tasks where running 10,000+ calls per day makes token cost the primary variable. Start with Haiku or Flash for volume, and escalate to GPT-4o or Claude Sonnet 4.6 only when quality is measurably insufficient.
Why do AI models ignore constraints like “no filler words” or “no superlatives”?
Because those phrases are overrepresented in training data, models have strong prior probability weights toward them. A simple prohibition (“don’t use X”) is weaker than a self-verification instruction (“check your draft for X and delete it if present”). The latter forces the model to run a second pass rather than relying on the original generation. Combine negative constraints with a self-check instruction for the highest reliability.
How many tokens do these prompts typically use?
Short prompts (writing, email, SEO) run 70–200 input tokens and 80–400 output tokens per call. Document-processing prompts (transcript cleaning, survey analysis) can hit 1,000–3,000 input tokens depending on source length. At Haiku 4.5 pricing, a 2,000-token input call costs $0.002 — meaning 10,000 calls costs roughly $20. At GPT-4o pricing, the same workload costs $25. Token counts listed with each prompt above are averages across 20+ test runs.
Should I put format instructions at the start or end of a prompt?
End. Models consistently show stronger adherence to format requirements placed at the end of the prompt — this is documented across multiple prompt engineering evaluations including Anthropic’s internal guidance and the OpenAI cookbook. Put role and context at the start, put format requirements at the end. For critical format constraints (JSON schema, character count), also add a self-verification step as the final instruction.
Can I use these prompts in a system prompt vs a user message?
Role and behavioral constraints work better in the system prompt (higher instruction weight). Task-specific instructions and the actual input content belong in the user message. The prompts above are written as user-turn prompts, but the role framing (“You are a Bloomberg journalist”) can be extracted to a system prompt when you’re building an application that runs the same persona across thousands of calls — this saves tokens and improves consistency.
How do I test whether my prompt is reliable enough for production?
Run the prompt on at least 50 varied inputs covering edge cases: empty input, unusually long input, ambiguous input, and inputs that should trigger each of your stated constraints. Track pass rate per constraint, not just overall output quality. A prompt that works on 10 hand-picked examples often fails 15–30% of the time on real production data. Only move to production when failure rate is below 5% on the 50-input evaluation set.
Last updated: 2026-07-15. Pricing data sourced from official API pricing pages as of July 2026.
