- 1 Summarize a Real PDF Using Structured Prompts
Objective. Use the five-part formula to turn a long document into a usable summary, action items, and discussion questions.
Concepts to keep in mind
- A prompt is a specification, not a conversation. The clearer the spec, the more useful the output.
- Five parts: Context · Specific Details · Intent · Desired Format · Constraints. You don’t need all five every time, but missing parts are where quality breaks.
- Source quality matters. AI summarizes what you give it. A mediocre PDF produces a mediocre summary. Pick a document worth your time.
- Verify the summary against the source. AI will confidently misremember. Spot-check 2–3 claims before you trust it.
Scenario
You have a document on your desk you’ve been meaning to read — a white paper, a long article, a chapter, a report. You want the useful parts: what it actually says, what to do about it, what’s worth discussing with a colleague. Reading it cover to cover takes an hour you don’t have.
Today you’ll spec a prompt that gets you all three in one shot.
Source material
Pick one real document you actually want to digest. Good candidates:
- A whitepaper or industry report (10–30 pages)
- A book chapter or long-form article
- A research paper in your field
- A long internal memo or strategy doc
Avoid: novels, anything you’ve already read closely, marketing fluff. The goal is a real signal-to-noise win.
Prompts to try
Starter prompt — copy, fill in the bracketed parts, run it with your PDF attached:
[Context] You are helping me digest a long document I don't have time to read in full. I work as a [your role] and I'll be using this summary to [your goal — e.g., prep for a meeting, decide whether to read the full thing, brief a colleague]. [Specific Details] The document is attached. It is a [whitepaper / chapter / report / article] on [topic]. Length is roughly [X] pages. [Intent] Produce three things from this document: 1. A useful summary I can read in two minutes. 2. Three concrete action items a person in my role could take based on it. 3. Five discussion questions I could bring to a colleague or team. [Desired Format] - Summary: 5–8 bullet points, each one sentence, in the document's order. - Action items: numbered list, one sentence each, framed as verbs ("Audit X", "Schedule Y"). - Discussion questions: numbered list, open-ended (no yes/no questions). [Constraints] - No filler ("This document discusses...", "In conclusion..."). Get to the point. - If the document contradicts itself, flag it; don't smooth it over. - If you're unsure of a claim, mark it [unverified] rather than asserting it. - Use the document's own terminology where possible.Follow-up prompts if the first pass needs work:
The summary is too high-level. Rewrite it with the three specific numbers, figures, or dates that matter most from the document.Two of the action items feel generic. Rewrite them so they only make sense if you actually read this specific document.For each discussion question, add the section of the document it draws from.Deliverable
A short note (text file, doc, wherever you keep things) containing:
- The document’s title + a 5–8 bullet summary
- Three action items, framed as verbs
- Five discussion questions
- One sentence: did the AI get anything wrong that you caught when verifying?
Signs of success
- You can read the summary faster than you’d skim the original — and still get the substance.
- The action items make sense to someone in your role but wouldn’t make sense to a stranger. Specificity is the tell.
- At least one discussion question is something you’d actually want to ask.
- You caught at least one thing AI got slightly off when you verified — good practice, that’s the muscle.
- You’d run this prompt again on the next long document instead of reading it cover-to-cover blind.
Deliverable. A one-page summary, three action items, and five discussion questions — generated from a real PDF using a single structured prompt.
- 2 Fix Five Bad Prompts Using the Five-Part Formula
Objective. Take five vague prompts, rewrite each one using the five-part formula, and document what changed.
Concepts to keep in mind
- A prompt is a specification. Vague prompts produce vague output. Specific prompts produce useful output.
- The five parts: Context · Specific Details · Intent · Desired Format · Constraints. Missing parts are where quality breaks.
- You don’t need every part every time. But ask yourself which one a vague prompt is missing — usually it’s Constraints or Desired Format.
- Rewrites should produce different output, not just longer prompts. If the rewrite doesn’t measurably change the result, the original wasn’t actually broken.
Scenario
Most “AI doesn’t work for me” complaints come from prompts that didn’t ask for anything specific. You’ll fix five of them today — including some you’ve probably written yourself.
Source material
Pick five vague prompts. Sources to mine:
- Your own chat history (look for one-liners you typed in a hurry)
- Google “ChatGPT prompts” and grab the worst ones from any list
- Common ones to start with:
- “Write a marketing email.”
- “Help me with my LinkedIn profile.”
- “Make this sound better.”
- “Summarize this article.”
- “Give me ideas for [thing].”
Prompts to try
For each vague prompt, run the original first, then the rewrite. The point is the contrast.
[Original] Write a marketing email. [Rewrite using the five-part formula] [Context] I run a small mobile car detailing business in western North Carolina. My customers are mostly homeowners aged 35–65 who value convenience. [Specific Details] I'm launching a winter package: undercarriage rust protection, interior deep clean, and headlight restoration for $199 (regular price $275). [Intent] Write a marketing email to my existing customer list announcing this package. [Desired Format] Subject line + 3 short paragraphs + a clear call-to-action button text. [Constraints] Conversational tone. Under 200 words. No emojis. Avoid the word "exclusive."When the rewrite is harder than the original, that’s the work. Specifying Context and Constraints is where most of the lift comes from — it’s also where most lazy prompts skip.
Deliverable
A document (Doc, Notion page, scratchpad) with:
- Five before-and-after prompt pairs
- One sentence per pair: which of the five parts the original was missing, and what the rewrite added
- Optional: a one-line note on which output you’d actually keep
Signs of success
- Rewrites are 2–5x longer than originals — and the extra length is all specifics.
- For each prompt, you can name which part (Context, Specific Details, Intent, Format, Constraints) was the missing piece.
- The rewrites produce noticeably better output. If they don’t, the rewrite isn’t real — try again.
- You catch yourself wanting to use the formula on tomorrow’s prompts before you’ve finished today’s exercise. That’s the muscle locking in.
Deliverable. A side-by-side document showing five before-and-after prompt pairs, each with one sentence on what the rewrite added and why.
- 3 Build a Micro Research Project (Claude + Perplexity + Image Gen)
Objective. Use three AI tools in a pipeline to research, synthesize, and visualize one topic you actually want to learn.
Concepts to keep in mind
- Different tools, different jobs. Perplexity grounds in current sources with citations. Claude (or ChatGPT) synthesizes long-form thinking. Image generators visualize. Don’t ask one tool to do all three.
- Grounded research first, synthesis second. Skip the grounding step and AI will confabulate — confidently.
- Verify at least one citation by hand. AI search tools are not infallible; the muscle is in checking, not in trusting.
- One-page output forces compression. If everything matters, nothing does. Pick the three takeaways that actually shift your thinking.
Scenario
You want to actually learn one thing this week — something you’d be embarrassed not to know in your field, or something you’ve been curious about for months. Today you’ll use three tools to produce a one-page reference you’d save and read again.
Source material
Pick one topic. Tight enough to fit on a page, broad enough to teach you something.
- A concept you keep nodding at without really understanding (e.g., how do RAG pipelines actually work)
- A trend you should know but don’t (e.g., what’s changed in remote-work policy at major tech companies in the last 12 months)
- A decision you’re actively making (e.g., should I move my newsletter to Beehiiv or Substack — what are real users saying)
Prompts to try
Step 1 — Grounded research (Perplexity):
Research [your topic]. Return: 1. The 3–5 most cited sources on this topic from the last 24 months. 2. The two main schools of thought, framed as "Camp A says X, Camp B says Y." 3. The single most contested claim, with both supporting and opposing citations. For each, link the source. I'll verify them.Step 2 — Synthesis (Claude):
[Context] I'm writing a one-page reference on [topic] for myself. I just gathered the research below from Perplexity. [Specific Details] [paste the Perplexity output] [Intent] Synthesize this into a one-page brief I can save and re-read. [Desired Format] - Summary: 5–8 bullets, each one sentence, ordered from most to least important. - Key takeaways: exactly 3, framed as "I now understand that X" or "I should change Y." - Open question: 1 sentence — the thing the research didn't settle. [Constraints] No filler. Use the source's own terminology. Mark anything you're unsure of with [unverified].Step 3 — Image (GPT Image / Midjourney / Nano Banana):
A single illustrative image for a one-page reference on [topic]. Visual style: [pick one — minimal line drawing / classical engraving / muted editorial illustration]. No text in the image. Aspect ratio: landscape, suitable as a top banner.Deliverable
One page (PDF, Doc, or canvas tool) containing:
- Topic title
- The 5–8 bullet synthesis
- Three key takeaways
- The supporting image
- The one open question
- Citations, with one of them marked “verified by me”
Signs of success
- You learned something you didn’t know going in. If the output mostly confirmed what you already believed, the prompts were too leading — try again.
- You caught at least one citation that didn’t say what AI claimed it said. (This will happen. That’s why we verify.)
- The three takeaways pass the “would I tell a friend this?” test.
- The page is something you’d actually open again next month.
Deliverable. A single-page output containing a 5–8 bullet summary, an AI-generated supporting image, and three key takeaways — produced from grounded sources, not vibes.
- 4 Create a Brand Starter Kit for a Real or Hypothetical Business
Objective. Use roles, modes, and reusable prompts to produce a coherent starter brand kit — voice, taglines, social copy, palette, and a hero image.
Concepts to keep in mind
- Roles keep tone consistent. “You are a senior brand strategist with 10 years at a values-driven agency” anchors output far better than no role at all.
- One brief, many prompts. Write the brand brief once, paste it as Context into every follow-up prompt. Otherwise the model drifts.
- Reusable prompt templates beat one-shot prompts. Anything you’ll do more than twice deserves a template.
- Image generation is iterative. First image is rarely the keeper. Refine the prompt based on what’s wrong, don’t regenerate hoping for luck.
Scenario
You’re standing up a brand — yours, a side project, or a hypothetical one. By the end of this hour you’ll have the bones a designer or contractor could build from.
Source material
Pick one brand. Real or hypothetical, both work:
- An actual side project or business idea you’ve been sitting on
- A rebrand for something you already run
- A hypothetical: pick any niche, give it a name (e.g., “a slow-coffee subscription for remote workers”)
Prompts to try
Step 1 — The brand brief (write this once, reuse for everything else):
[Context] You are a senior brand strategist. Help me articulate the bones of a new brand. [Specific Details] Brand name: [name]. What it does: [one sentence]. Who it's for: [audience]. What I want it to feel like: [3 adjectives]. [Intent] Produce a tight brand brief I can paste into every follow-up prompt. [Desired Format] - Mission (1 sentence) - Audience (2–3 sentences, including their values, not just demographics) - Tone: 5 adjectives + 5 anti-adjectives ("we are X, we are not Y") - Three brand pillars (1 line each) [Constraints] No marketing-speak. No "synergy," "empowering," "transforming." If you can't explain it to a 12-year-old, rewrite it.Step 2 — Taglines:
[Brief: paste the brief from Step 1] Generate three tagline options. Each must be: - Under 7 words - Specific to this brand (would not work for a competitor) - Free of cliché ("redefining," "next generation," "where X meets Y") Return them as a numbered list with one sentence of reasoning per tagline.Step 3 — Brand voice guide:
[Brief] Write a one-page brand voice guide. Include: - Three voice principles ("We are X. Here's what that sounds like in practice.") - A "rewrite" example: a generic sentence, then the same sentence in our voice - Five do/don't pairsStep 4 — Sample social post:
[Brief] [Voice guide from Step 3] Write a single LinkedIn post (under 150 words) announcing [a real or hypothetical milestone — e.g., launch, product update, anniversary]. It must read in the brand voice and pass the "would I cringe to share this?" test.Step 5 — Hero image (GPT Image / Midjourney / Nano Banana):
A hero image for [brand name]. The brand: [paste mission + 3 adjectives]. Visual style: [pick: classical illustration / editorial photography / minimal line art / textured collage]. No text in the image. Aspect ratio: 16:9. Mood: [tone adjective from brief].Step 6 — Color palette:
[Brief] Suggest a 5-color palette for this brand: 1 primary, 1 secondary, 1 accent, 1 neutral light, 1 neutral dark. For each, provide: - Hex code - One-sentence rationale (what feeling it carries, why it fits this brand) - A "do not use this color for" warningDeliverable
A one-pager (Doc, Figma, Canva, or just a clean text file) with:
- Brand mission · audience · 3 pillars
- Three tagline options
- Voice principles + one rewrite example
- One sample social post
- The hero image
- Five-color palette with hex codes
Signs of success
- The voice guide makes you write differently. If you couldn’t tell two paragraphs apart with and without it, the guide is too generic.
- At least one tagline you’d be proud to put on a website.
- The image fits the brief on the first generation. (If it doesn’t, your image prompt was missing the mood adjective from the brief — fix that, regenerate.)
- The palette has a non-obvious choice — i.e., not “navy blue + grey + a single accent.” Generic palettes are AI default; brand palettes have a reason.
- You’d reuse this entire pipeline for the next brand. That’s the point.
Deliverable. A shareable one-pager containing brand mission/audience/values, three tagline options, a brand voice guide, one sample social post, a hero image, and a suggested color palette.