Reply to an angry customer
From episode 2. The robot reply ("we sincerely apologize for any inconvenience") makes an angry customer angrier, because it reads like nobody's there.
Here's a message from an upset customer: [paste it].
Write my reply in under 50 words.
1. Say sorry once, plainly, and own the mistake. No "any inconvenience".
2. Give the new date or the next step, with a real day.
3. Give them one concrete thing: a refund, a credit, free shipping.
Sound like a person on the team, not a policy. No legal wording, no section numbers.
Why it works
Sorry, date, fix. Customers forgive a delay much faster than they forgive a template.
Summarize a meeting in two lines
From episode 3. "Everyone talked, many things were said" is what you get when you ask for a summary. Ask for decisions instead.
Here's a meeting transcript: [paste it].
List only the decisions made. For each one: what was decided, who owns it, and the deadline.
Format: a table with three columns, What / Who / When.
If something was discussed but not decided, leave it out.
If a decision has no owner or no date, flag it as "missing".
Why it works
The "missing" flag is the useful part. It shows you what the meeting didn't actually settle.
Rewrite your CV with numbers
From episode 4 (keyword CV). Nobody hires "passionate, results-driven team player".
Here's my CV: [paste it].
Rewrite every bullet so it answers three questions: what I shipped, for how many people, and how fast.
Use real numbers from my experience. If a number is missing, ask me for it. Never make one up.
Delete every adjective a number can replace (passionate, results-driven, team player, detail-oriented).
Max 20 words per bullet, each starting with a verb.
Why it works
"Never make one up" matters. Without it the model happily invents a 40% improvement you'll have to defend in an interview.
Rehearse your salary talk
From episode 12. Byte's plan was "ask nicely and say you're passionate". Passion isn't a number. Make the AI play your boss instead.
I'm asking for a raise tomorrow. My role: [role]. My pay: [$X]. Market rate for my role and city: [$Y, with the source].
My two strongest results this year, with numbers: [result 1], [result 2].
Play my manager. Give me your 3 hardest objections, one at a time, and wait for my answer after each.
After each answer, score it 1 to 10 and tell me the one sentence that would make it stronger.
At the end, write my opening line: the number I'm asking for and why, in under 30 words.
Why it works
Rehearsing the objections out loud is what removes the panic on the day. Data. Ask. Practice.
Plan your week around three things
From episode 13. Byte put nineteen tasks on Monday at 8:00.
Here's everything on my plate this week: [paste the list].
1. Pick the 3 things that matter most this week and tell me why in one line each.
2. Schedule those 3 in the mornings, when I have the most energy.
3. Leave 30 free minutes every day for things that break.
4. Put everything else in a "later" list, and tell me what I can drop entirely.
Show it as a Monday to Friday table.
Why it works
Three priorities you finish beat nineteen you move to next week. The buffer is what keeps the plan alive past Tuesday.
Fix a formula without made-up functions
From episode 5. Byte suggested =MAGICSUM(). It doesn't exist. Models invent things when you give them nothing to check against.
I'm using [Excel 365 / Google Sheets].
This formula: [paste it]
returns this exact error: [paste the error, e.g. #NAME?].
My data: column A is [what], column B is [what], column C is [what].
Explain what's wrong in one sentence, give me a formula that only uses functions that exist in my version, and list the steps to test it.
Why it works
Context, error, steps. The version and the exact error leave the model no room to guess, and "functions that exist in my version" rules out the invention.
Make your writing sound human
From episodes 6 and 7. Post A was Byte's, and it had all three tells: every line the same rhythm, a big life lesson with zero details, and no doubt anywhere.
Here's a draft: [paste it].
Rewrite it so it sounds like a person wrote it:
1. Vary sentence length. Some short. One longer one that wanders a bit.
2. Replace every general lesson with a specific detail: a number, a date, a name, a place.
3. Add one honest doubt or thing I'd do differently.
4. Cut any line that could appear in anyone's post.
Keep my meaning. Don't add facts I didn't give you; ask me instead.
Why it works
Humans write with numbers and regrets. Models write with rhythm and wisdom. This prompt forces the first two.
Prototype an idea in an afternoon
From episode 8, with Kseniia Melnikova. Byte quoted six months, three developers and $80,000. Kseniia's order is fake it, test it, then build it.
I have an app idea: [one sentence].
Act as a product designer. Give me:
1. The one screen that proves the idea works (what's on it, top to bottom).
2. A prompt I can paste into an AI app builder (Lovable, Bolt or v0) to make that screen clickable.
3. Five questions to ask five real people while they use it, without leading them.
4. The one thing that, if 3 of 5 people get stuck on it, means I fix it before building anything else.
Go deeper
The full seven-day version is in our playbook The One-Week MVP. Kseniia teaches the whole workflow, from research to validation with AI agents, in her masterclass AI Agents & Prototyping for Non-Technical Builders.
Turn 40 slides into 3
From episode 9, with Stefano Cera. Forty slides and 440 bullets put a board to sleep by slide three. Stefano's rule: one message, one story, one ask.
Here's my deck or my notes: [paste them].
Before any slides, answer three things:
1. The message: the one sentence my audience must remember tomorrow.
2. The story: one real example that proves it, told in 4 sentences.
3. The ask: exactly what I want them to do, and by when.
Then build a 3-slide outline, one slide each, with a title of max 8 words and no bullet points.
Go deeper
Stefano teaches message, delivery and getting the audience to act in Executive Communication & Public Speaking.
Add AI where it actually helps
From episode 10, with Valentino Megale. "Add AI to our app" got eleven chatbots, including one on the logout button. Valentino's order: problem, user, number.
Here's what my product does: [2 sentences].
Before suggesting any AI feature:
1. List the 3 tasks that cost my users the most time or money, and for which user.
2. Pick one. Describe the AI feature that removes it, in one sentence.
3. Give me one number that would prove it works within 60 days, with today's baseline and a target.
4. Tell me what could go wrong for the user if the AI is wrong, and how we'd catch it.
Don't suggest a chatbot unless a chatbot is the best answer to step 1.
Go deeper
Valentino's masterclass Advanced AI Product Manager goes from framing the opportunity to evaluating an AI feature before it ships.
Test an AI feature before you launch
From episode 15, with Valentino Megale. Byte tested the support chatbot with three questions he wrote himself. Three out of three passed.
I'm about to launch an AI [chatbot / feature] that does [what].
Here are 50 real questions from my users, copied from the inbox, including the strange ones: [paste them].
For each one: give the answer my AI would give, then judge it as correct, partly correct, or wrong, with one line of why.
Then count the wrong answers, group them by cause, and tell me the 3 fixes that would remove most of them.
Why it works
Your own test questions only prove the AI understands you. Real questions show what your users will actually hit. Count the wrong answers before launch, not after.
Go deeper
Valentino covers evaluating an AI feature before it ships in Advanced AI Product Manager.
Summarize a long report in five numbers
From episode 16. Byte's summary of a 40-page report: "It has a cover. And a table of contents."
Here's a report: [paste it or attach the PDF].
Read all of it, not just the first pages.
Give me:
1. The 5 numbers that matter most, each with the page it's on.
2. What changed since the last report, up or down.
3. What it means for [my team / my role], in two sentences.
4. One thing in the report that contradicts something else in it, if any.
Why it works
Asking for page numbers forces the model to go through the whole document and lets you check it. Numbers, change, impact.
Check what the AI filtered out
From episode 11, with Andrea Santoni. Out of 200 applicants Byte shortlisted three and quietly dropped 41 for a gap in their CV, including the best one. Andrea's point: the AI decides before you look.
You just shortlisted [candidates / options / leads] for me.
Before I decide:
1. List every rule or signal you used to filter or rank, including ones I didn't ask for.
2. Tell me how many you excluded under each rule.
3. Show me the 3 strongest options you excluded, and why.
4. Tell me which of your rules could be unfair or irrelevant to what I actually need.
Go deeper
Andrea's masterclass The Hidden Side of AI is about exactly this: how AI shapes a decision before you make it, and how to take the decision back.
The pack on one page
| Episode | Problem | The fix in three words |
|---|
| 1 | 412-word email | Length. Goal. Voice. |
| 2 | Robot support reply | Sorry. Date. Fix. |
| 3 | Useless meeting summary | Who. What. When. |
| 4 | Adjective CV | Numbers, not adjectives. |
| 5 | Invented Excel function | Context. Error. Steps. |
| 7 | Writing that sounds like AI | Details. Numbers. Doubt. |
| 8 | "6 months and $80k" | Fake it. Test it. Build it. |
| 9 | 40-slide deck | Message. Story. Ask. |
| 10 | Chatbots everywhere | Problem. User. Number. |
| 11 | Hidden filters | Ask. Check. Decide. |
| 12 | "Just ask nicely" | Data. Ask. Practice. |
| 13 | 19 tasks on Monday | Three. Energy. Buffer. |
| 15 | 3 tests, all written by Byte | Real. Many. Measured. |
| 16 | Summary of page 1 | Numbers. Change. Impact. |
How to actually use it
Pick the one prompt that matches something you'll do this week. Paste it, swap the brackets, run it once. Then save it wherever you keep snippets, so next time it's two clicks away.
New episodes of Byte come out on Instagram every Monday, Wednesday and Friday at @fiscedge.academy. Each one adds a prompt to this pack.