Prompting Guides for AI · member
15 prompts, checklists and workflows from the videos. The first six lines of every one are open; the full guides come with membership, and a new one lands with every long-form recording.
Stops a model answering as a generic assistant. The single highest-leverage change to any prompt you already use.
You are a {ROLE} with {N} years of experience in {DOMAIN}.
Context you need:
- Audience: {WHO WILL READ OR USE THIS}
- Constraint: {BUDGET / TIME / TECHNICAL LIMIT}
- What "good" looks like: {THE STANDARD YOU'D ACCEPT}Get a model to find the holes in its own output. Second-pass quality for the cost of one more message.
Review the output you just produced as if you were the person who has to live with it. Answer in this order: 1. What is factually shaky or unverifiable here? 2. What would the sharpest critic of this attack first?
For when the answers are generic because the model doesn't know your world. Front-loads the situation so it stops reasoning from general best practice.
Everything below is the context for this task. Read it all before responding.
## The situation
{WHERE THIS SITS — the project, the stage, why it matters now}
## Decisions already made (don't relitigate)Research you can actually check. Forces sources and separates what was found from what was inferred.
Research: {QUESTION}
Rules:
- Every factual claim gets a source. No source, no claim.
- Separate your answer into FOUND (in sources) and INFERRED (your reasoning
over them). Label each line.Turns an hour of recording into what was decided, what's owed and by whom — without the summary padding.
Here is a meeting transcript. Produce exactly four sections. No preamble, no summary of the summary. DECISIONS — what was actually decided, one line each. Only include things that were settled; a discussion is not a decision.
A blunt read on where the week actually went, what slipped and why, and the thing you've been carrying for three weeks without doing.
You're acting as my chief of staff. Below is my week.
{PASTE: calendar, inbox summary, task list, notes}
## Give me, in this order:
Write this before you build the workflow. It's also the shape the AI Automation Specialist assignment is graded against.
Before building, answer all eight. An automation that can't answer 5–7 will page you at the worst possible time. 1. TRIGGER — what starts it? (schedule / webhook / new row / manual) 2. INPUT — exactly what data arrives, and in what shape? 3. STEPS — each transformation, in order.
Most agent failures are tool-description failures. This is the checklist that catches them before you debug the model.
For each tool you give an agent: NAME — a verb and a noun. If you need a comment to explain the name, rename it. DESCRIPTION — write for someone who has never seen your system. State what it does, when to use it, and — the part everyone skips — when NOT to use it.
Ranked by consequence, with the failing case spelled out. Ends the review that always finds five stylistic things and no defects.
Review the diff below the way a senior engineer on this team would.
{PASTE THE DIFF}
## Context
What it's meant to do: {INTENT}Twenty rows that catch regressions. The thing everyone agrees they should have and nobody starts.
Don't build an eval framework. Build a table with four columns and twenty rows, then run it by hand the first time. INPUT | EXPECTED | ACTUAL | PASS? How to pick the twenty:
For when someone has proposed the most expensive model by default. Picks on the one dimension that actually decides it, then tests the cheap option first.
Task: {WHAT IT HAS TO DO}
Volume: {CALLS PER DAY}
Latency people will tolerate: {SECONDS}
Consequence of a wrong answer: {WHO IS AFFECTED AND HOW BADLY}
## Work through, in order:For a team that's already using AI without a policy — which is most teams. Written to be readable rather than defensible.
Fill in the blanks. One page. If it needs a second page nobody will read either of them. GREEN — do this freely e.g. drafting internal documents, summarising public material, code you'll review anyway, brainstorming.
Personalisation that holds up when they write back. Built to produce evidence, not adjectives.
Research {COMPANY} for an outreach message about {WHAT WE DO}.
Find and cite:
1. Something that changed in the last 90 days — funding, launch, hire, move.
2. What the person I'm writing to owns, in their own words if you can find
them (their posts, the job ad, a talk).A long post, four shorts, six clip candidates and a takeaway — every one traceable to something actually said in the source.
Source: {PASTE TRANSCRIPT OR LINK}
Audience: {WHO THIS IS FOR, and what they already know}
Voice notes: {THINGS WE DO AND DON'T SAY}
## First, find the spine
The 3–5 claims this actually makes. Quote the line where each is best said.For anyone using AI on numbers. Structures the verification instead of trusting the output.
Before any AI-produced number reaches a decision: 1. RECOMPUTE ONE — pick a single figure and calculate it by hand. If that one is wrong, discard the whole output; don't patch it. 2. TOTALS — do the parts sum to the whole? Models are fluent and bad at
Start with the free videos. Go deeper when you’re ready — a track for yourself, seats for your team, or learning live with us.