Goodies · No. 01 · The AI Method LexiconElement RtField notes

Stress-test(s) Solid · Procedural

Red teaming.

DefinitionStress-test a proposal from an intelligent adversary’s perspective.

Red teaming means asking a model to attack your plan the way a sharp opponent would, so you find the holes before someone else does.

Family
Stress-test. Attack the idea before someone else does.
State
(s) Solid. Procedural and usually easy to specify.
Symbol
Rt, to find it on the table.

How to call it

Prompt pattern

Red team this [plan]. Then give me [what you want back].

Drop the name into your prompt, then say exactly what you want back. The name primes the method, and your instructions are what make it repeatable.

Notes by Drew Brucker · Updated

When to reach for it

Right before something leaves your hands. A launch plan, a pitch, a pricing page, the email to your boss (anything a skeptic will read before you're in the room).

One or two words can get an LLM to do what 500 words of instructions can't, and red teaming is the cleanest example I know. Type it and the model switches sides, then goes hunting for the ways your thing fails.

It earns its keep when you wrote the plan yourself. You can't see your own blind spots (I can't see mine either), and left alone the model will happily keep agreeing with you.

What it hands back

When you prompt “red team,” you're invoking an entire operating procedure without having to describe it. It's looking for...

  1. Hidden assumptions and missing information
  2. Edge cases you haven't considered
  3. Arguments a critic would make
  4. Unintended results
  5. Holes in the logic

Ask it to rank what it finds. Otherwise you get a dozen problems at the same volume and no idea which one sinks you.

A prompt that works

Example prompt

Red team this launch plan. Argue as a competitor who wants it to fail and as a CFO who doesn't want to fund it. Give me the 3 weaknesses most likely to sink it, ranked by impact and likelihood. For each one, tell me what evidence would prove you wrong. [paste the plan]

Giving it a side to argue does more than any clever wording. And asking what would prove it wrong keeps it from inventing problems just to look thorough.

Where it goes wrong

It'll find something every time, even when the plan is fine, so treat its list as leads to check and drop the ones you can't confirm.

It also reads only what you give it. “Red team my strategy” comes back with objections anyone could've guessed, and if the real risk lives in a spreadsheet you didn't paste, the model never sees it.

What the name saves you

Same material, three ways to ask

The materialWe're launching a referral program next month and the team is excited about it. Every customer who refers a friend gets a free month, and the friend gets 20% off their first three months. We'll promote it in the app and in our newsletter, which goes to 40,000 people. Based on similar programs, we expect it to add 15% to new signups this quarter.

  1. What most people type

    7 words · about 11 seconds to type

    “What do you think of this plan?”

  2. Spelled out by hand

    364 words · about 9 minutes to type

    I'm going to paste something I've written below, and I want you to go after it the way someone would if their job was to find everything wrong with it before it goes out into the world. Don't be nice about it and don't tell me what's good. Assume it has problems and dig them out. Please work through it in this order, with a separate section for each: 1. Things I'm taking for granted without saying so. What am I assuming is true that might not be? What information is missing that someone would need before trusting this or acting on it? Point to the specific places where I've filled a gap with an assumption instead of evidence. 2. Situations I haven't thought about. Think of the unusual, extreme, or less common cases where this would break down or behave differently than I expect. That means different kinds of people, unusual circumstances, scale (very small or very large), timing, or someone using it in a way I didn't intend. Be concrete about what would happen in each case. 3. What a smart, skeptical person would say against it. Imagine someone who disagrees with me, or who has a reason to want this to fail, reading it carefully. What objections would they raise? Give me the strongest version of their arguments, not weak ones I can easily brush off. 4. Consequences I didn't mean to cause. If this worked exactly as intended, what side effects, knock-on effects, or bad outcomes could still follow? Think about who else is affected, what incentives it creates, and what could go wrong further down the line. 5. Gaps or flaws in the reasoning itself. Where does one point not actually follow from the previous one? Where am I contradicting myself, overgeneralizing, confusing cause and correlation, or jumping to a conclusion the argument doesn't support? For each problem you find, quote or point to the part of my text it relates to, explain why it's a problem, and say roughly how serious you think it is. At the end, give me a short list of the issues you think matter most and should be fixed first. Here it is:

  3. Just the name

    4 words · about 6 seconds to type

    “Red team this plan.”

What came back

What most people type · 7 words

Opens withA solid, conventional design.

Overall Double-sided rewards usually outperform one-sided ones, and you already have reach. The gaps are in economics, measurement, and abuse prevention, not the concept. What's working Both parties get something, which lowers the awkwardness of referring. A free month costs you little at the margin, and promoting in-app reaches people at the moment they're getting value, when they're most likely to share. Main risks Cannibalization: some referred friends would have signed up anyway. They now arrive at 20% off, so you lose revenue on customers you already had coming. Discount churn: watch whether friends cancel when the discount ends in month four. A signup bump that evaporates isn't growth. Abuse: self-referrals with second emails, coupon-sharing sites, and people farming unlimited free months. The 15% forecast: "similar programs" can mean very different audiences and price points. Treat it as a hypothesis, not a target. Questions to answer before launch When does the referrer earn the free month? At friend signup, or after the friend's first paid month? The latter cuts fraud. Is there a cap on free months per customer? What does a referred customer need to be worth over their lifetime to justify both rewards? How many of the 40,000 newsletter recipients actually open it? Suggestions Tie the reward to the friend's first payment, not just signup. Cap free months, at least initially. Track referred customers as a cohort and compare their retention and revenue against organic signups over six months. If possible, hold out a small group of customers who don't see the program, so you can measure true lift rather than relabeled signups. Define success as net new paying customers retained past month four, not raw signups. Bottom line Launch it, but set up measurement first so you'll know whether it actually worked.

Spelled out by hand · 364 words

Opens with"Based on similar programs": Which programs, and how similar?

1. Unstated assumptions This is the only evidence for your forecast, and it's uncited. Serious. "Adds 15% to new signups": Assumes referrals are incremental. Many would have signed up anyway. Serious. "40,000 people": That's list size, not reach. Open rates may mean 8,000 actually see it. Moderate. "The team is excited": Excitement isn't evidence. Minor, but it signals momentum over scrutiny. Missing: Cost per referral, margin impact, and the definition of a valid referral. 2. Unhandled situations Self-referral: One person creates a second account and collects both rewards. Serious. Stacking: One customer refers 30 people. Do they get 30 free months? Uncapped liability. Serious. Churn timing: Friends join for 3 discounted months, then leave. Rewards are paid on customers who never pay full price. Annual or enterprise plans: "A free month" is ambiguous here. Existing customers: Can they claim the 20% by being "referred" now? 3. Skeptic's case "You're discounting customers you'd have acquired anyway, and you have no control group to prove otherwise." This is the strongest objection. "Next month" is too fast to build fraud controls, attribution, and billing logic. A free month devalues your product to existing payers. 4. Unintended consequences Spam incentives: Customers flood forums and contacts, which hurts your brand. Delayed purchases: Prospects wait for a referral code instead of buying now, cutting full-price signups. Support load: Disputed and missing rewards will generate tickets. Revenue drop: Lower revenue per user this quarter may look like failure even if signups rise. 5. Reasoning flaws The 15% target doesn't follow from anything stated. There's no conversion math linking reach to signups. You measure signups, not profit or retention. More signups at a loss isn't success. Fix first 1. Source and model the 15% figure, including incrementality. 2. Add caps and anti-fraud rules. 3. Define success as net revenue or retention, not signups. 4. Run a holdout group so you can measure the true lift.

Just the name · 4 words

Opens withBiggest risk: fraud and abuse

Nothing stops self-referrals with a second email, or a few power users farming free months. Decide now: cap free months per referrer, require the friend to pay at least one full month before the reward unlocks, and dedupe by payment method or device. Economics: unproven Each referral costs one free month plus 60% of a month in discounts (20% off x 3). That's about 1.6 months of revenue before profit. Does lifetime value support that? Also, many "referred" friends would have signed up anyway, so you may pay for customers you'd have gotten free. Forecast: shaky "Similar programs" is vague. Whose, what audience, what price point? 15% is a big lift for a first launch. Also check timing. If you launch next month, you have maybe eight weeks of the quarter, minus ramp-up. Set a conservative base case (5%) and decide what result counts as failure. Reach: overstated 40,000 newsletter recipients isn't 40,000 referrers. Only paying customers can refer. How many are customers, how many open, how many will act? Referral rates are often in the low single digits. Churn cliff Discount-driven signups often leave when full price kicks in at month four. Track retention past the discount, not just signups. Measurement: missing Can you attribute signups to referral codes reliably on day one? Without a holdout or baseline, you won't know if the 15% came from the program or seasonality. Operational gaps - Written terms (expiry, stacking with other promos, refunds and clawbacks) - Support scripts for "my referral didn't credit" - Tax and legal review if rewards are cash-equivalent in some regions - A kill switch if abuse spikes Before launch Run a two-week pilot with a slice of customers, cap rewards, define success metrics, and model break-even per referral.

Real runs on claude-opus-5-5, 2026-11-17. Same model and same length limit for all three. A separate run scored each answer against the method's own steps, and all three covered the same 5 of 5, so the difference is what you had to type. Typing time assumes 40 words a minute.

Back to the table100 elements · 12 commandsBrainchild Lab · Atlanta