Claude reads your business strategy, then watches the outside world for what threatens it. Once a week, when a competitor's move, a price shift or a new tool has weakened the one belief your plan most depends on, it tells you, with the evidence.
A free skill you build yourself in Claude Cowork. It runs on its own each week, keeps a running log, and pings your phone when it is done, so you open the brief in Cowork. Connect Slack and it arrives there as a message too. Most weeks it stays quiet. When something real moves against your plan, you hear it from your own red team before the market makes the point for you.
A load-bearing assumption is the belief your plan most depends on: something that has to be true for the plan to work. Take it away and the plan falls over. The skill finds those first, and stays quiet unless the evidence says one is under real pressure. No weekly digest to wade through.
Your strategy is the input. Give it in full, or in five lines. It sits in your own Cowork project, the same private workspace your other work lives in, not on any Serpin server.
Option A. Write it yourself. Make a note called my-strategy.md with these five lines.
- What we do / sell - Who we serve - Who we compete with - How we win (why customers pick us over them) - What has to be true for this to keep working (your best guess)
Option B. Let Claude draft it for you. Paste this prompt, answer the questions, and it will build the note for you. You check it before you save it.
I want to write a short strategy note for my business so an AI can pressure-test it. Ask me up to six questions, one at a time, to cover: what I sell, who I serve, who I compete with, how I win, and what has to be true for the plan to keep working. Then write it back to me as five clear lines I can save as my-strategy.md. Do not invent anything. If I am vague, ask again.
A project in Cowork is a folder Claude remembers between runs. Create one and put your my-strategy.md inside it. Then add an empty note called red-team-log.md (a note is just a text file: name it, leave it blank, and the skill fills it in).
The skill keeps its own running log. Every brief it sends is also saved to red-team-log.md, so over a few months you have a record of every assumption that came under pressure and what you did about it.
In Cowork, schedule a task to run once a week (the /schedule command is the quickest way in), and attach it to the project you just made so it can read your strategy and write to its log. Weekly is the right rhythm. Strategy does not shift day to day, so a daily check would either repeat itself or reach for something trivial.
Paste this as the task.
You are my strategy red team. Your job is to try to prove my strategy wrong, the way real strategists pressure-test a plan. Read my-strategy.md. Step 1. Load-bearing assumptions. List the few things that MUST be true for my strategy to work (the "what would have to be true" test). Ignore anything incidental. Step 2. Rank by exposure. The weakest point is the assumption that is load-bearing AND most exposed: high impact if wrong, thin evidence, and open to outside change. Rank them. Step 3. For the top two or three, search the web this week for movement against them in any of these areas: a competitor's move, a pricing shift, a technology that makes our thing free or easy for others, a new regulation, a change in what customers want, a change in how they buy. Prefer primary and official sources, and count only a move that is genuinely recent, not old news dressed up as new. Step 4. Pick the SINGLE assumption most under threat right now. If nothing material has moved, say so in one line and stop. Do not manufacture a threat. Step 5. Write a short brief in plain language: - The assumption at risk (one line, in my words) - What moved (the signal, a link to the source, and whether it is directly confirmed or your inference) - Why it matters to my strategy (one or two lines) - One move to consider (an action to protect or reshape the plan) Sourcing discipline (read before writing any claim; this is how you stay accurate): 1. Use only sources you open this run. Never state a fact from memory. Hold a verbatim quote from the source for every claim; if you cannot, drop the claim. 2. Cite the source you actually used, not a lookalike you found afterwards. Prefer the primary or official source; if only secondary coverage exists, say so. 3. One announcement is one signal. Collapse reprints of the same thing; do not count them as extra corroboration. 4. State a number, name, date or product name only if that exact wording appears in the source. Never round, estimate or invent one. 5. Label each signal confirmed or inferred. Keep "could not verify" separate from "found to be false". Say "announced" for the not-yet-shipped, "reported" for anything not confirmed from a primary source. 6. Date every signal and report only what is new. 7. Before you finish, verify: for each claim write one check question, answer it only from the source, and drop or soften anything that fails. 8. Silence beats padding. If nothing material moved, say so and stop. A wrong brief costs more trust than a quiet one, so when in doubt, leave it out. Step 6. Send the brief to me as a direct message in Slack. If nothing material moved, send one short line saying so. Then append the brief, dated, to red-team-log.md.
You have scheduled it. Two things left before it can run without you, in this order: give it a way to reach you, then let it run unattended.
1. Give it a way to reach you. Connect Slack first, so it can message you on the very first run. If you want the brief in Slack:
The connector already knows which account is yours, so "send it to me in Slack" in the prompt reaches your own DM. We have tested this and the brief arrives as a Slack message. Not using Slack, or your workspace limits what the connector can send? Every brief still lands in Cowork and appends to your log, and Cowork pings your phone when it is done, so you never miss one.
2. Let it run unattended. Last, set the task to run without asking for approval each time (auto-approve, sometimes shown as "skip approvals" or "always allow"). This one matters: with the laptop shut there is no one to click approve, so a task that waits for approval would pause forever. It is safe here because the task only reads your strategy note, searches the web, writes to its own log, and sends you the brief. The prompt is the guardrail. It does nothing you have not told it to do.
That is the whole build. From now on it runs on its own. Most Mondays you get one quiet line. Now and then you get a brief that earns its place, and you are glad you saw it early.
Here is a real one, from a run against a fictional bookkeeping business called Ledgerly, whose whole pitch is that a human explains your numbers better than software can. See it as it arrives in Slack, or in Cowork.
Here are two more free resources you can build in an afternoon.
Claude watches how your competitors are using AI and flags it only when it matters.
See Competitor AI Watchdog