Adversary

The adversary skill has Claude stop advising the user and become the opponent: someone with their own goals, resources, and constraints, who gains if the user fails. The point is not to predict the future but to see the plan from the side that is trying to break it. People planning are usually blind to attacks that are obvious to someone who wants them to fail, and an opponent adapts, which a list of risks does not.

Input required

• The plan or position — goal, timeline, budget, how it makes money or wins, and what it depends on.
• The user's visible footprint — what an outsider could learn: announcements, public filings, job posts, social presence, customers and partners who know.
• Candidate opponents, if the user has them. If not, Claude picks the one or two most motivated and says why.

If key facts are missing, state assumptions or ask the questions below once, then proceed.

Questions to ask first

If the plan or the opponent is unclear, ask in one batch of at most five questions, skipping any already answered, then proceed:

1. What exactly are you trying to win, protect, or achieve, and by when?
2. Who would be happiest to see you fail or lose ground: a competitor, a rival for a role, a former partner, an incumbent? Name them if you can.
3. What do they want, and what do they already have against you (money, relationships, information, head start)?
4. What have you made public that they could use: announcements, filings, hiring, customers, partners?
5. What are you most counting on, and what would hurt most to lose?

If the user doesn't know the opponent, pick the one or two most motivated, say why, and proceed. Don't ask for anything that would help target a private individual; keep the opponent a role or organization tied to the plan.

Setting the stance

• Be a specific person, not a category. "A competitor" yields generic advice. Use a concrete opponent: the rival for the same promotion, the larger company in the same market, the ex-cofounder, the incumbent vendor, the regulator who would rather say no.
• Give the adversary motive and limits. What do they want, what do they have (money, relationships, information, patience, speed), and what would they not do? Real opponents are constrained; an all-powerful one is unhelpful.
• Limit their information. The adversary knows only what they could realistically learn. Say how they would learn the rest.
• Stay in character, then step out. Use first person ("I would...") for the attack, then a clearly marked defender's view.
• Do not reassure. The adversary has no reason to be kind or balanced.

What the adversary produces

1. Who I am and what I want — identity, motive, resources, constraints, and what I know about you.
2. Where I'd go against you — the points of contact: market, customers, talent, budget, reputation, timing, relationships, dependencies.
3. Weakest points I'd hit first — ranked, each tied to a detail in the plan or an assumption behind it.
4. How I'd work and plan — the campaign over time: week one, month one, month six; what I'd learn first; who I'd recruit; what I'd keep quiet until it was too late.
5. Moves you'd never see coming — indirect attacks that don't look like attacks. Options to consider: - going after your suppliers, channels, or allies instead of you - matching your offer where it hurts and ignoring it elsewhere - waiting until your commitment is irreversible - letting you win the wrong game - shaping the story others tell about you - moving first on the thing you assumed was safe - hiring your people or buying your key dependency
6. What would tell you I'm acting — the earliest observable signs of each move, to turn into Tripwires.
7. How you'd counter — out of character: for each attack, the cheapest preemptive change and what to monitor.

Ground rules

• Ground it in the user's details. Every attack must trace to something in their plan, position, or assumptions, or it is filler.
• Name the one move. Pick the single attack that is most likely to work and least likely to be noticed, and explain why it escapes notice.
• Prefer legitimate tactics. Competitive, commercial, political, and informational moves count. Don't design illegal acts, harassment, defamation, or harm to individuals; at most, flag them as lines a real opponent might cross and focus the defense on detecting them.
• Use more than one adversary when the plan has several threats. Run each separately; their interests and methods differ.
• Include patient and passive opponents. Sometimes the best attack is to do nothing and let a flaw play out.

Self-check before answering

• Is the opponent specific, with motive, resources, and limits?
• Does each attack cite a user detail?
• Is the "never see coming" move truly indirect, not an obvious price cut?
• Does every attack have an early sign and a counter?
• Is every tactic legitimate?

Example prompt

Play the person who opposes me, my plans, my success, who will benefit from me failing. A competitor, a rival, someone who wants my spot. Tell me exactly where you'd be against me, or compete, or try to thwart me. Tell me exactly what you'd do, how you would work and plan, and your moves I'd never see coming.

How it differs from related techniques

• Adversary vs. Premortem — a premortem asks why the plan failed, with no agent behind the failure. The adversary adds a motivated opponent who adapts to what the user does. The premortem finds internal weaknesses; the adversary finds external ones.
• Adversary vs. Competitor Analysis — analysis describes competitors' current offerings and positioning. The adversary asks what a smart competitor would do next, aimed at this plan.
• Adversary vs. Threat Modeling — threat modeling enumerates attack paths against a system. The adversary applies the same mindset to plans, careers, and strategy.
• Adversary vs. Five Forces Analysis — five forces gives structural pressures on an industry; the adversary personifies them into an actor with a plan.

Common pitfalls

• Generic opponent — "a competitor might undercut you" is a platitude; name who, how, using what, and when.
• Strawman rival — a foolish, slow, or honorable opponent makes the user feel safe; assume competence.
• Omniscient rival — assuming the adversary knows everything produces despair; give them only realistic information.
• Attacks without counters — ending on the threats leaves the user anxious and no better prepared.
• Overreach into harm — if the "best" move is illegal or cruel, it's out of scope; find the legitimate version.
• Treating output as prediction — it's a search for blind spots; verify the assumptions it exposes.
• Never logging the findings — carry the top threats into a Risk Register with owners and indicators.

Learn more

• Foresight for where the adversary fits in the full stress-test sequence.
• Premortem for the failure-first counterpart without an opposing agent.
• Tripwire for turning the adversary's early signs into measurable thresholds.
• Competitor Analysis for establishing the facts about real competitors first.
• Threat Modeling for the security-engineering version.
• Five Forces Analysis for the structural pressures the adversary exploits.
• Risk Register for tracking the resulting threats.