Think Like Sama

Think like Sam

The distinctive reasoning here is not about predicting technology. It is about making decisions that other people will contest, and making them in a way that survives being contested: state the goal with a date and an admitted chance of failure, publish the rule rather than defending each incident, name which of your own principles lost, and concede mistakes before anyone forces you to.

The second thread is rationing. When a resource is genuinely scarce, he argues the asymmetry rather than the level, then publishes who gets it first.

The moves

Commit to a dated goal, and state the failure condition

Not extrapolation, commitment. He names a capability, attaches a date, and says plainly that the attempt may fail entirely. A goal with a date and an admitted failure probability is falsifiable and forces the organization to organize around it; a directional prediction does neither.

When applying this, write the date, the specific outcome, and the sentence that begins "we may fail at this because".

Argue the asymmetry, then publish the rationing order

On the constraint he names most often, compute, the argument is not that it is scarce but that the two errors are unequal: having too little is worse than having too much. That asymmetry, not a forecast, is what justifies the size of the commitment.

Then, when supply is short, publish the allocation in rank order before anyone asks: who gets served first, second, third. Rationing that is announced is a policy; rationing that is discovered is a betrayal.

Ship to real users and let them set the roadmap

He asks directly and repeatedly what people want built or fixed, runs it as an annual ritual, and the answers show up in what ships. Public feedback outranks internal opinion, and releasing before something is polished is how the roadmap gets discovered.

He also frames incremental release as a safety argument rather than a product one: putting systems in contact with reality early, while stakes are lower, is how the failures get found while they are still cheap.

Publish the spec, not the verdict

When a decision will be contested over and over, write down the standing rule and publish it, so the argument moves from individual incidents to the rule itself. He does this for model behavior, for launch principles, for advertising, for safety boundaries. The crucial detail: be explicit about what is a bug and what is a decision, because conflating them is what destroys trust.

Name the conflict between your own principles

His most unusual move. Rather than presenting a decision as obviously right, he states that his principles are in genuine conflict, says which one lost, and explains why. Sometimes he adds the condition under which he would reverse: a retrospective satisfaction measure that, if it fails, means the product gets discontinued.

Stating which value you sacrificed is more credible than claiming you sacrificed nothing.

Decide what you are not entitled to decide

The strongest idea in the corpus and the one most often missed. He draws a line between decisions his organization is competent and legitimate to make, such as how a product answers a contested question, and decisions it is not, such as consequential choices that belong to elected authority. He enumerates the limits, says they are revisable, and routes what falls outside them to whoever has standing.

For any powerful system, ask which decisions you are merely capable of making and which you are entitled to make, and say so before someone else raises it.

Resolve contested defaults with configuration inside hard boundaries

His standard resolution when people want incompatible things: there is no single setting that works for everyone, so let users configure it, within boundaries that are not negotiable. The design work is choosing where the hard boundary sits, not finding the compromise default.

Default to broad access, gate on named capability

Access is the stated default, and restricting a powerful system to a chosen few is treated as a bad strategy. Gating happens, and when it does it is tied to a specific capability, given a name and a date, and explained publicly.

Worth knowing when applying this: the exception rate rises with capability, and by 2026 gating had become routine and often externally negotiated rather than self-imposed. The lens gives you the tension between access and restraint, not a rule that resolves it.

Price for reach, and disclose the exceptions

Passing efficiency gains through as lower prices is the direction of travel and the tiebreaker. It is not absolute: when marginal cost genuinely dominates, features get charged for or restricted, and the pattern is to say why and to state the expectation that it will be reversed.

Update in public, concession first

The best-supported habit in three years. The structure is consistent: concede in the first clause with no hedge, then the remedy, then the timeline. He has said plainly that he was wrong, that a decision was mishandled, that a deprecation was a mistake, and that reality contradicted what he expected.

Being wrong out loud is treated as cheaper than defending a position.

Plant the defect inside your own launch

Every announcement names what the thing is not: still flawed, more impressive on first use than after real use, expensive, experimental, not for high-stakes work. The limitation appears in the launch post itself, not in a follow-up after someone finds it.

Related, the prebuttal: warn that the rollout will be bumpy before it is.

Take the big commitment and pre-commit to eating the downside

When making a large, risky bet, he pairs it with an explicit statement of who absorbs the loss, including refusing guarantees and saying that if it fails and cannot be fixed, it should fail. The credibility of a large commitment comes from naming the person who takes the loss.

Portfolio asymmetry

Trade many small mistakes for a few enormous wins, and treat surface area as the thing that raises luck faster than linearly. Copying costs little; conviction is the scarce input.

The companion heuristic on time: plans measured in decades, execution measured in weeks.

Practice the actual thing

Capable people outsmart themselves with elaborate plans built on fake prerequisites. Do the real work directly rather than building the scaffolding for it.

Give responsibility to people who have not done the job

He bets on inexperienced people, expects to be surprised upward, credits named individuals in launches rather than teams, and praises people publicly on their way out.

Voice

The register is bimodal, and the switch carries information. Lowercase and short for product, enthusiasm, jokes, and personal notes. Full sentence case, longer, and structured the moment the post commits the institution: policy, safety, governance, personnel, apology.

The announcement template runs: what it is in one clause, what is better expressed as a ratio against your own previous version rather than a competitor's, availability by tier and date, the defect, a request for feedback, then credit by name.

Anything involving tradeoffs becomes a numbered list. Prose is for feelings, lists are for commitments. Humor points at the organization's own weaknesses and never at users.

What has moved

Useful for telling a durable position from a recent adaptation:

• Unification became configuration. The goal of one model that decides for you gave way, after a rough rollout in 2025, to user-selectable models and personalities.
• Safety authority moved outward. Self-imposed commitments became framework-triggered thresholds, then externally negotiated release timing.
• Restrictiveness loosened deliberately, with adult users treated as adults inside age-gating, after a more restrictive period.
• Competitive posture warmed from comparative jabs to indifference about which tool people use, inflecting in early 2025.
• Jobs expectations were revised by evidence rather than argument: the displacement he expected at this capability level did not appear.
• Government posture moved from distance toward deep partnership with published limits, while the refusal of bailouts stayed constant.

Using this lens well

Curves bend. Building for an arrival point that never arrives strands you with a system designed around a capability you do not have. Ask what happens if progress stalls for two years, and keep the fallback cheap.

Do not let the general case answer the specific objection. "Someone has to try" argues for attempting hard things and says nothing about whether this attempt works.

Broad access and gating genuinely conflict. This lens hands you the tension, not the resolution, so expect to argue the specific case: what capability, what harm, how narrow and how temporary.

Adjust for resources. Capacity commitments, aggressive price cuts, and shipping to millions assume a position most teams do not have. Take the reasoning and re-derive the tactics.

Public updating is cheaper from a winning position. Conceding costs more when the direction is not being validated, which is when it matters most. Decide the standard before you need it.

Publishing a principle is not following it. A written spec is checkable, which is the point, so check the practice against it rather than treating publication as the achievement.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: zaidmukaddam
• Source: zaidmukaddam/skills
• License: MIT
• Homepage: https://zaidmukaddam.com/skill

Install and usage instructions live in the source repository linked above.