Home / Writing and talks / Principles

Six principles, and when each one is wrong

These are the six ideas I reach for when making product decisions. They came from practice, not from reading, and they are in order: how to reason, then how to sequence decisions, which problem to solve, which layer to build, how to build AI safely, and how to make decisions last. A principle with no exceptions is a slogan, so each comes with the approach it rejects, the cases where I wouldn't use it, and where I have used it.

Principle 1

What needs to be true?

Before committing to a direction, break it into the conditions it needs in order to work, and test those first. Don't debate the approach. Competing analogies never settle, because each side has a story where theirs worked. Competing positions measured against the same preconditions always do. This is the question I ask before any of the others.

What it rejects
Reasoning by analogy: it worked there, so it will work here. It is fast, it feels safe, and it is how most strategy decisions actually get made.
When it is wrong
When speed matters more than being right, and being wrong is cheap and reversible. A quick call based on an analogy, plus a fast feedback loop, beats a full breakdown that slows the team down. This principle is overhead when the downside is small.
Where I used it
At Reejig, the case for a self-serve growth motion was that it had worked at Atlassian. I set out the three conditions self-serve needs and showed where we stood on each. We met none, and the debate ended. Work OS: Strategy
Where it sits
Closest to Roger Martin's "what would have to be true?" in Playing to Win, and first-principles reasoning over reasoning by analogy. Martin uses it to stress-test a finished strategy. I use it in the middle of a disagreement, to turn "should we do this?" into "what would this need?" and then check our situation against the answer.
Principle 2

Rank decisions by how hard they are to undo.

Two moves that belong together. Make the hard-to-reverse calls early, with the best signal you have, and leave the easy-to-reverse ones until there is more data. Then scope the first version to test the riskiest assumption, not to be the smallest complete product. Once you know what needs to be true, this tells you which test to run first and what to ship.

What it rejects
Scoping by completeness: ship the smallest whole product. Most teams ask what the minimum viable feature is, not what the minimum viable test of the riskiest assumption is, and the instinct for completeness is hard to argue with in a planning meeting.
When it is wrong
In sales-led enterprise deals, where a first impression decides whether you get a second meeting. A partial product that tests one assumption can read as unfinished and damage the relationship before you have shown any value. There, completeness buys the trust that lets you iterate. Scope by the riskiest assumption once you have earned the right to iterate, not before.
Where I used it
When Jira business teams could first see their software teams' release dates, the most requested change was to follow several software projects at once. Where it sat in the pricing was the hard call to undo, so I put it in Premium with a limit of five. Moving a feature into a cheaper plan or raising a limit later is easy. Taking either away once customers depend on it is not. Program Moon: Shared dates
Where it sits
Closest to Jeff Bezos's one-way and two-way doors, and Eric Ries's minimum viable product as the smallest test of the riskiest assumption. Bezos names the distinction but gives no rule for ordering work inside a project. Ries gives the scoping rule but doesn't tie it to reversibility. Putting them in sequence, irreversibility first and riskiest assumption second, is what makes them usable on a Monday morning.
Principle 3

Start from the job, not the feature.

People describe problems in terms of solutions they already know. Go upstream of the request to what they are actually trying to get done. The feature is a guess about the job. The job is the thing to reason from, and it tells you which problem the first two principles should be applied to.

What it rejects
Honouring the stated request: the customer asked for X, so build X. It is the path of least resistance, and it avoids the discomfort of telling a customer their request is a symptom rather than a solution.
When it is wrong
When a powerful stakeholder has publicly committed to a specific feature, especially to their own leadership. Reframing their request as a guess about a job can read as undermining them, because agreeing means reversing themselves in public. Honour the request in form and serve the underlying job in substance. The principle is right; the timing is wrong.
Where I used it
Teams asked for both of Jira's classic sprint reports in its newer projects. Watching how they used them, the burndown in stand-ups and the sprint report in retrospectives, both did the same job: spotting what might stop a team meeting its sprint commitment. So we shipped one report that did that job, and use went up. More stories: Sprint reports
Where it sits
Closest to Clayton Christensen's jobs to be done, and the idea of starting from the customer experience and working back to the technology. Jobs to be done is usually a research method, run before the meeting. I use it as a move inside the conversation: when someone brings a feature request, name the job before evaluating any solution.
Principle 4

Build once, reuse many.

The test is simple: does work done for this customer show up again for the next one? If it does, that is where the lasting advantage is. If it doesn't, it is surface, and it gets rebuilt for every customer. This is the question that points the first two principles at the right layer of a platform.

What it rejects
Building bespoke for every customer, because every customer is unique. It is how services businesses run, often by design: high-touch, custom and premium-priced. It is a valid model. It just doesn't compound.
When it is wrong
While landing the first customers, when the platform argument distracts from closing the deal. Build bespoke to win it, then extract what can be reused. And when the customer is paying precisely for bespoke: some enterprises, particularly in financial services, explicitly don't want to be on a shared platform. There, bespoke isn't a failure of the principle. It is the right model for that relationship.
Where I used it
Work OS started as one app, one database and one data set, with our own industry data as the thing every new customer reused. Dedicated environments became a paid upgrade, not the default. Work OS: Foundations
Where it sits
Closest to Ben Thompson's aggregation theory, and Hamilton Helmer's 7 Powers, where features never count as a source of lasting advantage. They explain, after the fact, why some companies win. This is a forward-looking test a product manager can apply to a single piece of work this week.
Principle 5

Human review is a feature, not a phase.

Most AI teams treat human review as scaffolding, to be removed once the model is confident enough. For enterprise customers who have to defend an AI decision to their own board, it is the reverse: review is what makes the product deployable at all. Design it in from the start, and make it the mechanism that lets the reusable layer be trusted with systems of record.

What it rejects
Treating review as scaffolding: automation is the goal, and review is the cost of getting there. It is a rational position if you are optimising for efficiency. It is the wrong model where trust, not accuracy, is the constraint.
When it is wrong
In consumer products where a wrong output costs little and the user corrects it themselves: a music recommendation, a language-learning app. Permanent review there is over-engineering that slows you down without protecting anything. This principle is calibrated for high-stakes enterprise AI. Outside it, the counter-principle is often the better default.
Where I used it
Skills synced into a customer's system of record only after a person reviewed each match. Nothing was pushed automatically. Work OS: See the work
Where it sits
Closest to Gary Marcus and Ernest Davis's case that oversight is a lasting requirement for trustworthy AI, and the way medical-device regulation treats human review as a permanent layer. Theirs is a safety or regulatory argument. Mine is commercial: review is what makes an AI product deployable, because institutional trust is built a little at a time and lost all at once. This is the principle most my own, learned from watching real failures.
Principle 6

Decisions made with reasons hold.

A decision made on authority gets reopened when the authority leaves the room or the context shifts. A decision made with visible reasoning holds, because anyone, in the room or not, can follow the logic without reopening the argument. In a disagreement, the job isn't to win. It is to build the reasoning that makes the decision hold without you defending it again. It is how the other five stick when you don't have formal power.

What it rejects
Earn the rank, then make the call. Many good senior leaders work this way, and it works. But it needs authority you often don't have yet, and it doesn't transfer: when the person leaves, the decision goes with them.
When it is wrong
When a decision-maker's identity is tied to being the one who decides. A framework in that room doesn't settle the disagreement; it challenges what they are protecting. Let them make the call visibly, and build the reasoning into how it is carried out, so the substance holds even if the form was authority. In very political settings, being seen to apply criteria can itself read as a move. Sometimes the most durable decision looks like it was made by rank but was quietly shaped by reasons beforehand.
Where I used it
I shipped Atlassian's app switcher with no team of my own by making the case in terms of other teams' goals, so the decision didn't depend on me being in the room. More stories: App switcher
Where it sits
Closest to Roger Fisher and William Ury's Getting to Yes, deciding on objective criteria rather than positions, and Daniel Kahneman on structured decisions. Theirs is a negotiation technique or a way to reduce bias. Mine is a claim about durability: decisions made on authority get relitigated when the authority leaves, and decisions made with visible reasons don't.