learn-ai-pm-with-phoebe / Leader session 3 of 6
Learn AI Product Management with Phoebe · Leader track · Session 3 of 6

Specs, decisions, accountability

You will never write a spec again, and you approve them constantly. That is the awkward position this session is written for. A spec is not a document, it is the record of six decisions, and a drafting tool will produce a beautiful version of the document with four of the six decisions never made. You cannot catch that by reading more carefully, because there is nothing on the page to catch. You can catch it with two questions asked before you open the file.

🟡 Leader track Heads of product · founders · stakeholders No spec to write Bring one you approved 45 min
0-3 · Welcome 3-14 · The spec as a decision record 14-24 · The two that are always missing 24-33 · Who signs, and what that means 33-42 · The two-question drill 42-45 · Q&A
Part 0

The one place your signature actually lands

a2 was about believing a finding. This session is about approving a commitment, which is the next thing that happens to the same evidence and the point where your name goes on it. Nearly every product decision you fund passes through a document called a spec, a PRD or a one-pager, and almost none of the value of that document is in the writing. It is in six decisions that either got made or did not, and a fluent draft is completely silent about which.

Your PMs run a lab on this in the practitioner track: six checks, five drafts of the same spec, scored 15, 35, 55, 70 and 100 as more input goes in, with only the top rung signable. You do not need the lab. You need to know which two of the six survive almost to the top of that ladder, why those two and not the others, and the two sentences that surface both of them without you reading a line of the document.

When a leader reaches this session they usually want to talk about dates, sequencing and sign-off gates. Those are delivery questions and they belong to learn-ai-project-management. This page stops at the moment the decision is recorded and somebody puts their name to it.

Live - presented in session Self-study - read after class ▶ Discussion exercise - run it in the room Sources covered
★ What you walk out with today The six things a spec has to contain, named so you can ask for them by name; the two that are almost always missing and the reason it is always those two; two questions that surface both without reading the document; and a clear answer to what a signature on a spec actually commits somebody to.
Part 1 · covers PRD and spec practice, the spec as a decision record

A spec is a decision record 9 min live

The most useful thing you can do with the word "spec" is stop treating it as a kind of writing. A spec is the place a team writes down what it has already decided so that somebody who was in none of the meetings can act on it without guessing. Everything else in the document - the context, the background, the tidy headings - is packaging around six decisions. Remove the six and you still have a document. You do not have a spec, and nothing about the file will tell you which one you are holding.

Two objects that arrive in the same file format A document · reads finished every heading a PRD is meant to have confident prose filling all of them approved in eleven minutes three readers, three readings What it proves: somebody wrote it A decision record · settles six things who it is for, and who it is not for what would count as having worked where the work stops, in writing what was given up, and who loses What it proves: somebody decided The two look identical, and they get approved at the same speed. Sections are where decisions live rather than what they are, so four of the six can be missing without leaving a mark on the page. Length, polish and structure are all evidence of writing, and of nothing else.
🔍 Click to zoom - the packaging is identical, the contents are not

Here are the six, with the column that matters to you. Not what each one is for - your PMs have that - but who finds out it was missing, and how late. Read the third column as a schedule of costs, because that is what it is.

The checkWhat it actually decidesWho notices it is missing, and when
1 · Problem evidencedwhether this is worth anybody's quarter, on evidence somebody could disagree withyou, in the review, if you ask. Otherwise nobody, until the thing ships and the numbers do not move
2 · User namedwho this is for, at a grain finer than "our users"a designer, in week two, asking which of three users to design for and being told all of them
3 · Success measurablewhat would count as this having worked, with a baselineeveryone, six months later, in an argument about whether it worked that nobody can win
4 · Non-goals statedwhere the work stops, and what is explicitly not in itthe engineering lead, in week four, when scope has grown by three tickets and nobody can point at when
5 · Edge cases decidedwhat happens on the unhappy path, decided in advancean engineer, at four on a Friday, who decides it alone and quite reasonably never mentions it
6 · Trade-off explicitwhat was given up to do this, and who loses by itthe retro, where it surfaces as a disagreement about what was agreed rather than about what to do next
LiveWhy the absence is invisible rather than sloppy4 min

The instinct is that a spec missing four decisions must look thin. It does not, and the reason is worth thirty seconds of your attention because it applies to every document that now reaches you:

  • A drafting tool fills sections, not decisions. Asked for a spec, it produces the sections a spec has. Where a decision was supplied it writes the decision down. Where none was supplied it writes something reasonable in the right register, and those two outputs are indistinguishable in tone, length and confidence.
  • The unmade decisions come out cleaner than the made ones. A real trade-off is awkward, because a real person loses something in it. An absent one produces a smooth paragraph in which everything improves. The smoothest section in a spec is the one most likely to contain nothing.
  • Approval rewards the shape. You have read several thousand documents in your career and you recognise a competent one in seconds. That recognition used to be a decent proxy for care taken, and it is now a proxy for nothing, in exactly the way a2 described for research.
The signable test, in one sentence Could an engineer who was in none of your meetings build the right thing from this without asking a question or inventing an answer? That is the whole bar. It is not about length, formatting or how confident the prose sounds, and it is answerable by anybody in the room in about a minute.
Self-studySix checks is a completeness bar, not a quality bar3 min read

Be precise about what this session gives you, or you will over-trust it in the other direction:

  • All six present does not mean the decision is right. A perfectly complete spec for a problem nobody has scores full marks. The checks catch absence, cheaply and reliably. Whether the problem was worth solving is judgement, it is yours, and nothing here can score it.
  • A metric being present does not make it the right metric. Check 3 is satisfied by a vanity number with a baseline attached. a5 is where the difference between a metric and a number lives.
  • Evidence being present does not make it good evidence. Check 1 is satisfied by eight interviews with your friendliest customers exactly as easily as by eight representative ones, which is what a2 spent a session on.

So the honest framing: completeness is what a leader can require, in public, uniformly, at no cost. Quality is what you have to think about, one decision at a time. Requiring the first is what buys you the attention to spend on the second.

Part 2 · the pattern, and the reason behind it

The two that require somebody to give something up 8 min live

The six checks do not fail at random. Four of them get written whether or not anybody thought hard, because writing them costs nothing. Two of them are almost always absent, and it is the same two every time, in AI-drafted specs and in hand-written ones. The pattern has a single cause: those two are the only checks where writing the sentence means a named person loses something they wanted.

1 · Problem evidenced Gets written anyway Somebody asked and gave a reason. Whether it is evidence was a2's whole session. 2 · User named Gets written, roughly "Users" passes as a segment until a designer asks which ones, in week two. 3 · Success measurable Gets written as a mood "Improve engagement" fills the section without setting any bar at all. 4 · Non-goals stated Almost always absent Nobody has to give anything up to leave this empty, and somebody does to fill it. 5 · Edge cases decided Gets written, sometimes If somebody technical reads it early. Otherwise decided on a Friday afternoon. 6 · Trade-off explicit Almost always absent Writing it means telling a named person, by name, that they did not get their thing. Two questions surface both, and you do not have to open the document to ask them. "What are we not doing here?" and "how will we know it worked?" Both are answerable in seconds when the decision was made, and unanswerable when it was not. Neither one requires you to have read a word.
🔍 Click to zoom - the same two, every time, for the same reason
LiveWhy it is always these two4 min

Four of the checks are descriptions. Two are concessions. That is the entire explanation, and it holds whoever or whatever wrote the draft:

  • Non-goals are the scope decision, and a draft cannot make one. Exclusion is not a writing problem. A drafting tool has no reason to leave anything out, because nothing in the input says what is out, so everything adjacent stays arguably in. If the non-goals section is empty, the team has not finished deciding. It has finished describing, and those two states produce the same document.
  • The trade-off is the sentence with a victim in it. "We chose this at the cost of that, and we are accepting it" names somebody who loses. Nobody writes that sentence casually, and a tool will not write it unsolicited because it is not in the input unless a person put it there. A spec in which everything wins is a spec in which nothing was chosen.
  • They fall together, which is why the fix is cheap. Your PMs measured this: adding the non-goals to the draft fixes two checks at once, because you cannot write down what you are not doing without exposing what you gave up. One instruction, two decisions, and it is the difference between the fourth rung of their ladder and the only signable one.
The uncomfortable corollary If your organisation produces specs with no non-goals, the tooling is not the cause. It is that nobody is willing to be the person who tells sales they are not getting the thing this quarter. That is a leadership problem wearing a documentation costume, and it is yours rather than your PM's.
LiveWhat each question is really asking for4 min

Two sentences, and they are deliberately plain. Neither reads as a challenge, both are unanswerable when the decision was never made, and both take four seconds to ask:

  • "What are we not doing here?" This is check 4, and it drags check 6 out behind it. A team that can answer it has an edge to its scope, and the answer to the follow-up - "and who wanted that?" - is the trade-off, stated out loud by the person who made it. A team that cannot answer it is about to spend a quarter discovering the boundary one ticket at a time.
  • "How will we know it worked?" This is check 3, and it doubles as a check on check 1, because a measurable success statement forces a baseline and a baseline forces somebody to have looked at the current number. "We will see improved engagement" is the answer that tells you nobody has.
  • Ask them in that order. Scope first, because it is the harder one and the room is freshest. If you ask the metric question first you will get a good answer, feel satisfied, and forget the other one, which is the one that would have saved you the quarter.

Neither question requires you to have read the document, which is the practical point. You can ask both while the spec is still on somebody's screen, in a corridor, or in the first minute of a review, and you will learn more than you would from twenty minutes of reading.

Self-studyWhen "no non-goals" is the honest answer3 min read

Occasionally a PM will say there are no non-goals worth writing, and sometimes they are right. Two cases, and one impostor:

  • The genuinely small change. A one-week fix to an existing flow does not need a not-doing list, and demanding one is ceremony. The size test is rough and workable: if being wrong costs less than the meeting to discuss it, stop discussing it.
  • The deliberate exploration. Sometimes the team is buying information rather than shipping an outcome, and the honest spec says so: this is a two-week probe, here is what we expect to learn, here is the decision it feeds. That is a legitimate document with a different shape, and it should say which decision it is a spec for.
  • The impostor: "everything is in scope for now." This is not an absence of non-goals, it is an unmade decision described in a reassuring way. It reliably becomes three extra tickets by week four, each individually reasonable, none of which anybody chose.

The useful discipline is not that every spec has non-goals. It is that no spec is silent about them. "We considered and excluded nothing, because the change is a week" is a perfectly good line, and it is a decision.

Part 3 · who signs, and what signing means

Three approvals, three different products 7 min live

Here is the failure that costs the most and looks the most like success. A spec is circulated. Three senior people read it, and every one of them approves, quickly and sincerely. Nobody disagrees. Six weeks later the disagreement arrives anyway, and it turns out that the three of them approved three different products, because the document was fluent enough to support all three readings and specific enough to rule out none of them.

One sentence. Three approvals. Three products. The line everybody signed off: "Summaries will surface the key outcomes of a meeting." Sales read it as something visible during the call, because "surface" sounds like it happens live Engineering read it as a paragraph generated after the call, from the transcript we already store Design read it as a list of decisions, pinned to the top of the meeting page, editable by the admin Nobody disagreed, because there was nothing specific enough to disagree with. Fluency created the consensus. The document was agreeable precisely where it should have been decisive, and the argument was not avoided - it was moved to week six and made more expensive. The fix is not a longer document. It is one named signer and one reading. Three approvals distribute comfort. One signature concentrates accountability, which is the point of it.
🔍 Click to zoom - agreement is not alignment, and fluency manufactures the first
LiveWhat a signature on a spec actually means4 min

Most organisations have never said this out loud, which is why approval is cheap and accountability is vague. Four sentences fix it, and they are worth agreeing with your team once rather than relitigating per document:

  • One name, not a list. A spec has exactly one owner, and it is the person who will answer the question "why did we build this" in two quarters. Everyone else on the document is a reviewer. Reviewers can object; only the owner decides, and only the owner is accountable.
  • Signing means "I read the six and I agree with them". Not "I have no objection", not "this looks fine to me". If a signer cannot say what the non-goals and the trade-off are without scrolling, they have not signed anything, they have waved.
  • Signing does not mean the outcome is guaranteed. This is the part that makes honest specs possible. The owner is accountable for the decision being well made on what was known, not for the world cooperating. Blur those two and the next spec you get will be written to be un-blameable, which is the worst kind of document.
  • Drafted with a tool changes nothing about who is accountable. Whoever signs, owns it, exactly as if they had typed every word. That principle needs stating once, plainly, before the first time somebody is tempted to explain an error by saying the draft came out that way.
The question that tests all four at once "If this goes wrong, whose call was it?" Asked in a review, in a neutral voice, it takes about three seconds and it is unanswerable when a document has three approvers and no owner. Ask it early enough and the answer is a name. Ask it late enough and the answer is a silence.
LiveYour own contribution to the problem: the spec by Friday4 min

This is the part of the session that is about you rather than about your PMs, and it is short because it does not need decorating. Consider what happens when you ask for a spec by Friday and get one.

  • You asked for a document, and a document is what a deadline can produce. Drafting is now fast enough that Friday is always achievable. The six decisions are not faster than they were: the trade-off still needs a conversation with sales, the non-goals still need somebody to accept a smaller scope, the metric still needs a baseline that somebody has to go and pull. None of that fits into the gap between your request and your deadline, and all of the writing does.
  • So the deadline selects for the half that got cheap. You will reliably receive the packaging on time, with the decisions missing, and it will look like responsiveness. This is not your PM cutting corners. It is you having specified the artifact instead of the decision, and them delivering exactly what you asked for.
  • A spec produced to a deadline is a description. A spec produced to a decision is a spec. The repair is one sentence in how you ask. Not "can I have a spec by Friday" but "by Friday I want to know what we are not doing and what we gave up - the document can follow".

The same repair works upward. When somebody asks you for a decision by Friday, the useful reply is to ask which decision, and what they will do differently once they have it. Half the time the honest answer moves the date, and the other half it removes the need for the document entirely.

Self-studyApproving faster is not the same as approving well3 min read

A quiet consequence of cheap drafting is that more documents reach you, each one better formatted than the last, and your available attention did not increase. Three habits keep that manageable without turning you into a bottleneck:

  • Read the non-goals first, then the trade-off, then stop. If those two sections are real, the rest of the document was almost certainly written by somebody who had decided things. If they are absent, nothing else in the document can compensate, so there is no point reading on.
  • Approve the decision, not the document. Say what you are agreeing to, in one sentence, in the thread. "Agreed: summaries for paid-workspace admins this quarter, not live notes, and we accept that sales loses the demo moment." If your sentence and the PM's spec differ, you have found the three-readings problem for free.
  • Let the small ones through without ceremony. A standard applied to everything at the same intensity gets abandoned within a month. Reserve the two questions for anything sized in weeks, and let the rest go by. Consistency on the things that matter beats thoroughness everywhere.

None of this makes you a better reader of documents. It makes you a person whose approval means something specific, which is a different and considerably more useful thing to be.

Discussion exercise · 9 min · everyone in the room

The two-question drill ★ run this on a spec you already approved

The whole point is that it works with the document closed. Everybody brings a spec they approved in the last quarter and did not write, and the first two questions get asked before anybody opens the file. Rehearsing it once in a safe room is what stops it sounding like an accusation the first time you do it for real.

Everyone names one spec they approved in the last quarter. Approved, not written - the reviewing seat is the one you actually occupy. Ideally something that shipped, so the answers can be checked against what happened.

Do not open it. Say the title out loud and nothing else. This is the constraint that makes the exercise honest, because it removes the option of scanning for the answer while somebody is asking the question.

Ask the two questions, in order, and time the answers. "What are we not doing here?" then "how will we know it worked?" Say nothing in the gaps. The silence is data and filling it is the most common way leaders destroy their own exercise.

Now open the document and find where each answer was meant to live. Write two words on a page: present or absent. Not "sort of present" - the value of the drill is that it produces a binary, and a hedged answer is the same as an absent one.

Ask the third question and get a name. "If this had gone wrong, whose call was it?" Then say, out loud and in one sentence, what you will require before you approve the next one. Commit to the wording in the room so it sounds like a standard rather than a reaction.

LivePrompt 1 · the scope question, word for word3 min
Say this "Without opening it - what are we not doing here?"

What a good answer sounds like. Immediate, specific, and slightly uncomfortable. "We are not doing live in-meeting notes, we are not touching the enterprise export, and we are not covering meetings with no clear decisions in them - those get an empty state rather than a generated guess." Three exclusions, named, delivered in under ten seconds because they were argued about rather than drafted. Then listen for the sentence that follows unprompted, which is usually the trade-off arriving on its own: "sales knows, and they are not happy about it". That volunteered second half is the strongest signal available to you in this whole session.

The failure mode to listen for. The reframe. "Well, we are focused on summaries" is not a non-goal, it is the goal restated in a way that excludes nothing. Its cousin is "nothing is really out of scope at this stage", which sounds flexible and means the boundary will be discovered by an engineer in week four. And listen for the pause followed by an answer constructed on the spot: that answer is being decided in the meeting rather than recalled, which tells you the decision did not exist until you asked. That is still a win, it is just a more expensive one.

LivePrompt 2 · the success question, word for word3 min
Say this "How will we know it worked? And what is the number today?"

What a good answer sounds like. A metric, a baseline, and a date, in that order and without hedging. "The share of recorded meetings where somebody opens the summary in the first week. It is 11% for transcripts today and nobody has ever set a target, so we are setting one at the review in four weeks." Note what makes that answer good: it contains a number that exists now. A success criterion with no current value is a wish, because you cannot tell afterwards whether anything moved.

The failure mode to listen for. The directional phrase - "improved engagement", "better retention", "higher satisfaction" - which is what gets written when nobody supplied a metric, formatted to look exactly like a metric. The subtler failure is a real metric with no baseline, which passes a quick read and fails six months later when two people argue about whether 14% was good. Do not accept "we will figure out the measurement during the build", either. That is the same sentence as "nobody will be wrong later", which sounds safe and is the opposite.

LivePrompt 3 · the signature question3 min
Say this "If this had gone wrong, whose call was it? And what were the rest of us agreeing to?"

What a good answer sounds like. One name, said without hesitation, followed by a one-line summary of what that person decided. "Mine. I decided summaries over live notes for paid-workspace admins, and the rest of you were agreeing that the evidence supported it." Then everyone else in the room says what they thought they had approved, in one sentence each, and the sentences match. When they match, you have alignment. When they do not, you have just found six weeks of future argument in ninety seconds.

The failure mode to listen for. The plural - "it was a team call", "we all agreed" - which is how a document with three approvers and no owner describes itself. Nobody is lying; the accountability genuinely was never assigned. Watch also for the reverse, where a junior PM is named as the owner of a decision they had no authority to make. If the trade-off cost sales a feature, the person who can absorb that conversation is the signer, and pretending otherwise sets somebody up. Naming the wrong owner is worse than naming none, because it looks resolved.

Real world

One line, three products, and nobody in the wrong. Cadence's summaries spec was approved by three people on the strength of a sentence that read "summaries will surface the key outcomes of a meeting". Sales approved a thing that appears during the call, because "surface" implied something happening while they were demoing. Engineering approved a paragraph generated afterwards from the transcript already in storage. Design approved a list of decisions pinned to the top of the meeting page and editable by the admin. All three readings are defensible from the sentence, all three people were paying attention, and the disagreement surfaced in week six as a conversation about what had been agreed, which is the most expensive conversation a product team can have.

What would have caught it costs one sentence each. Not a longer document - a longer document supports more readings, not fewer. If each approver had written one line in the thread saying what they were agreeing to, the mismatch would have been visible the same afternoon, for free. That is the whole mechanism: a spec is one reading or it is not a spec, and the cheapest way to test whether three people share a reading is to make them each state it in their own words before anybody starts building.

Take this with you

Questions to ask your PMs the whole point of the session

Seven questions, and the first two carry most of the value. Ask those two about everything sized in weeks, in the same words, until your team writes the answers into the document before the review - which is the actual goal, and takes about a month.

This week

Five things to do before a4 ◐ conversations, not artifacts

Source material

Sources covered

Full source map in materials/official-course-map.md. This page covers:

PRD and spec practice - the spec as a decision record, and what makes one signableParts 1 and 2 · the six checks, and who notices when each is missing
Scope control through explicit non-goals, and the trade-off that follows from themPart 2 · why it is always those two, and the two questions that surface both
AI-assisted knowledge work - fluency, automation bias, and where accountability sitsParts 1 and 3 · why absence is invisible, and who owns a drafted decision
Research quality and provenance - what evidence a spec must carryCheck 1 only · a2 was the session on believing a finding
Prioritisation frameworks - what a trade-off costs and how options get rankedNamed as check 6 only · a4 owns ranking and the theatre around it
Metric definition, baselines and targetsCheck 3 only, as "what is the number today" · a5 owns metrics and readouts
Delivery: estimation, sequencing, sign-off gates, rolloutOut of scope by design - learn-ai-project-management
Check yourself

Three questions before you go 🎯 ◐ 90 seconds

1 · Of the six checks, which two are almost always missing, and what do they have in common?

Four of the six are descriptions and get written whether or not anybody thought hard. Two are concessions: writing them means a named person loses something they wanted. That is why the same two survive to the top of the practitioner ladder, and why writing the non-goals fixes both at once.

2 · Three senior people approve a spec quickly and nobody objects. What have you learned?

"Summaries will surface the key outcomes of a meeting" was approved by sales as a live in-call view, by engineering as a paragraph generated afterwards, and by design as a pinned decisions list. Nobody disagreed because there was nothing specific enough to disagree with. The test is one sentence each on what they were agreeing to, and it costs nothing.

3 · You ask for a spec by Friday and get a well-structured one on Friday. What is the most likely thing to be missing?

A deadline selects for the half of the work that got cheap. The repair is one sentence in how you ask: request the decision rather than the document - what are we not doing, and what did we give up - and let the writing follow. A spec produced to a deadline is a description; a spec produced to a decision is a spec.

Leader session 3 cheat sheet · pin this

A spec isThe record of six decisions. Sections are where decisions live, not what they are.
The sixProblem evidenced · user named · success measurable · non-goals stated · edges decided · trade-off explicit.
The two missingNon-goals and the trade-off, every time, because both require somebody to give something up.
Your two questions"What are we not doing here?" and "how will we know it worked, and what is the number today?"
Signable testCould an engineer who was in none of your meetings build the right thing without inventing an answer?
Signing meansOne name. I read the six and I agree. Accountable for the decision, not for the world cooperating.
The fluency trapThree approvals, three readings, no disagreement. Fix: one sentence each on what you agreed to.
Your own fault lineA spec to a deadline is a description. Ask for the decision instead. Next: a4, prioritisation without theatre.