Kill List

What to stop doing, and how to run the kill so the org actually lets go.

• Is a Slack feedback channel an FDE to product loop?Salesforce's Agentforce org routes FDE feedback into a Slack channel with two tags. A channel is an inbox. The loop is a packet, dates, and a decision back.
• Is a hard budget cap the right cost control for AI agents?Salesforce, Snowflake, and Atlassian shipped the agent budget cap in one week. The cap fires after the spend. Re-quote the change that sets a cost-bearing key.
• How many skills should a PM skill pack auto-activate?Paweł Huryn's pack auto-activates 39 of 43 skills. Figma's PM OS has 31. In my library 30 of 51 commits have no author. Auto off by default, count fires.
• Should an AI-native CSM fix the customer's agent themselves?No. The CSM triages to the layer, names the keys holder at the customer whose fix it is, and knows the recall rule. Fixing it makes the CSM a second FDE.
• Where should AI inference cost sit on the P&L?Serving cost belongs in cost of revenue, split by feature. Anthropic's S-1 books compute as opex, and its margin reads 85% or about 45% as a result.
• What should a design leader build with their own hands each month?Not a component or a prompt. A constraint set for one live surface, with 20 real cases, and the percentage of surfaces covered as the number you report.
• Should product orgs still hire junior PMs when AI does the entry-level work?Yes, on purpose. Anthropic's junior tech hires fell from 27% to 9% and PM L1 share fell to 11%. A CPO who stops hiring at the bottom loses the bench.
• Should thumbs-up and thumbs-down feedback train your AI model?Thumbs helped make GPT-4o sycophantic. A thumb is one bit over a whole output. Join it to a rubric dimension, an output id, and a slice before anything trains.
• How should a business handle traffic from personal AI agents?Salesforce says personal agents will call every few hours; meet them with business agents. Publish the field they are polling at the URL the person sees.
• Should you build or buy your AI agent harness?Ramp built Laminar in six weeks. The build column collapsed, landing did not. Three lines before the memo: customer, trace reader, on call. A blank means buy.
• Who owns the eval harness, the CPO or the CTO?Both, split cleanly. The CTO owns the platform that runs and enforces it. The CPO owns the bar it checks. Both sign the gate to production.
• Who should fix a live AI agent after it is handed to the team?The team that uses it: a named person with hours, a versioned layer with an undo, and a go-live eval set the builder keeps frozen and reruns on every change.
• How do you measure the cost of reviewing AI agent output?Not by review time alone. Count the outputs nobody opened next to the outputs generated. 13 of my 39 agents produced work nothing ever read.
• Should an AI agent launch lead with its engagement rate?No. Engagement is true of any chatbot a visitor cannot get past. Lead with resolution under a stated definition, scored three ways, and the hand-off count.
• What should a CEO hold a CPO to on AI?Not dated AI features or hours saved. A growth measure per AI bet, cost per outcome, how product's own work changed, and one kill a quarter.
• Should long-running agents be priced on the final outcome?No. Keep the finish line as the renewal number and price the leg the customer will sign. One definition drew 200 disputes a week; a relay has one per handoff.
• What does a CPO own when prototypes are cheap?Not the roadmap. The operating model: the substrate, the listening loop, the eval bar, the org shape with the Product Builder as its unit, and the margin.
• Does keeping a human in the loop make an AI agent safe?Only if the gate is rare, carries context, and is measured. Past 95% approval or under 2 seconds per decision, the human has left the loop.
• Is price per call the right way to compare a decision model like Jev?No. Price per call is what you pay when the pick is right. Rank each branch by what a wrong pick costs to undo. My classifier fell from 89% to 61%.
• Should you build an internal tool because an AI agent can build it in 20 minutes?No. Build time stopped being a filter. I approved 39 cheap agent builds in 80 days and 13 became orphans. Gate on owner, workflow, and a switch-off date.
• Should a product manager hide schedule buffer?No. Plan at 70% of capacity and say so. A hidden buffer works once per stakeholder. Publish ETA variance on every date and report what consumed the buffer.
• Can AI turn a support cost center into a revenue engine?Only when unserved demand sits next to the automated work. IKEA had it: 8,500 staff, a €1.3 billion channel. Most support queues do not.
• Is commit drawdown a good measure of AI adoption?No. It measures spending against a commitment. I had 12 accounts paying committed minimums and using less than the minimum. Report outcomes bought instead.
• Should a SaaS vendor charge per API call for agent access?Cap the rate this quarter, let registered agents replace seats over the migration, price the outcome at month 18. A per-call meter on the seat skips all three.
• Is being the system of record still a moat?ServiceTitan cut off Podium, kept 1,000 customers, and added no revenue. The record is a retention moat, not a growth engine. Two lines replace the slide.
• Which design rituals should you kill, and what replaces them?Nine design rituals that produce a receipt instead of a decision, each with its replacement, plus the four-step protocol that makes a kill stick.
• Are OKRs still worth it for AI products?OKRs were built for weekly builds and monthly signal. For AI products with daily signal and four-hour prototypes, replace them with a Rolling Outcome Ledger.
• How do CPOs win the AI transition without stalling?The soft pivot is a fantasy. Win with a sunset date, rewritten comp, a pre-sold gross-margin trough, and the seven-decision sequence run in order.
• How do you decide whether to cannibalize your own product?Four diagnostic questions route the call to sunset, refresh, or split. Architecture, ICP, gross-margin trough, and buyer decide, not comfort level.
• How do you deprecate a feature?Deprecate on signal, not politics. Write a kill condition at launch, pick one of three tiers, and send individual notices to users from the last 90 days.
• How do you kill a feature without breaking trust?Kill a feature by running discovery on non-adopters first, then reorganizing before you rip out code. One kill lifted platform activation 15%.
• How do you run empowered teams when the CEO doesn't buy in?You don't need permission to run empowered teams. You need evidence. Four tactics, repeated for six months, until the results make the case for you.
• How do you sunset a pricing tier?Sunset a pricing tier over 18-24 months with a three-wave customer migration, a comp plan rewrite, and a post-sunset reorg. Five patterns kill most sunsets.
• Is spec-driven development (IDSD or SDD) worth it?SDD and IDSD are the same methodology layer with different stickers. Skip both. The prototype is the spec, the eval is the acceptance test.
• Should you kill AI office hours?AI office hours are 2026's agile transformation: high attendance, zero shipped output. Kill them and run paired shipping, eval reviews, and kill list reviews.
• Should you kill the PRD entirely?The PRD was a 2015 fix for a 2015 problem. Replace it with three artifacts: a working prototype, a one-page eval rubric, and a one-page README.
• Should you kill the roadmap?Kill the quarterly roadmap and run a one-page live bet portfolio instead: 5-7 active bets, each with a hypothesis, signal, kill condition, and decision date.
• Should you kill the status meeting?The status meeting exists because nobody trusts the dashboard. Replace it with one live product page and reclaim about six hours a week per PM.
• Should you still run an APM program?LinkedIn's CPO killed the APM program. Replace the two-year rotation with a 12-week Product Builder apprenticeship: four prototypes, one production ship.
• Were empowered product teams just a ZIRP feature?Empowered teams were a luxury of cheap capital, not a law. Expensive capital killed patience, AI killed the constraint. Autonomy survives, ceremony does not.
• What is a Kill List and how do you run one?A Kill List is the standing counterpart to a roadmap: a maintained list of things to stop. Run it on a cadence with one entry test: would we start this today?
• What is the anti-backlog?The anti-backlog replaces a bloated backlog with a live queue capped at two weeks, a hypothesis library, and a kill list. Migrate in four weeks.
• What should you do instead of a feature request queue?Replace the feature request queue with signal synthesis, prototype-first investigation, and a seven-item outcome bet ledger. Your 847-item backlog is debt.
• What is wrong with the 'I built it in a weekend' flex?Speed to demo is a vanity metric. The weekend-build flex measures the cheapest, most brittle thing you make. Ask if it was still running the weekend after.
• Why can killing a product beat launching one?A good kill often creates more value than a mediocre launch, but the org is blind to it. Killing well is the rarest and most valuable product skill.
• Why is a reorg a product decision?Conway's Law means your org ships its communication structure, so a reorg sets your roadmap ceiling before a ticket exists. The org chart is a product spec.