Product Field Work All writing
WritingProduct leadership

When the bottleneck moves

As models improve, it’s becoming common for complex features that once took weeks or months to be built in days, and sometimes in hours. That doesn’t make product leadership easier. It moves the hard part from “can we build it” to “should we, does it fit, and can our customers absorb it,” and it asks a team built for the old pace to work a new way without anyone feeling left behind.

The problem

For twenty years, the slow part of software was building it. Product teams wrote detailed requirements because engineering time was scarce, and every sprint had to count. Most of our rituals, from grooming and pointing to three-week sprints, exist to ration that scarcity.

That scarcity is easing. When a working feature takes days instead of a quarter, the rituals built to ration engineering time start rationing the wrong thing. The new scarce resources are judgment about what to build, the coherence of the product, and the customer’s capacity to absorb change.

1 : 0.5product managers to engineers, a staffing ratio one of Andrew Ng’s teams proposed, the inverse of the usual one to six or more
55%faster coding in GitHub’s controlled Copilot study; McKinsey found PMs gained about 40%, mostly on documents rather than discovery
2006the year “feature fatigue” research showed more features win the sale but lose the customer, a finding that matters more when features are nearly free
How I think about measuring the change
  1. Time from a validated idea to a customer using it, not story points or velocity
  2. Share of shipped work that ties to a stated outcome, the anti-grab-bag test
  3. Customer absorption, support contacts and retraining per release, not just adoption
  4. Change failure rate, so speed never quietly trades away quality
Why I wrote it

I lead product in regulated software, where every release carries validation, training and trust costs. I build with coding agents like Claude Code every day, and I see the gap between what’s now possible and how most teams still work. This is my attempt to describe the job on the other side of that gap, including the parts that are uncomfortable.

The ideaKeep the purpose; change the method. Use agents to see far more. Tie every feature to a stated outcome. When requirements leave a gap, trust the team to decide in line with the product’s goals and share the reasoning. Make quality a shared, automated bar. Write every decision into skills so no one has to learn it twice. And govern how much change customers get, not just how fast we can make it.

1What doesn’t change

Before anything about speed, the part that stays fixed. These were true before AI and they’re the only things that keep a fast team from becoming a fast feature factory.

The tenets stay. How we honor them changes.

Switch between how each tenet was practiced and how it’s practiced now.

2Seeing far more

Competitive analysis used to be a person reading a handful of competitors’ sites and release notes when they had a free afternoon. User feedback was a sample. Stakeholder requests lived in inboxes. Now agents can read competitors’ products, help centers, demo videos, reviews and press every day, across dozens of companies, and hand a leader a deep report each morning.

That changes what a product leader is responsible for. It’s no longer “did we look.” It’s “did we notice what mattered, and did it change what we build.”

A typical week of signals, before and with agents

Illustrative numbers for a mid-size software product, to show the order of magnitude. Your own will differ.

MarketCompetitor products, help centers, demo videos, pricing, press, job posts
CustomersReviews, ratings, tickets, call transcripts, usage and churn signals
InternalStakeholder and sales requests, error logs, support trends, subject-matter experts
DirectionCompany goals, board priorities, the executive team’s bets

3Strategy that stops a grab bag

When so much more can be built so quickly, the biggest risk is building everything. Each feature looks reasonable on its own. Together they turn a product into a drawer of gadgets. The defense isn’t a slower team; it’s a visible line from every feature to an outcome the company has chosen.

Does this feature attach to the strategy?

A company goal, the outcomes that serve it, and the bets behind each outcome. Propose a feature below and see where it lands.

Pick a proposed featureEach one either connects to a bet the team has made, or it doesn’t.

The question isn’t “can we ship it this week.” It’s “which branch does it grow on, and would a customer feel it as part of one product?”

The roadmap changes shape too. Quarterly lists of features go stale in days. What holds is a short set of outcomes and bets that the whole company can see, with the features underneath treated as experiments that can be added, changed or removed quickly. Feedback on the roadmap can come in the same way: a living page, reviewed weekly, with the evidence for each bet attached.

4The build path collapses

The old path is a relay. A product manager writes a spec, a designer draws each screen, engineering rebuilds the prototype from scratch, QA tests it, and the whole thing waits for the next sprint. The new path is one piece of work. A product manager builds with a coding agent such as Claude Code, using the design team’s skills and the engineering team’s standards, on a branch that engineers review. The prototype is the code.

One feature, two paths

The same small feature, a new filter on a busy list screen, with typical elapsed time on one shared scale. Press play.

The relay about 5 weeks

Product managerDesignerEngineers, 2–3QAScrum master

One piece of work about 2 days

Product manager with a coding agentEngineer reviewingCustomers, earlier

Illustrative. The new path assumes design and engineering standards are already encoded as skills, and automated tests are part of every change.

Notice what the new path still has: a review by an engineer, automated tests, a feature flag, and customers. Speed comes from removing handoffs and rebuilds, not from removing care.

5What building with a coding agent actually looks like

“Product managers write the code” sounds reckless until you see the steps. The product manager builds; engineering and QA still review what matters; customers see it earlier. Here is one feature, end to end, with the handoffs marked.

One feature, step by step

Who does each step, how long it usually takes, which shared skills the agent applies, and where the work passes to someone else.

    When engineering takes the wheel insteadNew or changed data schemas · new services or shared APIs · performance-critical paths · anything touching authentication, permissions or audit trails · work that crosses team boundaries. In these, the product manager’s branch becomes a working reference, and an engineer leads the build.
    Five questions before anything ships
    1. Should we build it, and what evidence says so?
    2. Which outcome does it serve?
    3. Does it feel like one product with everything else?
    4. Is this the best version, or just the first one that worked?
    5. Can our customers absorb it right now?

    When building costs so much less, “good enough” stops being a trade-off and becomes a choice. There’s very little excuse left for shipping anything less than the best version we can imagine.

    6The team’s shared memory

    The most important tactical change is the least visible. Every decision the team makes, about the customer, the design, the code, the domain or the quality bar, gets written into skills that the team’s coding agents apply to every future feature. Correct something once, and it stays corrected.

    That turns a familiar frustration into a process failure. If someone has to ask “didn’t we already agree on this?”, the fix isn’t another review comment. It’s updating the skill so the question never comes up again.

    How a signal becomes something every feature knows

    The team’s skill library, and the loop that keeps it current. Press “Next signal” to follow one through.

      Repeat corrections per featureIllustrative. The number of times a known decision has to be re-explained should fall toward zero.

      What current practice says about skills

      Skills are now an open standard that works across coding agents, not a single vendor’s feature. These are the practices I’d hold a team to, from Anthropic’s engineering guidance and the people building on it.

      StructureLoad only what’s neededA skill is a folder with a SKILL.md file. Agents see only its name and description until a task needs it, then read the body, then any bundled reference files. That’s why a team can keep dozens of skills without drowning every task in context.
      PortabilityOne library, any agentAgent Skills were published as an open standard in December 2025, and coding tools beyond Claude Code implement the same pattern. The team’s memory shouldn’t be locked to one tool.
      WritingOnly what the agent doesn’t knowKeep each skill short and specific to your product. Explain why a rule exists rather than shouting MUST, so the agent can apply it to cases the rule didn’t spell out.
      TestingEvaluations before the skillWrite a few realistic test tasks first, measure the agent without the skill, then add just enough to pass. One agent drafts the skill; a fresh one is tested with it.
      LearningCapture what worked, and what went wrongAfter real work, ask the agent to write its successful approaches and common mistakes back into the skill. That’s the “correct it once” loop, made routine.
      ScaleThe setup matters more than the modelAt large engineering organizations, the tooling around the agent (skills, project instructions, hooks, shared plugins) shapes results more than the model alone. Good skills spread by being packaged and shared.
      Skill gardening is now real work
      Every dayAgents propose skill updates from meeting transcripts, customer calls, competitive reports, review comments and incidents. Nothing changes until the skill’s owner approves, and every change has to pass that skill’s evaluations.
      Every weekEach owner reviews proposed changes, merges what’s been decided, and resolves conflicts between skills, with the product leader breaking ties.
      Every quarterPrune. Watch which skills and files agents actually read, retire rules nobody needs, merge overlapping ones, review any skill that came from outside the team, and check that the product tenets still read true.

      The energy that used to go into correcting the same thing in every review goes here instead. It’s the difference between a team that repeats itself and an engine that gets smarter with every feature.

      7Decide, then disclose

      A lot of delay hides in small gaps. The spec didn’t say what happens when someone closes a dialog with unsaved changes, so work stops, a question is logged, and it waits for a meeting. With the product goal and context available to the coding agent, the person building can make a well-reasoned call in minutes, in line with the product’s goals, and share it.

      The same missing detail, two ways

      An engineer hits a gap in the spec on Tuesday morning.

      Blocked and waiting

      Decide and disclose

      This only works if engineers know the customer and the goal, not just the ticket. That’s the real shift: from “implement exactly what was specified” to “understand why we’re building it, make good small decisions, and flag them.” Most gaps in a spec were never truly blocking. The process just required someone else to sign off before work could continue.

      8Every role, still essential

      None of this works if one department feels it’s being written out. Every role changes. None disappears. The work moves up: from producing each artifact by hand to encoding judgment once, so everyone’s work carries it.

      Who does what now

      What each role did, what it does now, and why the team is worse off without it.

      9Bringing engineering along

      Here’s the honest version. When product managers start shipping code with coding agents, many engineers feel two things at once, often without saying either. They worry their job is going away. And they worry they’ll be the ones blamed when fast, AI-written code breaks in production. So they pump the brakes: more requirements, more grooming, more questions before starting. That isn’t bad faith. It’s a reasonable response to carrying the quality risk alone.

      The answer is to take that weight off their shoulders, not to push harder.

      A shared quality bar, written before anyone speeds up

      So engineering and QA aren’t the only ones holding the line.

      Written quality guideCoding standards, security rules and review criteria, encoded as a skill every coding agent follows, for everyone.
      Tests in every changeThe agent writes unit and end-to-end tests with each feature; QA owns the suite and what “covered” means.
      Review by an engineerEvery product-built branch is reviewed. Engineers own architecture, data schemas and anything risky.
      Flags and fast rollbackNew work ships behind flags. If something breaks, it’s off in minutes, and everyone owns the fix.

      Quality stops being engineering’s burden and becomes the team’s standard. That’s what lets engineers be excited about speed instead of afraid of it.

      10How fast should we go?

      If we can release twenty features a week, should we? Usually not. Every visible change asks something of customers: attention, relearning, trust. In enterprise and regulated software it also costs retraining, documentation and sometimes validation. Just as always-on agents need a governor on spending, a product needs a governor on change.

      A change budget

      Pick where the product is and who uses it. The recommendation is a starting point for a conversation, not a formula.

      Product stage
      Who uses it
      –

        Two things make the budget go further. Most improvements can be invisible to users: speed, reliability, fewer errors. And visible change can be bundled into fewer, well-explained releases instead of a constant drip. An early product should sprint to its first plateau. A mature enterprise product should mostly refine, and make each visible change count.

        11Where the leader’s week goes

        A product leader’s week, then and now

        My own estimate of how the time shifts. Hover or tap a segment for the detail.

        Status meetings, spec writing and grooming shrink. Time with customers, strategy, reviewing what shipped and coaching the team grows. The leader also takes on a new job: curating the context the whole team and its agents work from, including goals, customers, principles and the design and quality skills. That context is now the most leveraged thing a product leader produces.

        The transition

        Teams don’t switch methods overnight, and they shouldn’t. This is the path I’d take with a team used to detailed requirements and three-week sprints.

        Days 1–30Lay the floor
        • Write the quality guide and test standards with engineering and QA, and encode them as skills.
        • Stand up the skill library: product context, personas, design, engineering, quality, domain rules, customer voice, market, decisions. Name an owner for each.
        • Pick one real feature and build it end to end the new way, together.
        Days 31–60Change the defaults
        • Make “decide and disclose” the norm for small gaps.
        • Product-built branches go through engineering review.
        • Share the goal and the customer evidence with every piece of work.
        • Start weekly skill gardening: agents propose updates from meetings, calls and reviews; owners approve.
        Days 61–90Retire the rituals
        • Replace story points and grooming with weekly outcome reviews.
        • Set a change budget per quarter with support and customer success.
        • Measure what matters, including repeat corrections, and show the team the results.
        Idea to customerTime from a validated idea to a customer using it
        Tied to outcomesShare of shipped work linked to a stated bet
        Change failure rateReleases that needed a rollback or fix
        AbsorptionSupport contacts and retraining per release
        Repeat correctionsTimes a decision already made had to be explained again

        Choices

        Encode judgment, don’t repeat itDesign and engineering put their thinking into skills once, so every feature carries it, instead of reviewing the same issues by hand forever.
        Decide and disclose over wait and askWhen requirements leave a gap, whoever is building decides in line with the goal and records why. Only real trade-offs wait for a group decision.
        Govern change, not speedThe team can go as fast as it likes internally. What reaches customers is budgeted.
        Where I’d push back on my own argument

        For some features, building in hours is realistic now; for many, days is. Either way, knowing it was the right feature still takes time with customers, and discovery doesn’t compress the way coding does. Product managers writing real code need real guardrails, and some teams won’t have the test coverage to start safely. In regulated software, validation is part of the product, not overhead. And the ratio debate is unsettled: several practitioners argue more PMs won’t fix a discovery problem. I think they’re right that the fix is better judgment, not more headcount.

        Sources

        1. Forbes Technology Council, March 2026: Andrew Ng’s team proposing one PM for every 0.5 engineers.
        2. Product Leaders Day India: Ng on the “builder’s block” and the six-to-one ratio.
        3. Allstacks, March 2026: GitHub’s 55% Copilot result and McKinsey’s 40% PM productivity finding.
        4. Notes on Marty Cagan’s “new standard” talk, April 2026: feature factories struggle more with AI; automated delivery leaves more time for discovery.
        5. SVPG, Product in the AI Era resource guide, and A Fresh Definition of the Product Role, August 2026.
        6. Product Compass, July 2026: most ideas still fail; discovery matters more when shipping is instant.
        7. Rust, Thompson and Hamilton, “Defeating Feature Fatigue,” HBR 2006, via the University of Maryland.
        8. Michele Galli, June 2026, and Lovex, August 2026: counterarguments on the ratio.
        9. Anthropic Engineering, “Equipping agents for the real world with Agent Skills”: skills as folders with SKILL.md, progressive disclosure, and capturing successful approaches and common mistakes back into skills.
        10. Anthropic, Skill authoring best practices: concise skills, evaluations before documentation, and one agent authoring while another is tested.
        11. Anthropic, Agent Skills overview: staged loading and using skills only from trusted sources.
        12. SwirlAI, March 2026: Agent Skills released as an open standard on December 18, 2025, and implemented by coding agents including Claude Code and Cursor.
        13. Generative Programmer, April 2026: authoring patterns, including explaining the why instead of all-caps rules.
        14. Anthropic, “How Claude Code works in large codebases,” July 2026: the harness around the model shapes results more than the model alone; skills load by relevance; plugins spread what works.
        15. Anthropic Engineering, “Effective context engineering for AI agents”: treating context as a scarce resource to curate.