We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
When the whole company runs on Claude
A senior PM three months into a new job described a company where "every single person" must use Claude heavily and usage is monitored. Questions get answered with "ask Claude." Engineers may only make small edits to Claude's output . What frightens them is that it works. In under a day they pulled user feedback and metrics, checked them against a new idea's assumptions, forecast outcomes, specified an MVP and published the PRD . A project now underway will feed surveys, feedback sessions and screen recordings into Claude to produce a prioritized roadmap. If it works, they expect the PM function to be cut .
The replies offered more useful ways to frame it than reassurance:
- What survives. One commenter said that at all-in companies they've looked at, two parts of PM work lasted: deciding what gets built, and checking output against what customers actually said. When nobody owns that second part, "quality drifts quietly for months." Their advice is to ask leadership whose name goes on the postmortem when an LLM-written spec ships a bug .
- Output isn't outcomes. Another PM said their company shipped 600 experiments in Q1 "and didn't change a single customer outcome" .
- Quality gates in practice. A practitioner pairs every agent with an "interrogation agent" that assigns confidence scores, sending low scores to human review. They test changes in batches of 10 before pushing 30,000-row updates, and they treat investigating mistakes as their own job .
- Where cuts may land. One commenter predicts product ops and junior PM roles will shrink, one PM will cover two or three squads, and senior strategic PMs will hold their value "because judgement becomes the bottleneck once synthesis is cheap." This is a forecast, not data .
Rekhi's list of process shifts
Sachin Rekhi lists what "AI-pilled" teams are changing. Specs are giving way to prototypes, which lose less in translation . Ad hoc research is becoming continuous discovery . PMs are moving from passing information around to building . Planning goes out a quarter at most, because capabilities can't be predicted further ahead . Acceptance criteria are being replaced by evals, which he calls "a necessary skill for all PMs" . Stripe's head of design added a counterweight: using AI only to make existing work faster misses the point. Her talk at Lenny's Summit urges having a point of view, encoding standards, and not confusing done with good .
Agents as buyers and builders
SaaStr gave Claude's beta Salesforce connector, Replit and Lovable identical prompts. All three got the facts right; they differed on what a seller should do next . Only Lovable ranked rescuing overdue deals first . Each tool asks for confirmation or limits itself to read-only access before writing back to Salesforce . The hosts also described a vendor's agent that deleted almost all of their Salesforce opportunities. A daily snapshot taken on Replit is what let them restore the data .
The bigger point for product and growth teams: an agent with app context picked Render for hosting without hesitation. For a CRM, it shortlisted HubSpot and Attio for a fresh start but advised against leaving Salesforce "without a clear reason" . The host calls this "agent-led growth" and urges vendors to monitor how agents rank them .
Buyers wanted the savings, not the AI
On a16z, Greer described pitching its AI reliability layer to customers who interrupted to ask about the 90% cut to their Datadog bills: "That's what we want to buy right now." Coding AI sells easily because engineering productivity is a large budget line. Operations value is harder to put a dollar figure on .
Careers
- Pay by AI focus: Aakash Gupta sorted 1,177 PM postings. Median base pay was $345K for model PMs, $258K AI-native, $231K AI platform, $222K applied AI and $216K outside AI, with equity excluded . Of 46 model PM roles, 76% mention evals, one mentions a PhD and eight ask for Python or SQL .
- GitHub as a signal: Gupta says only 24% of PM candidates have a GitHub, while his recent placements at OpenAI, Anthropic and Meta AI all did. What matters is one real-problem project, documentation that shows product thinking, and steady activity .
- Job search: Commenters describe slow hiring, a preference for candidates from the same industry, and referrals from the hiring team as "worth gold" . One person got a single interview from about 150 cold applications, while recruiter outreach led to interviews and offers .
- In an out-of-box Salesforce bake-off using identical prompts, Cloud Force, Replit and Lovable all produced working dashboards and broadly similar factual answers, but their top-five seller actions and prioritization differed; Lovable alone surfaced rescuing overdue deals as its top action. The comparison suggests evaluating agent products on recommendations and information hierarchy as well as connector and data correctness.
- The products had use-case-specific trade-offs: Cloud Force was in beta and restricted to read-only use despite having write capability, while Replit and Lovable asked for confirmation before writing to Salesforce. Replit and Lovable could connect to other apps; the hosts saw Cloud Force as more Salesforce-contained and potentially better suited to organizations seeking tighter guardrails.
- For this bake-off, reported platform/session figures were about $11.81 for Replit, $2.33 for Lovable, and an estimated $15 in Claude usage for Cloud Force, included in a Max subscription; model choices differed, and Salesforce API/MCP charges were still unknown.
- Agent recommendations may influence vendor selection: Saster Connect’s agent recommended Render for its workload, while 10K shortlisted HubSpot and Adio for a fresh CRM setup but advised against migrating existing Salesforce without a clear reason. The hosts argued that agents with application context can become a consequential route to product discovery and selection.
- In an internal agent-to-agent setup, 10K sent Muse daily briefs through a shared API-backed repository, with an agreed keyword required before Muse could take actions; separately, an unnamed connected agent deleted Salesforce opportunities, which 10K restored from a retained snapshot. The examples underline the value of explicit action controls and recoverable data when agents can write to business systems.
- LLMs can compress PM execution: the original poster says Claude took customer feedback and metrics through assumptions, forecasts, an MVP specification, and a published PRD in under a day. The harder-to-automate PM work is choosing what to build and using customer and market context to make and validate decisions.
- One commenter forecasts that AI will put product operations and junior/associate PM roles—whose work is often synthesis—under the most pressure, while expanding PM coverage and increasing the value of senior strategic judgment; this is a projection, not an established outcome.
- Measure customer impact, not AI-enabled activity: a PM reported shipping 600 experiments in Q1 without changing a single customer outcome.
- Keep human accountability and quality gates explicit. One practitioner describes using a second agent to challenge agent outputs, confidence thresholds to route work to production or human review, and test batches of 10 before updates to 30,000 rows; they also say investigating mistakes remains their responsibility. A useful leadership question raised in the thread is what Sr PM success means beyond LLM usage and who owns the postmortem when an LLM-written spec ships a bug.
- The speaker’s AI-assisted production workflow starts with an agent preparing a PR/design doc for tech-lead review; the human checks algorithms, task division, and PR dependencies before authorizing parallel work, then consolidates and reviews the results. A complex data-lake auto-promotion feature reportedly took about 10 days, compared with roughly six months for one person previously. Code reviews also shifted toward tests and golden expectations.
- Greer paired a telemetry-cost product (presented as reducing costs by 90%) with a later AI troubleshooting layer, but customers on sales calls prioritized the bill savings over the AI reliability feature.
- For operations AI, Greer’s approach is to detect new or changed patterns in logs and pass those signals to AI for investigation, rather than starting with an alert and asking AI to backtrack to a root cause. The speaker said it was being run internally and in a couple of customer environments; some ignored signals preceded production incidents, an encouraging but limited early report.
- Across 1,177 open PM postings, median posted base was $345K for model PMs, compared with $258K for AI-native, $231K for AI-platform, $222K for applied-AI, and $216K for non-AI PMs; equity was excluded. Model PMs work with research teams to set model-improvement targets, define evaluations, analyze failed runs, specify behavior, make ship decisions, and launch models. Of 46 open model-PM postings, 76% mentioned evals, while only one mentioned a PhD and eight asked for Python or SQL. A suggested way to practice evals is to choose a task a model struggles with, record 10 failures and their correct answers, then compare two models.
- Aakash highlights Addy Osmani’s view that a decent model with a strong harness can outperform a great model with a poor one. The proposed agent-building roadmap treats the harness as everything beyond the model and includes instructions, product/user/data context, on-demand skills, memory, permissions, tools, checks, and a completion loop; start with your own work, where mistakes are cheap.
- Aakash says 24% of PM candidates have a GitHub, while the latest PMs he placed at OpenAI, Anthropic, and Meta AI all had one. His recommended portfolio signal is a real-problem project, documentation that demonstrates product thinking, and consistent activity; copy-pasted code without tradeoffs, projects that prove little, and empty-commit activity are anti-signals. He frames self-built AI product work before receiving the PM title as “synthetic AI PM experience.”
- An IDFC First Bank PM with 2.5 years’ experience reported very few callbacks and no real interview traction after three months of searching for fintech or consumer-tech roles, despite regular applications, cold outreach, and an updated résumé.
- A commenter attributed the difficulty to slow hiring and employers favoring candidates from similar industries, and described hiring-team referrals as especially valuable; they said cold outreach may still work slowly and suggested banking PMs consider adjacent payments/fintech roles. Another commenter recommended networking more and applying less.
- A separate job seeker reported that about 150 cold applications yielded only one interview (for a Meta rotational PM role), while recruiter inbound of 2–4 contacts per month progressed to interviews and offers for them.
Lenny framed a talk by Stripe’s Head of Design as advice on preventing “zombie UIs.” The speaker argues that using AI only to make existing work faster misses its greater potential: raising the ceiling by having a point of view, encoding standards into the machine, not confusing “done” with “good,” and unleashing creativity and artistry.
For AI-enabled product work, don’t just use AI to make existing work faster; raise the ceiling by giving it a point of view, encoding standards, distinguishing “done” from “good,” and making room for creativity and artistry.
Your360 AI expected debriefs to last about 20 minutes, but its first customer exceeded Retell’s default 60-minute limit and broke the product; the median debrief is now 44 minutes, with some lasting 90 minutes . To support long calls, the team used state-machine orchestration, an interactive on-screen report, and a RAG layer for questions about past calls that keeps feedback anonymous .
A senior PM with about eight years of experience, mostly at smaller companies, says they are getting interviews but not offers and struggle to describe their work in expected PM terms such as strategy, product sense, and metrics; they also cite lacking big-tech experience as a source of imposter syndrome. One commenter described a tough market from personal experience: despite reaching final rounds in most interviews, their niche PM search took six months and many interviews. They found mock interviews and peer support through Stellar Peers (free) and Exponent useful.
B2B heads of product should plan for a 30-day onboarding period rather than the traditional 90 days, and are evaluated on actionable insights per session. Early impact depends on understanding and optimizing the product’s core flywheel; succeeding in the role also requires C-suite buy-in alongside CEO support.
Online attention is fragmented across thousands of app and community-specific “current things,” each with its own drama and figures; algorithms can make each community feel like the center of the universe. For product teams, this cautions against treating one platform’s discourse or perceived trend as representative of all users.
A PM said Elena Verna’s ProductCon SF talk described a non-engineer shipping to production weekly and IC work as a new career path, with ICs at Lovable earning as much as VPs; inspired by the talk, the commenter built a booking app in about two hours and started a weekly build community for PMs.
A non-technical PM at a startup building a social-media app is learning about APIs, GitHub, and SQL, including window functions and CTEs, to work more effectively with engineers. A commenter recommends asking an LLM to explain how things work and learning incrementally by starting with small pieces.
- For a freelance marketplace with freelancer supply but unresolved client acquisition, validate buyer demand rather than adding more supply: speak with people actively hiring freelancers to test whether the proposed benefits matter, and identify a real unmet need that differentiates the service.
- A suggested concierge-MVP path is to focus on a narrow category where clients already pay, sell a specific service, and handle matching and payment manually; one commenter suggests doing this for the first 20 clients. Track why buyers decline and whether they return, then use referrals from repeat clients as a possible early growth loop and productize the platform as trust develops.
In a discussion of essential product-marketing tools, one respondent called Claude connected via MCPs to product data (Snowflake, Jira) and CRM data (HubSpot, Salesforce) a must-have, emphasizing good data hygiene; they said this setup can support enablement, GTM, adoption, and competitive intelligence .
AI-assisted building can speed shipping, but distribution remains a major challenge; an existing audience may help a launch without showing that followers have the problem or will keep using the product. For product discovery, one commenter recommends focusing on a group you can reach, investigating problems people already spend time or money fixing, and learning why users stop; deep understanding of a specific group’s pain is presented as a way to build beyond a clone.
For a learning product with three daily topic suggestions and interactive lessons but only one paying user, commenters advised investigating the payer’s motivation before choosing a direction: ask what they were trying to solve, what they used before, and what they would do if the product disappeared. That can distinguish goal-driven demand from users who simply enjoy the product and help identify similar customers.
Before rebuilding around a narrow use case, test a defined outcome with the existing product: sell a one-week plan that promises a specific capability and check whether users achieve it, or offer a fixed-length plan for one goal to a small group in a community where those prospects already gather. A commenter argued that narrowing can improve targeting and messaging without necessarily shrinking the market, but recommended validating willingness to pay first.
For the audience-engaged book and movie podcast, commenters suggest treating the community as the product and starting with low-friction lists and recommendations, with affiliate links as a possible later revenue stream; ticketed events may fit if members are already interacting. Before building a list app, test a paid tier—such as early access and event recordings—and ask members what they already want; commenters caution that generic logo merch may be a weak fit unless the show has recognizable episode-specific jokes or imagery.
Sachin Rekhi describes five process shifts he is seeing in AI-focused product teams:
- Specs → prototypes: Prototypes can replace both Figma mockups and written specs, helping teams communicate with less information loss and reach product trade-offs faster.
- Ad hoc research → continuous discovery: Cheaper AI-enabled research and the research needs of AI prototypes are pushing teams toward ongoing discovery practices.
- Information mover → product builder: AI can take over coordination and information-management work, freeing PMs to spend more time building products.
- Long-range → short-range roadmaps: He says annual roadmaps are giving way to planning at most quarterly, as product velocity increases and future technology capabilities are hard to predict more than a quarter ahead.
- Acceptance criteria → evals: For non-deterministic AI products, PMs need AI evaluations suited to AI software; Rekhi calls this an increasingly necessary PM skill.
Sferical One launched as a shared AI supercomputer powered by NVIDIA, intended to give researchers and businesses more computing power to test ideas . Ek’s focus was on seeing what people build on the platform .