"Agent-native Product Management" — Marcus Moretti (Every guide)

Why this is in the vault

A working one-person PM system from Every's Spiral lead: a Rumelt-shaped strategy.md written by agent interview, ticketless Kanban run by the agent, and a daily one-page "product pulse" saved as dated markdown that becomes the product's memory. It is a clean template for how a solo operator closes the plan-ship-review loop with agents, which is exactly Ben's situation with Ray and the factory.

The core argument

Moretti runs Spiral alone (PM, code, support, marketing), so the old PM workflow no longer fits in the day. His claim: much of the routine, cross-tool PM work is now a conversation with Claude Code - "the conversation is the work." He writes no tickets and no SQL. The guide is framed as a point-in-time snapshot of pillars he expects to hold for months, not years.

A short history sets the stakes: product management began as Procter & Gamble's 1930s "Brand Man" (ownership and accountability), then absorbed customer liaison (HP), UX, agile, and A/B testing. The modern PM is buried under 100+ SaaS subscriptions; LLMs finally let one person cover the interdisciplinary load.

The loop

Plan -> ship -> review -> repeat. PM lives in plan and review; shipping is delegated to compound engineering. Every shipped thing is an experiment; once learnings pile up, revisit strategy. Citing Kieran Klaassen, the ratio has flipped from 20% planning / 80% execution to 80/20.

strategy.md (the ce-strategy skill)

Structure borrowed from Richard Rumelt's Good Strategy Bad Strategy, produced by an agent interview that pushes back on vague answers ("Whose situation specifically? What do they try today?"):

• Target problem - a recurring, expensive pain.
• Approach - one or two sentences of guiding policy that would intrigue someone with the problem; not a goal, not a feature.
• Who it's for - ideally one persona first (Moore's Crossing the Chasm), expand later.
• Key metrics - three to five, SMART, value-proving; at minimum "track people and dollars." Spiral's example: drafts exported. No page views or vanity metrics.
• Tracks - two to four multi-month capabilities; track one is usually core performance. More than four means lack of focus.
• Optional: Not working on and Marketing/positioning.

No features or statuses live here. Rerun the interview every few months; with accumulated context the agent's questions get "sharper, the conversation tougher."

Shipping without tickets

Use the ideate, brainstorm, and plan skills against the tracks. The issue tracker must have an MCP or agent integration (he uses GitHub Issues; Linear also fine). The agent writes, moves, and updates tickets; the human only talks about them. Statuses: now/next/later plus In Progress and Done. No sprints.

The product pulse (ce-product-pulse skill)

Prerequisite: instrumentation and stored logs (he suggests PostHog's setup wizard). First run interviews you on metrics (strategy KPIs, system health, key funnels), connects sources, then drafts a first pulse for feedback.

One page, roughly 30-40 terminal lines, four sections:

1. Headlines - the first three lines should tell you what matters.
2. Usage - engagement, value-realization events, conversions, each strategy metric with delta vs. prior window, AI-session quality scores if present.
3. System performance - p50/p95/p99 latency vs. prior window, top five error signatures with one-line explanations; omitted if no tracing.
4. Followups - one to five specific things to investigate; thin when nothing is anomalous.

Sources, up to four categories: product analytics (PostHog, Mixpanel, Amplitude), tracing (Datadog, Sentry, Logfire, Honeycomb), payments (Stripe, Paddle), and a read-only database connection. The interviewer is told to reject a write-capable DB connection; a read replica is ideal. Sections are skipped when sources are missing; "quality beats quantity." MCP is preferred, CLI or API works.

Behavior: the agent reads its own draft "like a founder," annotates anomalies, runs follow-up queries (one user? third-party outage?), and on a second pass adds a section pre-answering obvious follow-ups. No hard-coded thresholds by default; it judges against common sense and prior pulses. Every run saves to ~/pulse-reports/ as dated markdown, so the folder answers "when did this trend start?" Scheduled daily at 8 a.m. via Claude Code Routines, plus manual runs.

Qualitative feedback stays human: a conspicuous support email that lands in his inbox, and a 15-minute call link in every marketing email. Canny and Featurebase have MCPs as extra inputs.

Integration and gaps

When docs/strategy.md exists, ce-ideate, ce-brainstorm, and ce-plan read it as grounding, and planning should also consult past pulses. Not included: prioritization (he has a private daily /prioritize), pulse-to-pulse diffing (raw material exists, no skill yet), and per-stack customization beyond hand edits.

Mapping against Ray Data Co

The sharpest connection is to RDCO's review-bandwidth constraint: the pulse is a designed one-page read with headlines first and a capped followup list, which is the format discipline Ray's morning-prep brief and the factory's run reports need. Output volume is not the problem; an unbounded report is.

• Factory eval plans. The pulse's "read it like a founder, then pre-answer follow-ups" second pass is a self-critique step that maps onto the factory's fresh-eyes review gates. The read-only, reject-write-access DB rule is the same posture Ben should state explicitly for any Snowflake-connected agent (read replica or read-only role).
• Ray system. RDCO already has /finance-pulse and longitudinal vault indexes; Moretti's "folder of pulses is the memory" validates that, and his missing piece (pulse diffing) is something Ray could do cheaply because notes are dated markdown.
• Strategy doc. The Rumelt structure (problem, approach, persona, 3-5 metrics, 2-4 tracks, not-working-on) is a strong template for The Denominator's own one-pager and for the factory's per-use-case charter.
• Caveat. The context is a small consumer SaaS with clean product analytics; enterprise agent deployments rarely have a tidy "value-realization event" ready to count.

Why it matters for RDCO / The Denominator

• DO: write strategy.md for The Denominator using the five required sections, with an explicit "Not working on" list (for example: no consulting offers, per the phData rule). Pick 3-5 value metrics, not subscriber vanity counts.
• DO: add a fixed four-section, 40-line cap to factory run reports (headlines, usage, failures, followups ≤5), so Ben's review time per run is bounded.
• WRITE: a Denominator piece on "the pulse as the agent's eval surface": successful enterprise agents ship with a daily one-page health read that names its value event, and the hardest part is defining that event, not building the agent.

⚠️ Sponsorship

House promo, no paid third party. The guide is a delivery vehicle for two skills in Every's own compound-engineering plugin (install commands included) and showcases Spiral, the author's product. Bias implication: the workflow is presented as tool-agnostic, but every step routes through Every's plugin; the underlying method (interview-built strategy doc, scheduled one-page metrics read) works without it.

Related

• [[2026-05-01-every-claude-code-for-product-managers]] - Every's earlier PM-with-Claude-Code piece; this guide is the productized follow-on.
• [[2026-05-03-every-context-window-codex-goes-to-work]] - same-week digest where this guide first surfaced.
• [[2026-04-09-every-four-ai-agents]] - Every's internal daily-pulse agents; the source of RDCO's /financial-pulse idea.
• [[2026-05-04-every-chatgpt-manage-my-workweek]] - agent-native PM from the OKR side.
• [[2026-06-30-every-ai-strategy-bets-explicit]] - making strategy bets explicit, the same move as strategy.md.
• [[2026-02-09-every-compound-engineering-guide]] - the plugin these skills ship in.
• [[2026-01-28-every-stop-coding-start-planning]] - the 80/20 planning flip this guide cites.
• [[2026-01-17-every-agent-native-architectures-guide]] - the build architecture underneath.