PM Operating System: agent build spec

A build spec for standing up a PM operating system in Claude.

This is written for the agent implementing it inside a company: what to install, connect, and construct, in order, layer by layer. Click any bar to jump to that layer.

Distilled from Aakash Gupta's Product Growth podcast, featuring Jyothi Nookula, whose setup won an internal hackathon against 30 engineering teams.

tap a layer to jump ModelsL1 SurfacesL2 Knowledge baseL3 Integrations (MCP)L4 Agents & orchestrationL5
LAYER 01

Models: match intelligence to task

Route every task through this table instead of defaulting to one model for everything. That wastes either money or depth.

ModelUse forWatch out for
Haiku High volume, low depth work: generating variants, triaging a pile of documents, quick tagging or classification. Not enough reasoning depth for anything ambiguous. Good for keeping automation token cost down.
Sonnet About 90% of daily PM work: PRDs, research synthesis, competitive analysis, stakeholder briefs, roadmap thinking. Best ratio of quality to cost. Start here by default.
Opus High stakes reasoning: tradeoff analysis, reconciling contradictory research, long horizon planning with second and third order effects. Can loop on one reasoning path. If it stalls, start a fresh chat rather than pushing through.
LAYER 02

Surfaces: route tasks to the right interface

Each surface has genuinely different capabilities. Treat them as distinct tools, not interchangeable frontends for the same model.

SurfaceReach for it when
claude.ai (web)Quick conversational question, like "what's X versus Y." No local file access.
Desktop + CoworkScheduled automations, connectors to Gmail, Calendar, Drive, Jira, Slack. Only runs while the laptop is on.
MobileChecking a running automation from a walk. Not for building.
Chrome extensionComputer use: live competitive research, or user testing your own product with a real user persona.
Claude CodeBuilding: knowledge bases, MCP servers, agents, prototypes. VS Code is the gentle starting point; terminal or Cursor once comfortable.
claude.ai/designDecks, wireframes, prototypes, assets that match the company's brand via a Design System. Research preview, separate surface.
LAYER 03

Knowledge base: the layer that compounds

This is the highest leverage build: a knowledge base plus MCP server that gives Claude persistent, compounding context on the company, its people, and its history, instead of a blank chat window every time.

chief of staff architecture
Inputsmeeting transcripts, strategy docs, org charts, PRDs, emails
→
Extractiona prompt per document type pulls goals, people, decisions, signals
→
Context KBcompany / people / topics / meetings / my context
→
MCP serverread and write access to the KB
→
Claude Desktopchat with it as a chief of staff
kb folder structure 
context_kb/
├── company/
├── people/        # one file per person
├── topics/
├── meetings/
├── documents/
└── my_context/
    ├── priorities.md
    ├── okrs.md
    ├── insights.md   # patterns over time
    └── todo.md
why an mcp server, not just files

Markdown alone is passive. An MCP server lets Claude Desktop talk to the knowledge base: "I'm meeting my manager tomorrow, what should I know?" It also writes new extractions back in, so it compounds automatically instead of requiring manual upkeep. Keep it local if the content is sensitive org data. It never has to leave the laptop.

LAYER 04

Integration fabric: MCPs and skills

Connect Claude to the company's real tools through MCP, then encode repeatable playbooks as skills so they load only when a task actually needs them, about 50 words per skill until invoked.

connect in this order
  • Email and Calendar
  • Meeting transcripts (Granola, Meet, or Zoom)
  • Slack, often the richest context source after transcripts
  • Jira or Linear
  • Drive, Notion, or Confluence
  • Analytics and observability, CRM, domain tools, as needed
skills to build first
  • Backlog triaging
  • PRD writing (the company's actual template)
  • Customer interview synthesis
  • Support ticket to Jira ticket, triggered by an event

Rule: AI drafted skill files underperform human tuned ones. Let Claude draft, then add the team's own template and edge cases.

LAYER 05

Agents and orchestration: where it compounds

Scheduled automations, generative design, and the self testing loop that closed out the hackathon win.

cowork automations
AutomationCadenceOutput
Morning brief9:00 AM daily · haikuTop 3 focuses, calendar, inbox and Jira needing attention. Under 400 words, facts only.
End of day5:00 PM dailyWhat shipped, what slipped, what's new, tomorrow at a glance.
Standup briefSprint cadenceDone, in progress, blocked, new, sprint health. Under 250 words.
Ticket to PRTriggered by an eventSupport ticket → Jira ticket → Claude Code implements → PR opened for review.
adversarial evaluator: the hackathon technique

Inspired by GANs: a generator agent produces the system prompt, an adversary attacks and scores it, and feedback flows back until it clears the bar.

generatorAgent under test
score + feedback → ↻ revise ↻ ← attacks
adversaryRed team evaluator
ITER 1
8.52
caved on a format conflict attack
ITER 2
9.00
prompt revised
ITER 3
9.08
passed threshold (>8 mean)

The architecture is now easy to build. The leverage is entirely in what the adversary is told to test for. That is the actual product judgment.

SEQUENCE

Build order

The long pole is the knowledge base. Build it once and the benefit compounds every day after.

  1. 1
    Install and pick defaults Desktop app, Claude Code in VS Code, default to Sonnet.
  2. 2
    Ship one Cowork automation morning brief first, to confirm connectors authenticate.
  3. 3
    Build the chief of staff KB and MCP server highest leverage, do it once.
  4. 4
    Write 2 to 3 skills for the most repeated tasks.
  5. 5
    Layer in remaining automations end of day, standup, ticket to PR.
  6. 6
    Move decks and prototypes to Claude Design with a real Design System.
  7. 7
    Build an adversarial evaluator for the next agent shipped.