Claude Code Plugin

PM Operating
System

A Claude Code configuration that gives your AI assistant a structured productivity layer. Goals, tasks, projects, pipeline evaluation, and compounding knowledge — all in plain markdown. Runs in Claude Code, Codex CLI, Cursor, and Antigravity via the open Agent Skills standard.

Get Started View on GitHub
claude
0
Skills
0
Commands
0
Agents
0
MCP Tools
Philosophy
The LLM as an Operating System

Built on Andrej Karpathy's insight that large language models are best understood not as chatbots, but as the kernel of a new kind of operating system.

"Think about it more like an operating system."

— Andrej Karpathy, Intro to Large Language Models (2023)

🧠
Strategic Memory
GOALS.md is read every session to prioritize your work against your actual objectives.
⚡
Specialized Capabilities
32 skills the LLM can invoke — validation, risk analysis, build-ready technical specs (with task breakdowns), sprint planning, slide generation, and more.
🔄
Recurring Workflows
7 commands for daily standups, weekly reviews, quarterly OKR scoring, and repo analysis.
🤖
Autonomous Agents
3 agents that run in the background — deep research, batch evaluation, system diagnostics.
🔧
Structured Tool Use
MCP server with 11 tools for task management, project queries, and deduplication.
📡
Stays Current
Model/effort tiering, two currency watchers, and least-privilege scheduled reports keep the framework at the current Anthropic state of the art — catch-up becomes reading a report, not a research session.
📚
Long-term Memory
knowledge/ compounds from daily journals to session reviews to quarterly assessments.
Architecture
How Everything Connects

Three architectural views — system flow, project pipeline, and the compounding loop that makes each session smarter than the last.

System Overview

From raw ideas to daily execution — everything flows through a single inbox, gets classified, and ties back to your goals.

Capture
Brain dumps Slack threads Meeting notes Ideas on the go
↓
BACKLOG.md
Single inbox — append anything, process later
↓
Classify
/process-backlog
Dedup against existing items, classify by scope
↓
Execute
tasks/
Single-outcome work items
/morning picks Top 5 daily
Priority P0–P3, status tracking
Evaluate
projects/
Multi-step initiatives
/launch runs 7-stage pipeline
→ /user-stories promotes to tasks/
↓
Align & Manage
GOALS.md
knowledge/
MCP Server
Every task and project ties to a goal. Knowledge compounds across sessions. MCP Server manages state.

Project Pipeline

Each project passes through seven stages — five evaluation stages with a Go/No-Go gate, a non-blocking technical-spec stage, and a final decomposition stage that promotes to active. Skip ahead with --from.

STAGE 1
Validate
Market research brief
Go / No-Go
STAGE 2
Lean Canvas
Business model
Go / No-Go
STAGE 3
GTM Plan
Go-to-market strategy
Go / No-Go
STAGE 4
Competitive
Competitor landscape
Go / No-Go
STAGE 5
Pre-Mortem
Risk analysis
Go / No-Go
STAGE 6
Spec
Build-ready design (29 sections)
Build contract
STAGE 7
User Stories
Buildable stories
Ready to Build

The Compounding Loop

Three nested feedback loops. Each layer feeds the next — the system gets smarter over time.

Daily

/morning saves plans to journals. Next morning reads yesterday's actuals. Memories persist across sessions.

→

Weekly

/weekly compiles shipping summaries, reads journals for plan-vs-actual patterns, proposes workflow improvements.

→

Quarterly

/quarterly scores OKRs, archives stale projects, refreshes GOALS.md, cleans stale memories, audits AGENTS.md.

Features
Everything You Get

A complete productivity system — from ideation to execution to compounding knowledge.

🎯
32 Skills
Ideation, validation, build-ready technical specs, planning, execution, content output, and analysis — all invocable by your AI assistant.
⌘
7 Commands
Daily standups, weekly reviews, quarterly scoring, backlog processing, project launching, and repo analysis.
🤖
3 Agents
Deep research, batch evaluation, and system diagnostics — each runs autonomously in the background.
🔧
11 MCP Tools
Fuzzy deduplication, task queries, project pipeline status, system dashboard, priority alerts, and currency status.
📋
Prioritization
Goal-driven P0-P3 levels tied to your strategic objectives. ICE/RICE framework scoring.
🚀
Pipeline
7-stage project pipeline — 5 evaluation stages with Go/No-Go, a non-blocking spec stage, and story decomposition.
📖
Knowledge
Compounding loops from daily journals to session reviews to quarterly OKR assessments.
🔌
Integrations
Optional: Granola meetings, Slack messaging, Perplexity research, Google Workspace.
Cross-Tool
Runs in Any AI Coding Tool

Authored for Claude Code, generated to the open Agent Skills standard — so the same skills, agents, and MCP tools work in OpenAI Codex CLI, Cursor, Google Antigravity, and other conformant hosts.

uv run core/scripts/install_for.py --tool cursor --apply
🧩
Agent Skills standard
All 32 skills, 3 agents, and the analyze command generate to a committed .agents/skills/ tree that Codex, Cursor, Antigravity, Windsurf, and Copilot read natively.
📜
Shared AGENTS.md
One tool-neutral operating manual every host reads. CLAUDE.md stays the thin Claude-specific overlay on top.
🤖
Native subagents
The 3 agents also emit as real Cursor (.cursor/agents/) and Codex (.codex/agents/) subagents — separate context, background execution.
🔌
One-command wiring
install_for.py splices the manager-ai MCP server into each tool's config — dry-run by default, backs up first, never clobbers other servers.

Adapters are generated by build_adapters.py and kept in sync by the validator (check 38). Full setup: docs/portability.md.

Quick Start
Install in Minutes

Clone the repo and run setup. Two paths — guided or manual.

git clone https://github.com/Ninety2UA/pm-operating-system.git cd pm-operating-system ./setup.sh

setup.sh installs MCP server dependencies, creates workspace directories, and walks you through an interactive goals setup. Then run /refresh-goals in Claude Code to populate your goals.

# Clone git clone https://github.com/Ninety2UA/pm-operating-system.git cd pm-operating-system # Install MCP server dependencies cd core/mcp && uv sync && cd ../.. # Create workspace directories mkdir -p tasks projects knowledge/{research/projects,research/topics,meetings,journals,session-reviews,decisions,people,reference,updates,decks,voice-samples} # Set your goals (interactive bash flow) ./setup.sh
ToolVersionRequiredInstall
Python3.11+Yesbrew install python@3.13
uvlatestYescurl -LsSf https://astral.sh/uv/install.sh | sh
gitanyYesbrew install git
Claude CodelatestYesclaude.ai/download
Skills
32 Skills

Each skill teaches your AI assistant a focused capability — from market validation to sprint planning. 26 specialized skills below, plus 6 daily-workflow skills that appear as commands in the next section (the 7th command, /analyze, is standalone).

Ideation & Discovery
/discover-ideasSearch the web for project ideas and trending opportunities
/research-topicDeep research on any topic with web and social media signals
Project Evaluation Pipeline
/validate-projectResearch and validate a project idea against market reality
/lean-canvasCreate a Lean Canvas business model for a project idea
/gtm-planGo-to-market plan with ICP, beachhead segment, and channels
/competitive-analysisMap competitor landscape with strengths, gaps, and positioning
/pre-mortemRisk analysis that imagines the project has failed and works backward
Execution & Planning
/prdGenerate a Product Requirements Document for a project
/specSynthesize PRD + artifacts into a 29-section build-ready technical spec (pinned tech stack, C4 architecture, test-first task WBS, AI build-toolkit + design-tooling) — references the PRD, never restates it
/user-storiesDecompose a PRD into structured user stories with acceptance criteria
/sprint-planCreate a weekly sprint plan from current tasks and user stories
/plan-okrsCreate or refresh measurable OKRs aligned to your goals
/outcome-roadmapGenerate an outcome-focused roadmap from active projects
/prioritizeRank projects or tasks using ICE/RICE frameworks
/spin-upScaffold a project's CLAUDE.md with artifact links and recommended next skills
Analysis & Review
/ab-testAnalyze A/B test results with statistical rigor
/decisionDocument a decision with structured context, options, and rationale
/session-reviewCapture session learnings, prompts, and patterns for weekly analysis
/refresh-goalsReview and fill gaps in GOALS.md through conversation
Content & Output
/make-slidesBuild 1920x1080 HTML/CSS slide decks with a Playwright render loop (optional push to Google Slides)
/weekly-updateDraft an outbound stakeholder weekly memo — status, progress, blockers, asks
Integrations
/meeting-syncSync Granola meetings to local knowledge folder
/meeting-prepPrepare context for an upcoming meeting from People, transcripts, and tasks
/log-meetingCapture a meeting not synced by Granola — 1-on-1, interview, one-off, or standup
Currency (framework self-maintenance)
/cli-watchPlatform watcher — classify Claude Code / Anthropic changes since the baseline into a dated decision report (report-only schedulable; full owner-run)
/repo-watchEcosystem watcher — per-repo adopt/adapt/skip candidate lines over a registry of external agent frameworks
Commands
7 Slash Commands

Workflow automation at your fingertips — from daily standups to quarterly reviews.

/morning
Daily standup with meeting sync, top tasks, pipeline project, OKR check, and journal save.
/weekly
Weekly review with plan-vs-actual analysis, session patterns, and learning extraction.
/quarterly
Quarterly review: OKR scoring, project purge, goals refresh, system audit.
/process-backlog
Process BACKLOG.md with duplicate detection against existing tasks and projects.
/launch
Full evaluation pipeline with Go/No-Go gates at each stage. Skip with --from.
/write
Generate content — blog posts, emails, social media — in your authentic voice.
/analyze
Deep compatibility analysis of an external repo/resource against our system.
Agents
3 Autonomous Agents

Each runs in its own context window, autonomously performing focused tasks in the background.

Research
deep-research
Multi-source background research using Perplexity. Saves structured briefs to knowledge/research/.
When to use: Thorough investigation of a topic, market, or technology before making decisions.
Evaluation
batch-evaluator
Parallel project evaluation with comparative ranking across 5 criteria and weighted scoring.
When to use: Multiple project ideas to evaluate at once — get a ranked comparison.
Diagnostics
system-health
Diagnostic scan of tasks, projects, goals, and backlog for issues, staleness, and misalignment.
When to use: Things feel off and you want a system-wide health check across all dimensions.
MCP Server
11 Structured Tools

The manager-ai MCP server provides programmatic access to tasks, projects, and system state — with fuzzy deduplication built in.

ToolDescription
list_tasksQuery tasks with filters (priority, status, category)
get_task_summaryPriority/category/status counts with time estimates
check_priority_limitsAlerts if P0 > 3 or P1 > 7
prune_completed_tasksArchive done tasks older than 30 days to tasks/archive/
list_projectsQuery projects with filters (status, priority, category)
get_pipeline_statusCount of projects at each pipeline stage
get_project_artifactsCheck which evaluation artifacts exist for a project
get_project_summaryAggregate project stats and artifact coverage
get_system_statusFull dashboard — tasks + projects + backlog + time insights
get_watcher_statusCurrency watcher status — last report, days since, undecided candidates
process_backlog_with_dedupDeduplicate backlog items against existing work
Usage
How It Works in Practice

The workflow is simple — brain dump, process, evaluate, execute, compound.

  1. Syncs new meetings from Granola (if configured)
  2. Shows your top 5 tasks sorted by priority
  3. Flags blocked tasks and suggests unblocking actions
  4. Recommends one project to advance through the pipeline
  5. Notes which OKRs today's work advances
  6. Saves the daily plan to knowledge/journals/
$ claude > /morning Syncing 2 meetings from Granola... Top tasks for today: 1. [P0] Finalize client proposal 2. [P1] Review campaign metrics 3. [P1] Update sprint board ✓ Journal saved to journals/2026/04/06.md
  1. Reads every item in BACKLOG.md
  2. Checks for duplicates against existing tasks and projects
  3. Classifies each item — task vs. project
  4. Asks for clarification on anything ambiguous
  5. Creates files and clears BACKLOG.md
> /process-backlog Found 7 items in BACKLOG.md Checking duplicates... 2 matched existing Created 3 tasks, 2 projects ? "Build newsletter tool" — need more context. What's the scope? ✓ Backlog cleared
  1. Stage 1: Market validation with web research
  2. Stage 2: Lean Canvas business model
  3. Stage 3: Go-to-market strategy
  4. Stage 4: Competitive landscape analysis
  5. Stage 5: Pre-mortem risk analysis
  6. Stage 6: Build-ready technical spec (non-blocking)
  7. Stage 7: Decompose into buildable user stories
> /launch my-project Running validation research... ✓ Stage 1 passed — market exists Building lean canvas... ✓ Stage 2 passed — viable model ... ✓ All 7 stages complete. Ready to build.
  1. Compiles a shipping summary from completed tasks
  2. Reads journals for plan-vs-actual patterns
  3. Reviews session learnings for recurring prompts
  4. Proposes workflow improvements
  5. Suggests new commands or skills based on patterns
> /weekly Shipped: 12 tasks, 2 projects advanced Pattern: You consistently underestimate research tasks Suggestion: Add 30% buffer to research estimates ✓ Weekly summary saved
Integrations
Optional Power-Ups

Connect external services to supercharge your workflows. All are optional — the core system works standalone.

🔍
Perplexity
Powers research for validation, competitive analysis, and topic research. Required for /validate-project, /research-topic, and /discover-ideas. Uses the official Perplexity MCP server (@perplexity-ai/mcp-server).
1. Get an API key from the Perplexity API Portal
2. Install using the official Claude Code install command:
claude mcp add perplexity --env PERPLEXITY_API_KEY="your_key_here" -- npx -y @perplexity-ai/mcp-server
Add -s user before --env to register at user scope so every Claude Code project picks it up.
💬
Slack
Post standups to channels, read message history, search conversations, and draft announcements. Used by /morning and /weekly.
Install via Claude Code plugin:
/plugin install slack
Includes MCP tools + skills like /slack:summarize-channel, /slack:standup. OAuth prompt on first use.
📝
Granola
Sync meeting notes and transcripts into knowledge/meetings/ automatically via /meeting-sync. Requires paid plan.
1. Install the Granola desktop app (paid plan required)
Already wired up in .mcp.json. Authenticates via OAuth on first use. Setup guide
📧
Google Workspace
Gmail and Calendar for enriching /meeting-prep with email history and checking your schedule.
Install the official gws CLI:
brew install googleworkspace-cli
CLI tool (not MCP). Run gws auth setup once to configure Google Cloud OAuth (requires gcloud). See README for full auth steps.
Self-Audit
The Framework Validates Itself

Run one command after any change to skills, agents, hooks, or MCP tools — 46 deterministic checks catch drift before it ships.

uv run core/scripts/validate.py

Inline # /// script metadata auto-installs pyyaml. No venv setup.

📋
Frontmatter
Every skill, agent, command, project, task, and memory file checked for required fields and well-formed YAML.
🔗
Cross-references
Broken markdown links, orphan reference files, skill-to-skill invocations that point nowhere.
🔀
Registry parity
Bidirectional: every skill on disk appears in CLAUDE.md, and every entry in CLAUDE.md exists on disk.
🔧
MCP integrity
Servers reachable, declared tools match server.py registrations, each tool has a wired dispatch branch.
🏗️
Workspace shape
AGENTS.md tree matches reality; init-workspace.sh scaffolds the documented paths.
🛣️
Pipeline conformance
Projects at evaluating/ready/active flagged when expected artifacts (validation, pre-mortem, PRD, stories) are missing.
📦
External deps
Skills that call npm, gh, or gws warn if the CLI is not on PATH.
🧹
Hygiene
Tracked .DS_Store, stray TODO/FIXME markers, stale lock files, hardcoded user paths in the validator itself.
🤖
Model currency
No file references a retired model ID; every skill, agent, and command carries a deliberate model: assignment.
🪂
Degradation coverage
Every Claude-native capability ships a fallback rule, so the portable .agents/skills/ tree stays free of Claude-only tokens.
🔐
Secret hygiene
Blocking scan of tracked, portfolio-public artifacts for credential-shaped strings.
🛡️
Automation guard
The report-only guard hook stays wired and validated — local wiring is checked whenever present.

Exit 0 clean, 1 on findings. Warnings (non-blocking) are reported separately. Run it before every PR.

Staying Current
The Framework Keeps Itself Current

Two currency watchers track the Anthropic platform and the wider agent ecosystem, write dated decision reports, and leave every adoption decision to you — so the framework never drifts a generation behind between manual catch-ups.

Staying Current architecture — Anthropic releases and external agent frameworks flow through /cli-watch and /repo-watch into a dated decision report, an owner adopt/adapt/skip decision, and a transactional baseline advance, with currency visibility surfaced in /morning and /weekly

Watchers fetch only the delta since the last baseline. Adopting anything is a separate, owner-gated step — the baseline advances only after adopted work is committed.

🎚️
Model & effort tiers
Every skill, agent, and command carries a deliberate model: assignment — mechanical work runs cheap, judgment work pins higher. Adapters map tiers per host.
📡
Two watchers
/cli-watch tracks Claude Code / Anthropic releases; /repo-watch tracks a registry of external agent frameworks. Both write dated decision reports.
👁️
Currency visibility
The get_watcher_status MCP tool plus a step in /morning and /weekly surface days since the last run and undecided candidates.
📐
Capability baseline
docs/capabilities.md records the verified Claude Code capability surface — models, effort, workflows, scheduling, hooks — as the watcher's first baseline.
Least-Privilege by Construction

Scheduled watcher runs are report-only behind five independent layers — restricted tool profile, write fence, egress pin, fail-closed guard, and defang on ingestion — adversarially reviewed through nine rounds of bypass attempts.

Defense-in-depth automation guard — nested layers: restricted tool profile, write fence to knowledge/currency/, egress pin to allowlisted hosts, fail-closed PreToolUse guard, and the report-only watcher run at the center; defang, validator enforcement, and adversarial hardening annotated alongside

The fail-closed guard denies on any parse doubt, fetched web content is defanged before it can reach a report, and validator checks 39–46 keep the whole path enforced.

FAQ
Frequently Asked Questions
Clone the repo and run ./setup.sh. It installs MCP server dependencies, creates the workspace (tasks/, projects/, knowledge/) plus a blank BACKLOG.md and a GOALS.md template, then walks you through an interactive goals setup. Then run /refresh-goals in Claude Code for a guided walkthrough that fills in your goals, and /morning to start your first standup.
No. The system creates new directories (tasks/, projects/, knowledge/) and markdown files. It never touches your existing code or documents. All personal data is gitignored by default.
No. All integrations are optional. The core system works entirely with local markdown files and the MCP server. Granola, Slack, Perplexity, and Google Workspace add extra capabilities but aren't required.
It's built for Claude Code, but the underlying system is just markdown files and an MCP server. Any AI assistant that can read markdown and call MCP tools can use it — including ChatGPT, Cursor, and other MCP-compatible tools.
Every session, the assistant reads GOALS.md (your strategic objectives) and AGENTS.md (the instruction set). Tasks and projects reference goals in their context sections. The AI cross-references to suggest what advances your objectives most.
Yes. Skills live in .claude/skills/<name>/SKILL.md and agents in .claude/agents/<name>.md. Each uses YAML frontmatter for configuration. Everything that you invoke as a slash command is also a skill — there's one consistent pattern. You can modify existing ones or create new ones by following the same shape.
The system learns through three loops: daily (journals), weekly (pattern detection + shipping summaries), and quarterly (OKR scoring + system audit). Each layer feeds the next — so your assistant gets smarter about your work patterns, estimates, and priorities over time.
Everything is plain markdown with YAML frontmatter, stored locally on your machine. No cloud sync, no databases, no external storage. Personal directories (tasks/, knowledge/, etc.) are gitignored by default. You control exactly what gets committed.
Run uv run core/scripts/validate.py. The validator runs 46 deterministic checks covering frontmatter, cross-references, registry parity, MCP tool wiring, workspace shape, pipeline conformance, external CLI deps, and hygiene. Exit 0 means clean. Warnings (non-blocking) are reported separately. Run it before every PR, or any time something feels off.
Scheduled watcher runs are report-only by construction: a restricted tool profile denies Bash and mutating MCP tools, writes are fenced to knowledge/currency/, fetches are pinned to allowlisted hosts, and a fail-closed PreToolUse guard denies anything it can't parse cleanly. Fetched web content is defanged before it lands in a report, and the whole path went through nine rounds of adversarial review. Nothing runs unattended unless you explicitly choose it during setup.