Project OS
A set of skills that spins up a complete project environment in Notion, then operates it autonomously. Agentic project managers run sprints, write defects, and manage the full delivery lifecycle.
Claude.ai and Claude Code are two different surfaces with no shared memory. The strategy lives in a chat thread. The code lives in a repository. Neither can see the other, and every new thread starts from nothing.
That gap is the real problem, and it is not a tooling problem. It is a continuity problem. Project OS closes it by making Notion the intelligence layer that spans both.
The gap this closes
Each surface is fluent in half the project and blind to the other half. Without something between them, the human becomes the integration layer, re-explaining the project on every new thread.
Claude.ai
Strategy, requirements, personas, the reasoning behind every decision.
Cannot see the codebase, the commits, or what actually got built.
Claude Code
The repository, the build, the defects, what the code actually does.
Cannot see the thinking that produced the specification it is building.
Notion
One source of truth both surfaces read from and write back into. Context stops dying at the end of a thread.
Design principles
Four rules hold the system together. Everything in the architecture below exists to enforce one of them.
One memory, two surfaces
Notion is not a reporting layer bolted on after the fact. It is the working memory both surfaces share. The thinking done in conversation becomes the specification Claude Code builds against, and what the build reveals flows back into the same documents.
Separation of powers
Every skill has a narrow, declared write scope. Discovery never touches the Backlog. The Backlog skill never edits requirements. Claude Code never rewrites the specification it was given. Nothing in the chain can quietly overwrite the reasoning that produced it.
Approval gates, not autonomy theatre
The digest skill audits, reports, and asks. It writes inline annotations with links for anything it was not cleared to change, so deferred decisions stay visible instead of evaporating at the end of a thread.
A regenerable synthesis layer
The LLM Wiki is never hand-edited. It is rebuilt on a schedule from commits, Backlog state, and the Daily Digest, so the summary can never drift away from the system it describes.
System architecture
Three layers, one source of truth. Claude.ai runs the skill chain, Notion holds the project state, and Claude Code executes against it.
Claude.ai
Project OS skill chain
/pm-project-new
Whenever a new project needs a Notion project container and complete specification templates, backlog, and supporting documents.
Scaffolds project docs, Backlog, TRD, PRD, Persona Profiles, Feature Requirements, Project Overview, and Daily Digest.
Write any project content, Digest rows, or ticket content. This is a templatized project HQ.
/pm-project-discovery
Once a project is scaffolded and the concept is understood well enough to begin capturing design and technical requirements, via natural language conversation in voice mode.
Project intake interview. Writes requirements, market research, value proposition, and technical specifications into the project documents. This intelligence is consumed by Claude Code during build and becomes the foundation for the Backlog.
Write to the Backlog.
/pm-project-backlog
Once a project has sufficient detail across multiple core documents to warrant an actionable Backlog.
Bulk translates project documents into Epics and actionable tickets. Writes User Stories, Acceptance Criteria, How to Test, and Claude Code prompts where necessary.
Edit anything except the Backlog.
/pm-project-sprint-activate + -close
Once the human Project Director decides it is time to sprint. Not necessarily code: it could be user research, field testing, or feature exploration.
Loads a sprint backlog. Moves tickets through sprint start and stop transitions.
Touch the Daily Digest.
/pm-project-thread-digest
When a thread is nearing its natural conclusion and its emergent intelligence needs to be backfilled across all project documentation.
Acts as a hyper-aware project manager keeping documentation in parity with the decisions and events of a single thread. Generates an end-of-thread synopsis, audits thread-relevant documents for conflicts and gaps, delivers a numbered action report, executes approved changes, writes inline annotations for anything deferred, and updates the Daily Digest.
Edit or repair content in any document without approval.
Notion
Project documentation
Product Overview
Product Roadmap
Feature Requirements
Persona Profile
Project Requirements (PRD)
Technical Requirements (TRD)
Market Analysis
Backlog
Claude.ai Daily Digest
Thread synopses. Decisions, additions and changes, contradictions and conflicts, open questions.
LLM Wiki page
Regenerable synthesis layer: never hand-edited.
Single source of truth
Claude Code
Execution layer
Building Software
Writes code and defects.
Edit PRD, TRD, Backlog, or Daily Digest.
Second Brain Routine
Planned, scheduled
Reads GitHub, Backlog, and Daily Digest, then regenerates the LLM Wiki page.
Edit PRD, TRD, Backlog, or Daily Digest.
GitHub
Repository: commits and pull requests.
How intelligence moves
- /pm-project-new→NotionScaffolds the new project container and document templates.
- /pm-project-discovery→NotionWrites requirements, research, and specifications into project documents.
- /pm-project-backlog→BacklogSynthesises project documentation into actionable Epics and Stories.
- /pm-project-sprint-*→BacklogMoves tickets and updates ticket states across sprint transitions.
- /pm-project-thread-digest→NotionReads, writes, and edits documentation. Appends to the Daily Digest per run.
- Claude Code→NotionReads all documentation including the Backlog as context for the build.
- Claude Code→BacklogWrites defects, new tickets, and story updates.
- Claude Code→GitHubCommits code with comments and opens pull requests.
- Second Brain→LLM WikiReads commits and emergent intelligence, then regenerates the wiki.