DM
Claude Code replaced 80% of our roadmap planning. We're a leaner team now. Wild times to be building.
👍🎉💡
1,247·312 comments
It's not the tool. It's the context you're feeding it.
OS-Intelligence is how you fix the context.
Last time, Maya pushed back on the Q3 roadmap. You said you'd come back with the stakeholder map. She's waiting.
Your Claude Code surfaces it before she sits down - what she said, what's stuck, what she's waiting on you for.
Six weeks since you touched it. And you can't remember the details. Three decisions happened in between.
Your Claude Code rebuilds the picture in a paragraph - the position, the in-flight work, the reasoning behind every call since.
Three meetings this week. Two stakeholders flipped positions. One open question still nobody's answered.
Your Claude Code reads across every meeting, chat, and decision - surfaces the contradictions, the shifts, the unresolved.
OS-Intelligence is a reasoning layer for AI-native PMs. It captures your meetings, PRDs, stakeholder notes, and chats, synthesises them into reference state, and surfaces the right context at the start of every session. Your stated goal becomes the lens. Every claim traces back to source.
/os-welcomeMeetings, PRDs, decisions, chats, open questions - captured from everyone who contributed, surfaced when you need them, ready for your AI to reason against.
Four of six pilot users were PMs mid-build of something similar themselves. If that's you, this is the working version of what you were going to build.
Install today. First synthesis on real PM work in one sitting.
/os-welcomeWe're building this in the open, together.
Get started
A real current-state.md
- the project's live context, ready when you start.
# Atlas - current state ## Project Position Q3 roadmap is contested. Tom (eng) wants to defer "agentic flows" until v2; Maya (design) wants them in v1 to test the core hypothesis. Decision pending. ## Stakeholder Dynamics - Maya Chen (design lead) - pushed back Apr 14. Wants stakeholder map before next sync. - Tom Reilly (eng lead) - flipped Apr 18. Now open to v1 if scoped to one surface. ## Standing Decisions - Ship v1 to internal users only (Apr 09) - Defer mobile until v1 lands (Apr 11) ## In Flight - Stakeholder map for Maya - owed - Cost estimate for agentic v1 - w/ Tom ## Open Questions - Are we measuring activation or retention? - Who owns the v1 / v2 cut decision?
Claude reasoning against it
- every answer traced to source.
› What is Tom's objection? Tom's objection isn't to agentic flows in principle - it's to scope and timing. As of Apr 18 he flipped: he'll accept v1 if it's bounded to one surface. The unresolved piece is cost. He's working on an estimate for the scoped version; until that lands, he can't commit. There's also a second-order objection: he wants the v1 / v2 cut decision owned by someone - currently it isn't, and that's blocking Maya's stakeholder map. If you want him on board this week, the move is to surface the ownership question before the cost estimate, not after.
First cohort installed May 2026. More real-user quotes landing this week.
For the dev-fluent reader who scrolls. The boring, important parts.
Synthesis splits across specialised sub-agents and reconverges, so a 2-hour transcript doesn't block a 30-second answer.
Routine reads go to Sonnet; structural reasoning escalates to Opus. You don't pay Opus prices for a name lookup.
Every decision is a new entry, never a rewrite. Your past selves stay legible - diffs over time, not overwrites.
When the schema evolves, your existing trail migrates forward - no re-ingest, no losing what you already wrote.
The log is a file in your repo. Nothing leaves your machine unless you push it.
Upgrades are git pulls. Breaking changes ship migration scripts. You stay in control of when, and whether, to move.
No. The reasoning trail is a folder in your repo. Synthesis runs locally through your own Claude Code install - same model calls you're already making, no new endpoints, no new accounts.
If you push the repo to GitHub, that's on you and where you push it. OS-Intelligence itself adds zero telemetry.
Claude Code is the only environment where a long-horizon reasoning layer can sit next to your work, not in a separate tool. Skills, sub-agents, model routing - the primitives needed already exist. OS-Intelligence is the connective layer.
If those primitives change, OS-Intelligence changes with them. That's a feature, not a risk.
Not yet. Claude Code is the substrate OS-Intelligence runs inside - install it first, then come back. /os-welcome takes you from clone to first synthesis in about 20 minutes.
It's built for PMs already living in Claude Code - that's the substrate, that's where the reasoning happens. If you don't use Claude Code yet, install it first (10 minutes from claude.com/code), then clone OS-Intelligence and run /os-welcome. The guided walkthrough takes about 20 minutes and lands you with your first synthesis on real work.
Not a memory layer. Not a retrieval tool. OS-Intelligence is a reasoning layer with structure, traceability, and governance built in.
Granola captures meetings. Notion AI summarises documents. Team OS structures shared team context. OS-Intelligence is the layer underneath: it synthesises your raw material (meetings, PRDs, stakeholder notes, chats) into a reasoning trail your AI reasons against, not a wiki you click through.
Different shape, different layer - they stack rather than replace. If you already use one of those, OS-Intelligence is what makes your AI sessions reason against that captured context instead of starting cold every time.
Scroll up to the proof section - that's a real current-state.md file (project name and people anonymised), and the second frame is what your Claude Code session looks like reasoning against it.
The full Acme Corp example also ships in the public repo at projects/acme-corp-example/ - clone, browse it, see the structure on synthetic data before you bring your own.