Skip to content
Aakash GuptaAakash Gupta

The Unfair Way to Build an AI Native Product Team

Together AI's product team walks through their entire product development cycle, from a 5-min customer research brief to validating that agents can actually use what they shipped. Skills, shared context, a cross-repo orchestrator, and agent evals. Full Writeup: https://www.news.aakashg.com/p/how-to-build-a-shared-context-repo Transcript: https://www.aakashg.com/how-to-build-a-shared-context-repo-any-ai-harness-can-read/ Timestamps 1:44 - Why individual productivity was the wrong goal 3:26 - Inside the Together product repository 10:41 - Where team skills end and personal skills begin 13:16 - Ads 15:17 - Running feature research live 19:56 - The PRD writer that interviews you 27:31 - What a good PRD looks like now 35:07 - Orchestrator, one view across every repo 40:24 - Shared context is a hierarchy, not a flat pool 44:19 - Agent evals, testing the product as an agent 52:26 - Where the PM ends and the developer begins 🏆 Thanks to our sponsor 1. Arize: Trace, evaluate, and fix your AI agents before your users find the failures - https://bundle.aakashg.com/ Key Takeaways: 1. Individual productivity can move a company backwards - The team's starting question was not how to make each person faster. It was whether everyone generating unlimited code and content actually added up to progress. Charles calls the failure mode flooding your coworkers' context windows, where everyone launches slop at each other. 2. The shared repo holds context and skills - Markdown and YAML files covering customer intelligence, sandboxes, and the output of strategy meetings broken down by mission and milestone. Anything tied to a specific codebase stays out. The payoff is that a PM can read another team's context and draft a real proposal before taking up that PM's time. 3. Skills live closest to the work they touch - If a skill references code inside one team's repo, it stays colocated there. Everything else goes to a personal or shared repo. Test it on a branch, use it a few times, and only push to main once it proves repeatable. Niche ones never get pushed. 4. Shared context is a hierarchy, not a flat pool - The team abandoned the idea that everyone should carry everyone's context. Most people have no motivation to learn the depth of someone else's area. They want the one answer they came for. Some people live at the bottom of the hierarchy; most just traverse the top. 5. The PRD stopped being a gate - Historically, it was the document everyone aligned on before building started. Together treats it as a trigger for ideation and problem-solving instead. One to two pages, defining the customer problem, a few solution options, and a sample user journey. 6. A prototype replaces the bulk of the long document - Pavneet came from Amazon and used to write 20-page PRDs. A separate skill takes the one-pager and produces a prompt for a design tool. That visual is where the sharpest feedback shows up, from engineering and marketing alike. 7. Discovery collapsed from half a day to five minutes - The research agent pulls from the support platform, the project tracker, and internal docs at once. It surfaced 19 tickets filed in two months, flagged that the feature had been partially built and abandoned, and gave verbatim with sources. The value is not the summary. It is not duplicating work someone already started. 8. Automate execution, keep decisions human - Defining the feature, the API surface area, and the abstraction layer stay hands-on. Code writing is the part that runs on its own. The PRD skill is explicitly instructed to challenge the PM's assumptions rather than accept them. 9. Agents are already the majority user, so validate for them - Agent evals spins up a sandbox, gives an agent a real task against the product, and watches. It caught that agents couldn't find the fine-tunable models page because it wasn't linked from the quick start. Dozens of doc fixes came out of this. Charles calls agent success the new bar for UX. 10. They refused to oversell the gains - No story points, so no proof, but velocity is up more than 5%. Charles finds 3x claims suspicious because discovery, debate, and coordination don't get magically better with AI. Costs stayed manageable partly through open-weight models, partly because optimizing for collective output never produced the runaway token budgets others report. 👨‍💻 Where to find our guests Charles Zedlewski: https://www.linkedin.com/in/charleszedlewski/ Necoline Hubner: https://www.linkedin.com/in/necoline/ Pavneet Singh: https://www.linkedin.com/in/pavneetsingh/ Hassan: https://www.linkedin.com/in/nutlope/ 👨‍💻 Where to find Aakash Twitter: https://x.com/aakashgupta LinkedIn: https://www.linkedin.com/in/aagupta/ Newsletter: https://www.news.aakashg.com/ #AIProductManagement #ProductManagement #AgentEvals 🧠 About Product Growth: The world's largest podcast focused solely on product + growth, with over 200K+ listeners. 🔔 Subscribe and turn on notifications.

Charles ZedlewskiguestNecoline HubnerguestPavneet SinghguestAakash GuptahostHassanguest
Sep 8, 202659mWatch on YouTube ↗

At a glance

WHAT IT’S REALLY ABOUT

Together AI’s playbook for building a collectively productive AI-native team

  1. Together AI redesigned product work around collective productivity, aiming to prevent AI from flooding coworkers with low-signal documents and code.
  2. They maintain a central product repository of context and reusable agent “skills,” stored as Markdown/YAML and updated via standard pull requests with clear domain ownership.
  3. Live demos show AI-assisted feature discovery that pulls evidence from support tickets, customer calls, and issue trackers to rapidly validate demand and avoid duplicated efforts.
  4. They redefine the modern PRD as a short ideation trigger paired with a shareable prototype, rather than a long gating document used to start engineering work.
  5. They built internal tools—Orchestrator for cross-repo sandboxed querying/PR creation and Agent Evals for automated agent-based product tests—to keep quality high as agent usage becomes the dominant UX.

IDEAS WORTH REMEMBERING

5 ideas

Optimize for team throughput, not individual AI-generated output.

Together AI found that boosting each person’s output with AI can actually slow the company down by spamming teammates with low-signal drafts. Their north star became “collective productivity”: shared context, reusable skills, and workflows that reduce duplicated work and improve decision quality.

A lightweight, version-controlled product repo can be the backbone of an AI-native team.

They created a central “Together product repository” (mostly Markdown/YAML) containing product strategy, team context, and reusable “skills” that agents can run (e.g., sprint updates, competitor news, feature research). This repo is designed to be compatible with multiple harnesses and models (Claude Code, OpenCode, Hermes, etc.).

Treat context like code: explicit ownership + PR-based maintenance prevents rot.

Context is maintained by the owners closest to the domain (e.g., PMs own their area docs), with updates flowing via normal pull requests. Leadership direction becomes a clear trigger event to refresh strategy artifacts, and gaps are fixed when someone tries to use the context and finds it missing.

Create a promotion path from personal experiments to shared, standardized skills.

They distinguish between team-level repeatable skills and personal experiments: prototype a workflow in a branch, promote to main if it’s repeatable and broadly useful, and keep code-adjacent skills co-located with the relevant engineering repo. This avoids dumping everything into one shared “misc” bucket.

Wire customer evidence directly into the PM workflow to compress discovery time.

Pavneet demonstrates a “feature research” workflow that pulls from Pylon (support), Linear (execution), Notion (docs), plus a Customer Insights system that summarizes Gong/Slack/Pylon. The result: a fast “smell test” with counts, verbatims, and links to source evidence—cutting a half-day task to minutes and preventing duplicated work when something was partially built before.

WORDS WORTH SAVING

5 quotes

It felt like after a while that the new party foul was flooding your coworkers' context windows. Where we all just start launching slop at one another.

— Charles Zedlewski

So what we set out to do as a team is ask ourselves, what would it mean to use AI, but not to use AI to just make ourselves individually productive, where we all just start launching slop at one another, um, but rather AI to make us collectively more productive.

— Charles Zedlewski

If I had to do this manually, this would have easily occupied half of my day, if not more.

— Pavneet Singh

What replaces the bulk of that is actually a prototype.

— Pavneet Singh

So we sort of, we sort of set aside the idea that there was one big, broad, flat set of context that we're all gonna swim in, and it's much more like a context hierarchy, and some of us belong all the way down to the bottom of the depths of that hierarchy, and some just wanna traverse the top.

— Charles Zedlewski

Collective vs individual productivityShared product repository (Markdown/YAML)Context ownership and updates via PRsReusable skills and harness/model choiceCustomer evidence aggregation (Gong/Slack/Pylon/Linear)PRDs as short ideation + prototypesOrchestrator cross-repo sandboxing and model routing

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.

iOSAndroidClaudeChrome