Duet AI for Founders: Founder Mode as an Operating System

Founder Mode used to be a personality trait. With Duet AI, it's an operating system.

Paul Graham coined the term in September 2024 after Brian Chesky told a room of YC founders that the standard advice, "hire good people and give them room to do their job," nearly killed Airbnb. The fix was not hiring better managers. It was running Airbnb the way Steve Jobs ran Apple: skip-level access, relentless detail obsession, zero layers between the founder and the product.

The thing nobody said at the time: Founder Mode had a physical ceiling. One human cannot sit in 40 code reviews, 12 customer calls, and 8 design crits a week. At some point every founder hires a VP, then a director, then a manager for the managers. The bloat layer begins and velocity dies.

Agents break the ceiling.

This guide shows you how to run Founder Mode as a system, not a vibe. You'll learn the skill stack, the memory repos, the daily cadence, and the exact prompts to paste into Duet this week.

What is Founder Mode?

A few things define Founder Mode in practice:

• Skip-level everything. Founders talk to ICs directly. No pre-meetings, no filtered updates, no org-chart theatrics.
• Small, direct teams. Steve Jobs ran annual retreats with the 100 most important people at Apple, regardless of where they sat on the org chart. Impact was the filter. Seniority was not.
• Detail obsession. Founders stay in the code reviews, the copy edits, the customer calls, the pricing math.
• No modular delegation. Manager Mode treats every team as a sealed subtree and hopes for the best. Founder Mode opens the boxes whenever the founder wants.

Founder Mode is not micromanagement. It is the refusal to pretend that distance from the product makes you a better leader.

Why Manager Mode fails

Every founder has lived this story.

You start with a team of three. You ship features on Tuesday and close paying customers on Friday. Growth is good, so you raise a Series A. You hire a VP of Product because someone told you that's what grown-up companies do. Six months later, shipping a feature takes three months, and you can't remember the last time you talked to a real user.

That is Manager Mode.

The people architecting the product are not the people coding it. Product managers schedule pre-meetings for the sync meetings. The company optimizes for internal politics instead of customer velocity. Vision drifts because no one in the execution chain has the founder's full context.

Garry Tan, CEO of Y Combinator, has been one of the loudest voices against this default. His argument: hiring a layer of professional managers almost always introduces bloat, slows shipping, and puts the founder on the wrong side of the product. The best founders do not graduate out of the trenches. They buy better shovels.

Founder Mode vs Manager Mode

Why agents break the Founder Mode ceiling

The reason Founder Mode had a ceiling was physics. One founder cannot sit in 40 code reviews, 12 customer calls, and 8 design crits a week without burning out. So at some point, most founders hire a COO, a VP, a director, and the bloat layer begins.

AI agents kill the physics problem.

An agent can sit in the code review. An agent can draft the design spec. An agent can write the cold email sequence, read the customer interview transcript, and flag the churn pattern. Each one runs in parallel. Each one takes your full context as input. None of them ask for a performance review or a Notion page about "team ceremonies."

The goal is no longer to scale headcount. It is to scale output.

This is what Duet enables: the founder stays the primary orchestrator, and a fleet of specialized agents handles execution. Velocity stops being a function of how many humans you hired last quarter and starts being a function of how sharp your skills, memory, and briefs are.

The Founder Mode stack in Duet

A chat window is not Founder Mode. Asking vague questions in a text box is the AI equivalent of hiring a junior generalist. You re-explain context, re-review bad output, and slowly drift back toward Manager Mode because the agent cannot finish anything.

Three layers change the equation.

1. Memory repos: institutional knowledge that never quits

A normal hire takes six weeks to get up to speed. They read docs, shadow calls, fumble through the codebase, and ask the same questions twice. A memory repo removes that step entirely.

Your agents load your full context on every session:

• Architectural decisions and the reasons behind them
• Brand voice, tone guidelines, banned words
• Past customer complaints and how you resolved them
• The pricing experiment you killed in March and why
• The positioning narrative you approved last quarter

Memory is not documentation. Documentation is for humans who might read it. Memory is the substrate your agents run on every single session.

2. Skills: specialist ICs on the bench

Skills are pre-trained individual contributors. A skill is not a chatbot. It is a specialist with opinions, frameworks, and taste baked in. Drop one into a Duet session and it behaves like a senior IC who already read your brand book.

You are not teaching the agent how to design, write, or sell. You are pointing it at the problem.

For the full mental model, see our deep-dive on Agent Skills 101: Tools vs MCP vs Skills.

3. Orchestration: you, staying in the review loop

You are the taste filter. Not the task assigner. Your job in Founder Mode with Duet:

• Write sharp briefs with outcomes, not step lists
• Review PRs, drafts, and designs against taste, not grammar
• Send work back with a one-line correction and a pointer to memory
• Update the brief or the memory when the output drifts, not the one-off prompt

The agent handles the execution loop. You handle the judgment loop. That trade is the entire Founder Mode unlock.

Real repos and skills to install this week

You do not need to build this stack from scratch. The ecosystem already shipped the pieces.

Official Anthropic catalogs

• anthropics/skills. Anthropic's official skills repo. Includes starter skills for document generation, data analysis, research, and creative work. Load these into Duet as your default bench.
• anthropics/knowledge-work-plugins. Pre-built plugins for sales, marketing, finance, and general productivity. Drop-in quality for non-engineering roles.

Duet team's own stacks

• anianroid/claude-code-starter-pack. The full starter pack we use internally. Skills, CLAUDE.md templates, MCP configs, and the patterns that actually stick.
• Sawyer-Middeleer/cc-gtm. Our GTM skill pack: market research, contact enrichment, copywriting in your voice, CRM cleanup. Every skill from the Claude Code for Founders guide lives here.

Community power packs (steal these today)

• garrytan/gstack. Garry Tan's opinionated Claude Code setup. 23 slash-command specialists (CEO review, eng review, design review, ship, QA) that turn one agent into a virtual engineering team. 20k+ stars. The closest open-source analog to running Founder Mode.
• garrytan/gbrain. Garry Tan's open-source AI knowledge system. Long-term memory for agents, stored as markdown in a git repo. People, companies, meetings, transcripts, ideas. Agents scan and enrich it while you sleep. Exactly the memory repo layer above.
• obra/superpowers. Jesse Vincent's agentic skills framework. Composable skills for TDD, systematic debugging, brainstorming, and subagent-driven development. 355k+ installs. Enforces red-green-refactor and root-cause debugging before any fix.
• blader/humanizer. Claude skill that strips AI-generated tells from writing. Based on Wikipedia's "Signs of AI writing" guide. Runs a second audit pass to catch lingering AI-isms. Drop this in your writing channel before any draft leaves the building.
• nextlevelbuilder/ui-ux-pro-max-skill. Design intelligence skill. 50+ styles, 161 color palettes, 99 UX guidelines, 10 stack targets (React, Next, Svelte, SwiftUI, Flutter, etc.). Activates automatically on any UI/UX brief. Closest thing to a senior designer on the bench.
• gsd-build/get-shit-done. Spec-driven meta-prompting system by TÂCHES. Spawns fresh subagent contexts for each task so long sessions do not degrade. Built-in schema drift detection, security enforcement, TDD pipeline mode. The execution engine that keeps long-running workstreams from drifting.
• travisvn/awesome-claude-skills. Curated discovery list of every serious Claude skill in the wild. Your feed for when the starter bench isn't enough.

Watch and read (context on why this works)

• Garry Tan announcing gbrain on X. The original "I might give it away as open source" thread. Context on what long-term agent memory looks like in production.
• Garry Tan launching gstack on LinkedIn. The viral launch post that turned skill packs into a public category.
• TechCrunch: Why Garry Tan's Claude Code setup got so much love, and hate. Read both sides before adopting.
• Jesse Vincent: How I'm using coding agents. The backstory behind Superpowers. Read it to understand why TDD-enforced agent workflows beat free-form prompting.
• Simon Willison's notes on Superpowers. A second expert's read. Great for quick skim.
• SitePoint: GStack Tutorial for 10K LOC/Week. Guided install if you want more than the README.

Your own stack

The real power comes from the skills you build around your product and your voice. Start with three:

1. A memory skill that loads brand voice, positioning, and past decisions into every session.
2. A writing skill trained on your best public writing (landing pages, tweets, blog posts). Every first draft should already sound like you.
3. A review skill that runs your personal pre-ship checklist against PRs, drafts, and designs before you look at them.

For the full build pattern, see How to Build a Shared AI Skill Library.

The 7-day Founder Mode sprint

Founder Mode as a system does not need a quarter. It needs a week.

Day 1: Pick your bottleneck channel

Open duet.so and create one channel. Do not create seven. Pick the single function eating the most founder time right now. Design. Copy. Cold outreach. Customer research. Whatever makes you groan on Sunday night.

Day 2: Load memory

Upload or paste into your Duet memory the following:

• Your top 5 public writing samples (tweets, blog posts, landing page copy)
• Your positioning doc or a 200-word "what we sell and why" brief
• The last 3 customer calls as transcripts or notes
• Your 10 most important past decisions and the reasons behind them

This takes two hours. It pays back every session after.

Day 3: Install one skill

Pick one skill from anthropics/skills or cc-gtm that maps to the bottleneck. Install it in the channel. Do not install five. Five at once means debugging five at once.

Day 4: Run one real job end-to-end

Not a demo. Not a toy. A real piece of work with a real deadline. Brief the agent with an outcome, not a task list:

"Ship a 3-email re-engagement sequence for users who signed up, created an account, and never came back. Length: 80 words each. Goal: get them to log back in this week. Voice: match the samples in memory. Ready to paste into [your ESP] by EOD."

Review the output. Redirect. Ship.

Day 5: Tighten the brief, tighten the memory

Whatever went wrong, fix the root cause:

• If the tone drifted, add 3 more writing samples to memory
• If the CTA was weak, update the brief template with a CTA rubric
• If the agent hallucinated a feature, add the product spec to memory

Do not fix the one-off prompt. Fix the system.

Day 6: Spin up a second channel

Same pattern, different function. If Day 1 was design, Day 6 is growth. If Day 1 was writing, Day 6 is code review.

Day 7: Run parallel

Open both channels on a split screen. Brief both agents. Review both outputs. This is the first day you will feel the physics break.

By Day 8 you will stop asking whether agents can run execution. You will start asking how many parallel channels you can sanely orchestrate.

The daily Founder Mode operating cadence

Cadence is where Founder Mode either becomes an operating system or dies as a good intention.

Morning: 30-minute review block

Open Duet. Skim every channel. You are looking for three things:

• What shipped overnight that needs your taste filter
• What is stuck waiting on a founder decision
• What the agents flagged that you would have missed

Respond in-line. Ship the good stuff. Send the bad stuff back with one-sentence corrections.

Midday: outcome-based briefs

Write 2-3 new briefs for the afternoon. Each one names an outcome, acceptance criteria, and the relevant skills. Skip anything that does not need you to write the brief yourself.

Afternoon: parallel execution (you are not in this loop)

The agents work. You do the thing only you can do: talk to a customer, close a deal, make the strategic call, or delete a feature.

Evening: ship and curate memory

30 minutes at end of day. Ship what's ready. Then curate memory:

• One positive example into the "good output" folder
• One correction into the "voice rules" file
• One decision into the "decision log"

Memory that does not get curated decays. Memory that gets curated compounds.

Weekly: the Sunday sharpen

One hour every Sunday. Audit the week:

• Which briefs worked? Why?
• Which outputs missed? What did the memory or skill lack?
• Which channel is ready to hand off to the team?
• Which skill needs to graduate from experimental to default?

This is the loop that turns Founder Mode from a vibe into a flywheel.

Founder Mode prompts to paste in Duet

Steal these. Adapt the bracketed pieces.

The outcome brief

Outcome: [final state, one sentence]
Success metric: [measurable thing, numeric if possible]
Skills to load: [list the skills from your bench]
Memory to load: [brand voice, last 3 customer calls, positioning doc]
Constraints: [length, tone, banned words, deadlines]
Out of scope: [what NOT to do]
First deliverable: [what I want to see in my next review]

The review-and-redirect

Good: [what worked, 1 line]
Drifted: [what missed, 1 line]
Fix: [specific correction, 1 line]
Update memory: [the rule or sample to add so this never happens again]
Ship when: [new acceptance criteria]

The memory curation prompt

Read the last 5 shipped deliverables in this channel.
Extract: voice patterns that worked, phrases I repeated, CTAs that landed.
Update the voice memory file with a "confirmed patterns" section.
Flag: any pattern I used that contradicts the existing voice guidelines.

The "fire the bloat layer" prompt

Audit my last 30 days of work in this channel.
List every decision that waited more than 24 hours for me.
For each: suggest a skill, memory update, or pre-approved rule that would let an agent handle it next time without founder involvement.
Flag the 3 highest-leverage changes.

What good looks like: the 3-person company that out-ships the 50-person company

The era of building a management pyramid to scale a software company is ending. The companies winning now are not the ones that hire the fastest. They are the ones that stay close to their product the longest.

A 3-person team running Founder Mode in Duet can out-ship a 50-person org if:

• Memory is curated weekly
• Skills are specialized, not generic
• Briefs are written as outcomes, not tasks
• Founders review, and never become the assignment desk
• Every human hire replaces an agent that hit a ceiling, not an agent that worked

Founder Mode used to be a personality. With Duet, it is an operating model. Load your bench. Keep your team small. Stay in the details. Ship at the speed you did on day one, at any size you want.

FAQ

Keep reading

• How to Build Your AI Command Center with Duet. The full setup for your Duet workspace.
• Claude Code for Founders. The GTM, product, and ops skill pack that started it all.
• Agent Skills 101: Tools vs MCP vs Skills. The mental model for how skills actually work.
• How to Build a Shared AI Skill Library. Package tribal prompts into reusable team capability.