# Garry Tan System Reverse Engineering Dossier - 2026-05-16 ## Purpose This dossier reverse-engineers Garry Tan's public agent system down to product specs, JTBDs, personas, use cases, surface area, funnel/CTA strategy, and system architecture. It builds on: - projects/ren/daily/2026-05-16-garrytan-repos-map.md - projects/ren/daily/2026-05-16-garrytan-library-dissection.md - Local clones: gstack, gbrain, gbrain-evals - Public YC pages fetched on 2026-05-16: ycombinator.com, ycombinator.com/apply, ycombinator.com/software ## Executive Read Garry is not only publishing tools. He is publishing an operating model for a one-person or tiny-team AI software factory. The system has four jobs: 1. Prove a new work style: one builder with agents can ship at team-scale. 2. Package the work style as reusable methodology: gstack skills. 3. Give agents continuity: gbrain memory/runtime. 4. Make the claims credible: gbrain-evals benchmarks and gstack LOC methodology. The YC CTA is not bolted on. It is part of the funnel: gstack demonstrates the kind of high-agency, AI-native engineering YC wants around it, then routes engineers/founders toward YC. ## The System In One Sentence gstack is the workflow OS, gbrain is the memory substrate, gbrain-evals is the proof layer, and YC is the distribution/recruiting endpoint. ## Product Spec - gstack ### Product Name gstack ### Category AI engineering workflow OS / agent skill suite / browser-enabled software factory. ### Promise Turn Claude Code and adjacent coding agents into a virtual product/engineering team with roles, review loops, browser QA, security checks, docs, shipping, retros, and memory. ### Primary User A technical founder, CEO, staff engineer, or AI-native builder who still wants to ship product directly. ### Core Pain Blank-prompt coding agents are powerful but structurally weak: they do not know when to plan, when to push back, when to test, when to open a browser, when to ship, or how to maintain project memory. ### Desired Outcome The user can say what they want to build, and the system imposes a disciplined sprint loop: product framing, architecture, design review, implementation, code review, QA, security, docs, ship, deploy, and retro. ### Core Workflow Think -> Plan -> Build -> Review -> Test -> Ship -> Reflect. ### Key Product Surfaces - Markdown skills: named specialist roles exposed as slash commands. - Browser daemon: real Chromium with CDP bridge, sidebar, screenshots, cookies, terminal/pairing. - Host adapters: Claude, Codex, OpenCode, Cursor, OpenClaw, Hermes, Factory, Kiro, Slate, GBrain. - Setup/install: 30-second personal install plus team mode. - Safety layers: careful/freeze/guard, CSO, prompt-injection defenses. - Memory bridge: setup-gbrain and sync-gbrain. - Docs and release automation: document-generate, document-release, ship, land-and-deploy. ### gstack Success Metric Reduced time from idea to verified PR/deploy while preserving quality. Publicly, Garry supports this with LOC methodology, install counts, test counts, and visible shipped repos. ## Product Spec - gbrain ### Product Name gbrain ### Category Personal/agent memory runtime / local knowledge brain / MCP tool surface. ### Promise Give agents durable memory they can search, update, cite, and maintain across sessions, machines, and sources. ### Primary User A power user or agent operator whose work depends on accumulated context: people, companies, decisions, meetings, notes, tasks, code, research, and original ideas. ### Core Pain Agents forget. Notes rot. Search misses intent. Conversations and files become unstructured sludge. Personal context is spread across meetings, docs, chats, repos, email, social, and local files. ### Desired Outcome The agent checks the brain first, retrieves relevant context, writes back new facts with citations and backlinks, and keeps the knowledge graph healthy. ### Core Data Model - Files/pages are the user-visible source of truth. - Postgres or PGLite is the runtime/query index. - Sources map external corpora into the brain. - Chunks and embeddings power retrieval. - Links/backlinks/timeline/source attribution keep context navigable. - MCP and CLI expose the brain to agents. ### Key Product Surfaces - CLI: init, sync, embed, search, query, recall, think, doctor, serve, serve-http, sources, pages, files, import, export, upgrade, jobs, agent, autopilot, dream, eval. - MCP: typed agent operations over stdio/HTTP. - Skills: ingestion, enrichment, query, maintenance, briefing, publishing, agent ops. - Storage engines: Postgres/pgvector and PGLite. - Agent orchestration: Minions durable job queue. - Integrations: meetings, email, calendar, voice notes, media, webhooks, credential gateway. ### gbrain Success Metric Retrieval correctness, freshness, citation quality, graph quality, and operational health. Public proof includes LongMemEval, BrainBench, source-swamp tests, and comparison-system notes. ## Product Spec - gbrain-evals ### Product Name gbrain-evals ### Category Benchmark corpus and reproducibility harness for agent memory systems. ### Promise Make memory quality measurable and publishable. ### Core Pain Every memory product can claim it remembers. Few can prove retrieval, temporal reasoning, source resistance, and agent-loop behavior with reproducible numbers. ### Desired Outcome A gbrain release or architecture claim can cite benchmark evidence. ### Key Assets - LongMemEval public benchmark report: gbrain-hybrid 97.60% R@5 on the _s split. - BrainBench synthetic corpora and category runners. - Source-swamp resistance benchmark. - Minions vs OpenClaw subagent benchmark. - Comparison systems page separating retrieval recall from QA accuracy. ## Product Spec - YC Funnel Layer ### Product Name YC CTA / software recruiting / founder application surface. ### Category Distribution, recruiting, and community conversion layer. ### Evidence gstack README ends with: - Free, MIT licensed, open source. No premium tier, no waitlist. - Come work at YC - ycombinator.com/software. - MIT. Free forever. Go build something. YC software page positioning: - Engineers at YC are at the center of the startup ecosystem. - They work near YC founders, partners, investors, and guest speakers. - Many YC engineers have founded companies or plan to. - CTA: email software@ycombinator.com with a brief note about yourself and what you would want to build. YC application page positioning: - YC invests $500k in selected startups. - Batch is in-person in San Francisco. - Applications are submitted online. - Late applications can still be considered. - YC continues helping founders after the batch. ### Funnel Interpretation gstack attracts AI-native builders. The README demonstrates credibility, gives them a useful free product, then points the best-fit users toward YC as employees, founders, or community members. ## JTBD Map ### Job 1: Ship Like A Team While Staying Small User: technical founder / CEO / solo builder. Situation: I have product ideas and limited time, and blank coding agents create as much management work as code. Job: Help me turn a feature idea into a shipped, reviewed, tested implementation without hiring a full product/engineering team. Current alternatives: manually prompting Claude/Codex, hiring contractors, waiting for team bandwidth, using Cursor ad hoc. gstack solution: office-hours -> CEO review -> eng/design/DX review -> implementation -> review -> QA -> ship. ### Job 2: Make Agents Remember What Matters User: founder/operator with recurring relationships, decisions, meetings, and projects. Situation: Every agent session starts cold and loses prior context. Job: Give my agent a persistent memory it can search, update, and cite before acting. Current alternatives: grep, Notion, Obsidian, long prompts, chat history. gbrain solution: files/pages as source of truth, indexed runtime, MCP tools, brain-first skills, source attribution, backlinks. ### Job 3: Turn Personal Context Into Operational Leverage User: executive, investor, founder, chief of staff, or EA-like agent operator. Situation: Meetings, emails, voice notes, social signals, and docs contain relationship and decision context, but none of it compounds. Job: Convert ambient signals into structured memory pages that improve future responses and briefings. gbrain solution: meeting-ingestion, enrich, signal-detector, daily-task-prep, briefing, voice-note-ingest, data-research, webhooks. ### Job 4: Trust Agent Output Before It Touches Production User: engineer, founder, product lead. Situation: AI can make fast changes but misses edge cases, security issues, UI bugs, docs drift, and slop. Job: Add a disciplined quality gate before merge/deploy. gstack solution: review, cso, qa, design-review, devex-review, document-release, canary, benchmark, careful/freeze/guard. ### Job 5: Prove Memory Quality User: system builder, AI infra evaluator, maintainer. Situation: Retrieval systems can look good in demos but fail on temporal, noisy, or long-context recall. Job: Measure whether the memory system retrieves the right context under realistic conditions. gbrain-evals solution: LongMemEval, BrainBench, source-swamp resistance, comparison-system caveats. ### Job 6: Join The AI-Native Builder Circle User: ambitious engineer/founder. Situation: I want to work at the frontier of startup software and AI-native development. Job: Find the people and institutions building this way. YC CTA solution: free product first, then recruiting/founder application routes: ycombinator.com/software and ycombinator.com/apply. ## Persona Map ### Persona A: Technical Founder Who Still Codes Needs: speed, taste, quality, product challenge, shipping discipline. Uses: office-hours, autoplan, plan-ceo-review, plan-eng-review, ship, gbrain memory. Trigger: wants to build a feature or startup wedge quickly without expanding headcount. ### Persona B: Founder/CEO Who Wants A Product Staff Needs: product reframing, strategic pushback, execution options, scope control. Uses: office-hours, CEO review, design consultation, retro. Trigger: has a product direction but needs better thinking before code. ### Persona C: Staff Engineer / Tech Lead Needs: review, architecture, QA, security, docs drift control. Uses: plan-eng-review, review, cso, health, document-release, benchmark. Trigger: AI-generated changes need senior-level gatekeeping. ### Persona D: Agent Power User Needs: persistent context, cross-session memory, brain-first lookup, exact citations. Uses: gbrain query, enrich, signal-detector, daily-task-prep, maintain. Trigger: repeated work where missing context wastes time or damages quality. ### Persona E: Executive Assistant / Chief Of Staff Agent Operator Needs: meeting context, task prep, relationship memory, summaries, reminders. Uses: briefing, meeting-ingestion, daily-task-manager, daily-task-prep, enrich, reports. Trigger: preparing for calls, tracking open loops, remembering people and companies. ### Persona F: AI Tool Builder / OSS Contributor Needs: architecture examples, eval harness, host adapters, reproducible benchmarks. Uses: adding hosts, gbrain-evals, skillify, functional-area-resolver, OpenClaw integration. Trigger: wants to extend the system or compare it to their own agent stack. ### Persona G: YC Software Candidate Needs: signal that YC is building seriously with AI and values high-agency engineers. Uses: gstack as proof artifact; CTA to ycombinator.com/software. Trigger: sees Garry's open-source stack and wants to work near the founders and products using it. ### Persona H: YC Founder Applicant Needs: ambition, speed, proof that small teams can do more, access to YC network. Uses: gstack narrative as belief reinforcement; CTA to ycombinator.com/apply. Trigger: realizes AI-native software factories change what a tiny startup can build. ## Use Case Catalog ### gstack Use Cases 1. Start a product idea: office-hours asks forcing questions and reframes the problem. 2. Review a plan strategically: plan-ceo-review expands, reduces, or challenges scope. 3. Lock implementation architecture: plan-eng-review produces data flow, edge cases, and tests. 4. Improve visual/product quality: plan-design-review and design-review catch slop. 5. Improve developer experience: plan-devex-review and devex-review test onboarding and APIs. 6. Build with multiple reviews: autoplan runs CEO, design, eng, DX, and adversarial review. 7. Debug properly: investigate enforces root cause before fixes. 8. Review code before landing: review analyzes diff risks and can auto-fix simple issues. 9. Test in a real browser: qa and browse drive Chromium. 10. Ship a PR: ship syncs base, runs checks, bumps version, opens PR. 11. Deploy and verify: land-and-deploy merges, waits for CI/deploy, runs canary. 12. Security audit: cso runs infrastructure-first threat modeling. 13. Keep docs current: document-release and document-generate. 14. Coordinate agents: pair-agent shares a browser with OpenClaw/Hermes/Codex/etc. 15. Preserve context: context-save/context-restore and learn. 16. Protect scope: careful, freeze, guard, unfreeze. 17. Benchmark performance: benchmark and canary. 18. Set up persistent memory: setup-gbrain and sync-gbrain. ### gbrain Use Cases 1. Initialize a brain: setup/init with PGLite or Supabase. 2. Search memory: search/query/recall. 3. Index code and docs: sync/import/embed. 4. Serve agents: MCP stdio/HTTP server. 5. Ingest meetings: meeting-ingestion with attendee enrichment and timeline merge. 6. Ingest ideas/links: idea-ingest and article-enrichment. 7. Preserve voice notes: voice-note-ingest with exact phrasing. 8. Enrich entities: enrich creates rich person/company pages. 9. Prepare the day: briefing, daily-task-prep, daily-task-manager. 10. Maintain graph quality: maintain, citation-fixer, frontmatter-guard. 11. Run durable jobs: minion-orchestrator and jobs. 12. Publish/share pages: publish and brain-pdf. 13. Migrate existing notes: migrate from Obsidian/Notion/Logseq/etc. 14. Build skills: skill-creator, skillify, testing, functional-area-resolver. 15. Research with brain context: perplexity-research, data-research, academic-verify. 16. Build ambient capture: signal-detector and webhook-transforms. ### gbrain-evals Use Cases 1. Validate retrieval changes before release. 2. Compare hybrid vs vector vs keyword retrieval. 3. Publish credible benchmark claims. 4. Test source-swamp resistance. 5. Benchmark durable job queues vs subagent workflows. 6. Separate retrieval recall claims from QA accuracy claims. ## Skill Inventory - gstack Total: 52 SKILL.md files found. - autoplan: auto-review pipeline. - benchmark: performance regression detection. - benchmark-models: cross-model benchmarking. - browse: fast browser QA and dogfooding. - canary: post-deploy monitoring. - careful: destructive-command warnings. - codex: Codex CLI wrapper and review path. - context-save / context-restore: working-context persistence. - cso: chief security officer audit. - design-consultation: product design exploration. - design-html: production HTML/CSS generation. - design-review: visual QA and slop detection. - design-shotgun: multiple design variants and comparison board. - devex-review: live developer experience audit. - document-generate: create missing docs. - document-release: update docs after changes. - freeze / unfreeze: edit-scope lock. - gstack-upgrade: update installed gstack. - guard: careful + freeze. - health: code quality dashboard. - investigate: root cause debugging. - land-and-deploy: merge, deploy, verify. - landing-report: release queue snapshot. - learn: manage project learnings. - make-pdf: markdown to publication-quality PDF. - office-hours: YC-style product interrogation. - open-gstack-browser: launch browser UI. - OpenClaw-native skills: office-hours, ceo-review, investigate, retro. - pair-agent: shared browser for cross-agent coordination. - plan-ceo-review: founder/CEO plan review. - plan-design-review: design plan review. - plan-devex-review: developer-experience plan review. - plan-eng-review: engineering plan review. - plan-tune: question sensitivity and developer psychographic. - qa / qa-only: browser QA with or without fixes. - retro: weekly engineering retrospective. - review: pre-landing PR review. - scrape: web data extraction. - setup-browser-cookies: import real browser cookies. - setup-deploy: configure deployment settings. - setup-gbrain: install/init/register gbrain. - ship: PR shipping workflow. - skillify: codify scrape flows into skills. - sync-gbrain: keep code/memory current in gbrain. ## Skill Inventory - gbrain Total: 43 SKILL.md files found. - academic-verify: verify research claims/citations. - archive-crawler: crawl personal archives. - article-enrichment: structure article dumps. - ask-user: reusable decision pattern. - book-mirror: personalized book analysis. - brain-ops: core read/write cycle. - brain-pdf: render brain pages via gstack make-pdf. - briefing: daily briefing. - citation-fixer: citation audit/repair. - cold-start: day-one data bootstrap. - concept-synthesis: concept map and idea evolution. - cron-scheduler: schedule management. - cross-modal-review: second-model quality gate. - daily-task-manager: task lifecycle. - daily-task-prep: morning prep. - data-research: structured trackers from sources. - enrich: person/company enrichment. - frontmatter-guard: YAML/frontmatter validation. - functional-area-resolver: compress routing tables. - idea-ingest: links/articles/ideas to brain. - ingest: content router. - install: deprecated in favor of setup. - maintain: brain health checks. - media-ingest: video/audio/PDF/book/screenshots/repos. - meeting-ingestion: meeting transcripts and entity propagation. - migrate: import from external note systems. - minion-orchestrator: durable jobs and subagents. - perplexity-research: brain-augmented web research. - publish: password-protected HTML page sharing. - query: answer with brain knowledge. - repo-architecture: filing rules. - reports: save/load timestamped reports. - setup: set up GBrain. - signal-detector: always-on inbound signal capture. - skill-creator: create skills. - skillify: turn features into tested skills. - skillpack-check: health report for host agents. - smoke-test: post-restart health and auto-fix. - soul-audit: generate identity/user/access/heartbeat config. - strategic-reading: apply source text to a strategic problem. - testing: conformance and test-suite intelligence. - voice-note-ingest: exact-phrasing voice capture. - webhook-transforms: external events to brain pages. ## Funnel And CTA Map ### gstack Funnel 1. Credibility hook: Karpathy quote and Peter/OpenClaw reference. 2. Founder authority: Garry identifies as YC CEO and former builder. 3. Proof: 3 production services, 40+ shipped features, LOC methodology. 4. Product promise: virtual engineering team in Markdown skills. 5. Free/open-source commitment: MIT, free forever, no premium tier. 6. Immediate CTA: install gstack in 30 seconds. 7. Team CTA: add gstack to current project so teammates get it. 8. Platform CTA: install for OpenClaw or other AI agents. 9. Memory CTA: set up GBrain. 10. Recruiting CTA: Come work at YC - ycombinator.com/software. 11. Founder/community CTA: YC homepage/apply pages say it is never too early to apply and point to online application. ### CTA Psychology - Try it first: reduces skepticism and converts critics into users. - Fork/improve/make it yours: OSS community activation. - Team mode: turns individual install into org propagation. - OpenClaw/other agents: avoids Claude-only ceiling. - GBrain: upsells from process to durable memory. - YC software: routes high-agency engineers into YC's hiring funnel. - YC apply: routes ambitious builders/founders into YC's founder funnel. ## Strategic Narrative ### Narrative 1: The Solo Builder Can Be A Software Team Claim: one person with the right agent workflow can ship like a team of twenty. Proof assets: gstack README, contribution graphs, LOC methodology, open repos. ### Narrative 2: Methodology Beats Raw Model Power Claim: blank agents are not enough. The leverage is in roles, gates, prompts, and tooling. Proof assets: gstack's specialist skill suite. ### Narrative 3: Memory Is The Next Missing Layer Claim: agents become far more useful when they remember decisions, people, code, and context. Proof assets: gbrain, gstack setup-gbrain/sync-gbrain, skillpack. ### Narrative 4: Benchmarks Create Trust Claim: memory/retrieval must be measured, not demoed. Proof assets: gbrain-evals and LongMemEval results. ### Narrative 5: YC Is The Natural Home For This Workstyle Claim: YC funds and hires people building at the frontier of tiny-team leverage. Proof assets: gstack's YC identity, YC software recruiting page, YC application page. ## Reverse-Engineered Product Requirements ### gstack Requirements - R1: Install in under a minute for one user. - R2: Install into team repos with auto-update. - R3: Expose named skills that map to real product/engineering roles. - R4: Run inside Claude Code first, but support other host agents. - R5: Provide real browser operation, not only text plans. - R6: Defend against prompt injection and destructive commands. - R7: Support end-to-end PR workflow from plan through ship. - R8: Keep docs synchronized with code changes. - R9: Preserve learnings across sessions and machines. - R10: Integrate with gbrain for durable memory and code search. ### gbrain Requirements - R1: Work locally with PGLite and in cloud/shared mode with Postgres/Supabase. - R2: Make files/pages user-visible and portable. - R3: Index content into searchable chunks and embeddings. - R4: Expose CLI and MCP operations. - R5: Support source attribution, backlinks, timelines, and citations. - R6: Ingest many formats and external signals. - R7: Provide health checks, migrations, and repair paths. - R8: Support agent orchestration through durable jobs. - R9: Benchmark retrieval and memory quality. - R10: Avoid silent corruption: validate schemas, frontmatter, migrations, and embedding dimensions. ### gbrain-evals Requirements - R1: Keep benchmark bulk outside production gbrain. - R2: Pin corpora and comparison methodology. - R3: Separate retrieval recall from answer accuracy. - R4: Support adapter comparisons. - R5: Publish reproducible reports with caveats. ## Product Gaps / Open Questions 1. gstack and gbrain are both moving extremely fast; the product story is ahead of install reliability on Windows and some remote/MCP paths. 2. gstack's strongest moat is methodology, but methodology can be copied; durable memory and eval proof are the harder-to-copy layers. 3. gbrain has broad surface area; the risk is operator confusion and health noise. 4. gbrain-evals is powerful but under-marketed inside the main funnel. 5. YC CTA is present, but there may be a stronger founder-facing CTA: use gstack to build your YC application/product demo faster. 6. OpenClaw integration is strategically important: OpenClaw can be orchestration, gstack can be methodology, gbrain can be memory. ## Ren/OpenClaw Implications Ren should study this as an integrated reference architecture: - gstack gives Ren a language for role-based contribution workflow. - gbrain gives Ren a model for durable knowledge and agent memory. - gbrain-evals gives Ren a model for proving memory/retrieval quality. - YC funnel shows how OSS proof can double as recruiting and community acquisition. Possible Ren product wedge: OpenClaw as the orchestrator that routes work across agents; gstack as optional methodology pack; gbrain as durable context engine; Ren as the contribution/opportunity intelligence layer. ## Next Dossier Modules To Build 1. gstack product teardown by skill: one page per skill with trigger, user job, input, output, dependencies, and quality gate. 2. gbrain data model map: pages, sources, chunks, embeddings, links, facts, takes, timelines, jobs, auth. 3. gbrain skillpack JTBD map: every skill mapped to user job and brain directory. 4. Funnel teardown: exact copy/CTA map from README, YC software, YC apply, and Garry social proof. 5. Competitive map: gstack vs Cursor/Copilot/Claude Code workflows; gbrain vs Mem0/Supermemory/MemPalace/Obsidian/Notion; gbrain-evals vs LongMemEval/LoCoMo/ConvoMem. 6. Contribution map: unclaimed issues that improve the system's strategic story. ## Working Thesis For Hiten Garry's system is an OSS proof-of-concept for AI-native company-building. It packages his personal operating system as software, proves it with public artifacts, gives away enough value to create adoption, and routes the best-fit users back into YC. The product is not just gstack. The product is the belief that a tiny team with agents, memory, and rigorous workflow can out-ship a traditional team. gstack is the method, gbrain is the continuity, gbrain-evals is the proof, and YC is the network where that belief compounds.