Aakash GuptaBuild a Growth Operating System in Claude Code, Step by Step
CHAPTERS
- 0:00 – 5:21
Why AI unlocks “full‑stack growth”: shipping experiments without waiting on engineers
James frames how AI agents (especially via Claude Code) let growth practitioners run experiments and ship changes dramatically faster, even as a solo operator. Aakash introduces James’s growth track record (ElevenLabs, consulting) and sets the episode’s focus: applying agentic workflows to growth work like SEO, ads, and marketing.
- •AI enables growth teams to execute end-to-end experiments faster with fewer engineering bottlenecks
- •Rise of “full‑stack growth” roles that can ship directly with the right guardrails
- •James’s background: CEO-led growth, ElevenLabs hypergrowth, now teaching agentic growth workflows
- •Episode scope: turning Claude/agents into a practical growth operating system
- 5:21 – 6:52
The “Growth Agents” repo: tools + skills packaged as an agentic growth operating system
James introduces the Growth Agents GitHub repo and how it installs a working scaffold for agentic growth. The core idea is to combine reusable “skills” (PRDs, audits, retros) with tools that have strong CLIs/APIs so outputs flow back into the agent reliably.
- •Repo provides an installable scaffold for running growth work inside Claude Code (and other agents)
- •Emphasis on tools with strong APIs/CLIs to keep workflows automatable
- •Skills library includes routines (standups, retros) and execution skills (PRDs, SEO audits, etc.)
- •System can customize itself based on your role/goals and improve via scheduled routines
- 6:52 – 12:05
The goal cascade: mission → OKRs → monthly/weekly retros → daily focus
James explains his anti-rabbit-hole structure: long-term mission, quarterly OKRs, monthly retros driven by connected data sources, and weekly retros to keep execution honest. He highlights how hooks and reminders in Claude help prevent distraction and ensure work ladders up to real outcomes.
- •Mission file anchors everything; quarterly OKRs tie execution to strategy
- •Monthly retro pulls quantitative data from tools (e.g., Stripe/PostHog/CRM/accounting)
- •Weekly retro blends personal reflection with hard metrics to reduce self-deception
- •Session hooks keep Claude constantly reminding you of priorities and pushing back on distractions
- 12:05 – 14:44
Are AI standups/retros actually useful? Fixing context gaps and spotting patterns
Aakash challenges whether AI retros are worth the overhead given fast-changing context. James argues the value comes from iteration on the skills and the agent’s ability to detect behavioral patterns (e.g., procrastination via “overbuilding”) using integrated to-do lists and repos.
- •AI retros can fail if context isn’t structured; skills must be iterated
- •Weekly retros are highest ROI; daily standups can be optional/low value
- •Agents can detect recurring avoidance patterns by inspecting tasks/commits
- •Accountability benefit: agents enforce alignment to stated goals and metrics
- 14:44 – 25:30
Live setup: why James runs Claude Code in the terminal (not Cursor/Desktop)
James demonstrates spinning up an agent in a demo folder and explains his preference for terminal-based Claude Code. He describes how onboarding interviews and structured context creation improve downstream outputs, even if the initial setup feels slow.
- •Terminal workflow is faster/simpler; less distraction vs IDE-centric workflows
- •Claude Heads-Up Display and multi-window terminals for parallel agent work
- •Onboarding can be “rapid” or a deeper 1–2 hour interview for better long-term performance
- •Agent audits existing folders rather than forcing a new “second brain” structure
- 25:30 – 30:30
Skill library spotlight: SEO audit workflow with Firecrawl + DataForSEO (vs Ahrefs/Semrush)
James walks through a practical SEO audit skill and the tools that make it reliable and cheap. He explains how Firecrawl avoids scraping issues and how DataForSEO can replace expensive subscriptions for many core SEO tasks.
- •SEO audit checks technical SEO, crawlability, metadata, and performance (Lighthouse)
- •Firecrawl converts sites to Markdown to bypass scraping/Cloudflare issues
- •DataForSEO enables low-cost keyword research, gap analysis, and keyword universe building
- •Argument: many paid SEO suites are clunky for building actionable keyword plans vs agentic workflows
- 30:30 – 36:23
Autonomous agents reality check: scheduled skills vs “agents that print money”
Aakash asks whether autonomous agents are overhyped. James clarifies his system is an operating system with scheduled routines and on-demand agent work, not a fully autonomous growth team, and explains what “autonomy” realistically looks like today.
- •Most “autonomous” claims are summarization or periodic monitoring, not self-directed execution
- •Practical autonomy today: run jobs on schedules (e.g., weekly site regressions) and report back
- •Human still steers priorities; agent executes chunks of work for hours at a time
- •Operating system framing: structure + skills + integrations drive results more than autonomy hype
- 36:23 – 37:45
SEO audit results on Aakash’s site: paywalls, missing descriptions, crawl issues—and fast fixes
James reviews the audit output for Aakash’s newsletter site and highlights issues like paywalled content implications, missing SEO descriptions, and crawlability problems. He contrasts the old world (agency report + engineering backlog) with the new world: turning the report into a PRD and fixing rapidly with agents.
- •Common findings: missing descriptions, long/duplicative titles, crawlability, paywall considerations
- •Historical contrast: weeks for agency report + weeks for engineering fixes
- •New workflow: convert audit into an actionable PRD and execute quickly in-repo
- •Aakash notes real SEO traffic decline, making the audit immediately relevant
- 37:45 – 41:23
Multi-agent refine: using a panel of critics to stress-test plans before execution
James introduces “multi-agent refine,” spawning multiple subagents with different lenses (technical correctness, simplicity, missing gaps, feasibility). He explains why adversarial review consistently surfaces blind spots compared to a single agent reviewing its own work.
- •Subagents critique from different perspectives and return structured weaknesses
- •Useful for both product work (payments infra) and growth strategy/planning documents
- •“What’s missing?” critic is especially effective for unknown unknowns
- •Practical tip: use cheaper models for subagents to avoid rate limits
- 41:23 – 45:38
Multi-agent refine demo on OKRs: turning messy goals into measurable, aligned outcomes
James runs multi-agent refine on a demo OKR document and shows how quickly critics flag stale timelines, weak measurability, and lack of alignment to company goals. He recommends using refine during the planning phase (audit → PRD → refine → build).
- •Demo flags: stale OKRs, unclear measurement, misalignment with company objectives
- •Refine synthesizes P0/P1/P2 issues for prioritization
- •Workflow recommendation: refine the PRD after audits and before building/fixing
- •Option to apply fixes autonomously or keep the human in the loop for decisions
- 45:38 – 52:43
Experiment discipline + PostHog + Claude: templates, closure loops, and event hygiene
James explains how the system enforces rigorous experimentation: hypothesis templates, baselines, confounders, stop conditions, and forced documentation. He then dives into PostHog best practices, emphasizing API-driven workflows, read-only Git access for growth, and fully documented event definitions.
- •Experiment templates enforce statistical rigor, confounders, and proper closure/documentation
- •Retros/standups check for stale experiments and force learnings into a knowledge base
- •PostHog + Claude enables creating dashboards/experiments via API rather than UI
- •Best practices: read-only Git access for event clarity; CI tests to require instrumentation; document all event definitions
- •Suggested prompt: have Claude audit telemetry/events, find gaps, and update definitions with correct permissions
- 52:43 – 1:07:29
Biggest mistakes, connecting every tool, the next frontier—and how James built a 6‑figure cohort business
James shares early mistakes (overtrusting agents, shipping bugs) and how tooling like code review and error reporting mitigates risk. He argues the largest multiplier is connecting every relevant tool via MCP/CLI/API so agents have full context. He closes with where agentic growth is heading (more self-directed agent “companies”), plus how he turned workshops into a cohort business and what learners commonly miss: asking agents to improve their own prompts/skills.
- •Early pitfall: overtrusting Claude; mitigate with code review + error reporting
- •Key leverage: integrate everything (CRM, email, enrichment, analytics) so workflows become one prompt
- •Future: a self-improving, 24/7 “company of agents” that can chain experiments end-to-end
- •Business story: workshop → online cohort → referrals → $100K/mo without full-time hires
- •Learning mistake: people don’t ask agents to improve their own system prompts/skills and iterate via reflection
