AI Catalyst C3·Core Session - Week 12·3:10:11

Session 24: Building a Marketing OS — Sector-Agnostic Command Center on an Open-Source Stack

Harshith Vaddiparthy Trainer — Outskill; builds a complete marketing operating system live (website + dashboard monorepo) and opens his entire real playbook: self-hosted PostHog analytics, a 178k-contact Twenty CRM, his Resend/React-Email newsletter infrastructure with real open-rate numbers · Shivani Cohort manager — logistics, CSAT poll, next-morning resource posting

▶local recording · LMS lesson

Session map

The frameThe machineThe craft

click a node — its card pops up (drag it anywhere, × to close)

Concept

The map reads left to right — the frame flow into the machine, then into the craft. Click any node to open that idea here; every timestamp jumps into the recording.

The short version

The Marketing OS is a command center, not a tool: analytics + CRM + newsletter + SEO/GEO in one dashboard, 'everything interconnected... that's why it's an operating system' — with the architecture split that makes it reusable: the WEBSITE app is swappable per industry, the OS 'doesn't really care about skincare or real estate.'
The session's most portable move is the HANDOVER DOCUMENT: when Claude Code went rogue (Playwright unasked, bad output, token burn), he demanded 'a handover document of everything that you have just done' and moved to Codex without losing the project — the day-shift/night-shift pattern under fire.
The stack doctrine is open-source-first with honest exceptions: PostHog (analytics, session replays, funnels — free cloud tier or VPS self-host), Twenty CRM (self-hosted, 178k contacts), React Email (open) — but Resend is paid because SMTP self-hosting 'is very, very complex.'
His audience machine, shown raw: a viral tweet's engagers harvested via Apify into the CRM (Twitter handles + emails), permission-first outreach ('I would always take permission before adding them'), the site pop-up feeding Resend, sources attributed, list health watched to protect the domain he 'spent a lot to acquire.'
Newsletter craft with real numbers: ~40% open rate against an 'industry standard' of 30-35% for elite lists; the subject-line lesson learned on his own data — 'I went viral on X' flopped ('nobody cares'), 'The Stripe for AI' won ('it makes them curious') — value-add beats bragging.
Meta-observations worth keeping: 'Claude has an attitude since Fable 5'; 'the first version of Fable was so good... After that, it's not as good'; enterprises choose Claude for data security while he leans Codex for speed — the cohort's live model-market read, one day after the sprint's subsidy math.

The concepts

The Marketing OS: sector-agnostic command center over swappable websites

'The marketing OS doesn't really care about skincare or real estate. It's the website. If the website is built out, the marketing OS will just connect to it.'

The architecture is one monorepo, two apps, two deploys: APPS/WEBSITE — the public marketing site (today: 'Velora' skincare, built with Three.js/3UI for consumer-grade interactivity, brand assets generated via GPT Image) — deliberately swappable ('my learners can change the entire website from skincare to supply chain and it should still work'); APPS/DASHBOARD — the OS itself (shadcn + TweakCN, dark mode), permanent, holding pages for analytics, CRM, and newsletter that INTEGRATE the real tools rather than reimplement them ('you don't have to add in all the tools, just build the dashboard, set up the pages... I'll tell you exactly what tools to integrate').

The integration thesis is the point: 'your CRM would be connected to your analytics. Anyone who signs up to your newsletter from your website lands in your CRM... everything is interconnected' — and the payoff compounds with agents: 'let's say we give it access to our Claude, it can make better informed decisions, as everything is in one place.' Deployment shape follows: monorepo ≠ one deploy — two Vercel projects, each with its own root directory. Crowd-sourced scope at the open: leads, content, campaigns, CRM, funnels, traffic, brainstorming→posting, analytics, plus yesterday's SEO/GEO folded in.

Worked example · from the session

The live gamification layer: his Excalidraw agenda converted into a checklist mini-site (one-shot by Codex on a spare port) with done-buttons he clicks as each stage lands — the session's own progress rendered as a product.

Why it matters

It's the operations counterpart to s23's visibility work — the same week teaching how to be FOUND and how to RUN what the finding feeds.

People get this wrong

A marketing OS is a SaaS you buy.

It's a thin, owned command center over open-source tools you self-host or free-tier — the expensive part (the tools) is free; the valuable part (the unification) is yours.

One command center over swappable surfaces, open-source floor underneath

Everything is interconnected, guys. That's why it's an operating system.

For your projects

The dashboard-as-integration-pane echoes your own KB site: thin pages over real data sources.

This idea elsewherebuilds onweek 12's pair: visibility layer then operations layer

Go deeper

In one line: Marketing OS = monorepo {apps/website (swappable, per-brand, interactive), apps/dashboard (permanent command center: analytics + CRM + newsletter pages integrating PostHog/Twenty/Resend)}; two Vercel deploys; integration-not-reimplementation; agent-readable by design.

Sector-agnostic split answered live to a cohort question — the OS ships once, websites rotate ()

Dashboard as 'one-stop shop... to look at everything and make informed decisions' — the mock-data version lands in minutes, integrations follow ()

Three.js/'3UI' as the consumer-site aesthetic lever ('SHADCN is for dashboards; 3UI is for interactive websites') ()

Deploy discipline: 'even with one Git repository, you deploy two apps on two separate Vercel projects' ()

The whole repo open-sourced to the cohort at session end — his standing pattern ()

▶ Watch this taught:

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

Why keep the dashboard tool-agnostic pages instead of building analytics/CRM features into it?

The real tools (PostHog, Twenty, Resend) are mature and open; the dashboard's value is UNIFICATION — one pane, one context for humans and agents. Rebuilding features would trade integration for maintenance.

The handover document: firing an agent without losing the project

'Hey Claude, you are very bad... Can you just give me a handover document of everything that you have just done, and I'll just move on to Codex, please.'

The incident, unedited: Claude Code (Opus 5) starts strong, then opens Playwright unasked, generates 'a very, very bad website' ('Build something people remember' — 'what is even that?'), and burns tokens — echoed by the room ('it seems to go off in its own direction, burn a lot of tokens, and then come back with less quality work'). The exit protocol: BEFORE killing the session, demand a MARKDOWN HANDOVER DOCUMENT summarizing everything done; verify it lands in the project folder; then hand the folder + document to the next agent, whose first instruction is 'read the entire handover document.'

Codex resumed cleanly and one-shot what Claude fumbled. The pattern generalizes his multi-tool workflow ('the only thing common is my folder' — from the sprint) and got stress-tested again later when PostHog's install wizard threatened his Claude Code credentials — the fix again explicit instruction: 'please do not mess with my Claude settings... the Claude settings are very, very important for me.' Balanced coda kept in: a cohort member argues they should have debugged Claude instead ('recalibration and guardrails... would have been a good exercise') and he half-agrees — swapping was a time call, not a verdict.

Worked example · from the session

The full arc on camera: rogue Playwright → stop → handover request → document verified in-folder → 'Alright, Claude, buh-bye' → Codex reads it and continues — maybe four minutes of ceremony for zero lost context.

Why it matters

Every multi-agent workflow eventually hits a misbehaving agent mid-task; this is the cheapest known exit that preserves the work.

People get this wrong

Switching agents mid-project means restarting the project.

It means externalizing state first. The handover document is the ten-minute bridge that makes agents interchangeable.

Can you just go ahead and give me a handover document of everything that you have just done, and I'll just move on to Codex, please.

For your projects

Direct kin to your project-handoff-todo skill — same artifact, agent-to-agent instead of chat-to-chat.

This idea elsewherereappears inthe day-shift/night-shift log, now used under fire

Go deeper

In one line: Agent-swap protocol: on misbehavior, (1) stop new work; (2) demand a markdown handover document of all work done, written into the project folder; (3) verify it exists; (4) instruct the successor agent to read it first. State lives in files, not in the dying session. Consider debugging-in-place when time allows — swapping is triage, not doctrine.

The trigger symptoms named: unrequested tool use (Playwright), off-brief output, token burn without progress ()

The document must land IN the folder — 'one good thing it did was the handover document is inside the same folder' ()

Successor's first prompt: 'read the entire handover document' before any building ()

Same-family incident: the PostHog wizard vs Claude credentials — explicit do-not-touch instructions as guardrails ()

Model-market color: 'enterprises use a lot of Claude' (data security) vs his Codex speed preference; 'the first version of Fable was so good' ()

▶ Watch this taught:

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

Why demand the handover from the FAILING agent rather than writing it yourself?

The failing agent still holds the session context (what it tried, changed, and why) that isn't in the files yet — the handover extracts that context into files before it evaporates.

The open-source marketing stack: PostHog, Twenty, React Email

Every 'secret tool' he reveals turns out to be open source — 'the things that I share are going to be open source' — with one honest exception, and the exception teaches the rule.

ANALYTICS — PostHog: open source, self-hostable, 'the best web analytics tool.' Full-fat product analytics: dashboards (DAU/WAU, funnels, drop-off, time-to-convert), SESSION RECORDINGS (cursor movements, rage-click detection, per-source filters like Google-Ads-only or newsletter-referred visitors — with emails/PII masked), heatmaps, feature flags, an MCP, and an AI wizard installer ('posthog wizard self-driving' — copy one command, the agent integrates the SDK). Hosting reality: it's heavy — 'think of it like n8n' — VPS not Vercel for self-host (his: insights.harshith.com); the cloud free tier is 'very generous' (~1M events/month). Privacy duties named: cookie disclosure, privacy-page declaration (his own shown).

CRM — Twenty ('20.com'): 'the number one open source CRM on GitHub' (~55k stars), self-host or managed; his instance holds 178,000 contacts at crm.harshith.vc on a VPS. Companies, people, opportunities, tasks, notes, lead-source tracking, even call recording. His take on its UI ('maybe their UI isn't that good') loops back to the OS thesis: keep Twenty as infrastructure, surface it in your own dashboard.

EMAIL — the honest exception: Resend is PAID because 'you can't really host your own SMTP. That's very, very complex' — but its composition layer, React Email, IS open source, with a template gallery whose 'Copy for AI' buttons emit ready prompts ('just paste the entire prompt and say: create the email') — Codex one-shot a branded Velora newsletter from one.

Worked example · from the session

The PostHog recordings demo on his live site: a visitor from Australia, Windows/Chrome, five clicks two minutes ago, newsletter pop-up dismissed — 'that gives me a data point. Why did they cross out the newsletter pop-up?'

Why it matters

The stack turns s23's traffic into understanding (PostHog), relationships (Twenty), and owned distribution (email) — at near-zero licence cost.

People get this wrong

Marketing tooling = expensive SaaS subscriptions.

Analytics and CRM are genuinely free at his scale (open source + generous tiers); only delivery is bought — the stack costs a VPS and one email subscription.

Think of it like n8n... the way you host n8n, very similarly, you would be hosting your PostHog as well.

For your projects

PostHog's newsletter-pop-up-dismissed event is exactly the consent/UX telemetry your email-deliverability and site-health skills reason about — with data instead of guesses.

This idea elsewherebuilds onthe open-source money model, now consumed from the user side

Go deeper

In one line: Stack: PostHog (product analytics + session replay; cloud free tier or VPS self-host; wizard/agent install; privacy disclosure required) · Twenty CRM (open source, self-host; infrastructure behind your own dashboard UI) · Resend (paid delivery — SMTP isn't self-hostable in practice) + React Email (open composition; Copy-for-AI templates). Heavy tools live on VPSes, not serverless.

PostHog per-source replay filters are the growth loop: watch exactly what ad-clickers or newsletter-readers do ()

Rage clicks as UX telemetry: 'the button has an issue... it could mean a lot of different things' ()

The n8n hosting analogy: PostHog and Twenty are VPS-class apps; Vercel won't hold them ()

Twenty's 7-day trial card-gate on cloud vs fully-free self-host — the recurring open-core shape ()

React Email's Copy-for-AI buttons: templates shipping their own integration prompts — a pattern to watch ()

PII posture: recordings mask emails/identities; declare analytics in the privacy page ('neither PostHog nor my visitors receive any raw IP address') ()

▶ Watch this taught:

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

Why is email the stack's one paid seam?

Deliverability. SMTP self-hosting means IP reputation, warming, and blocklist wars — the operational cost dwarfs the subscription. Composition (React Email) stays open; DELIVERY is bought.

The audience machine: capture loops and the permission line

One sarcastic viral tweet became 178,000 CRM contacts — and then he asks the room, on the record: 'do you guys think what I've done is in a gray area?'

The loops, all shown live: (1) SITE → EMAIL: the newsletter pop-up (whose absence he catches and hotfixes mid-session) feeds Resend; sources attributed (newsletter page vs pop-up vs footer) so the funnel is measurable. (2) SOCIAL → CRM: his viral tweet (mocking an AI job-application tool with a fictional auto-rejector) drew mass engagement; Apify harvested the engagers — handles, companies, emails — into Twenty: 178k people, 95k companies, lead-source tagged. (3) CRM → RELATIONSHIP: outreach framed as consent-first — 'you've interacted with my tweet... would you want to subscribe?' — 'I would always take permission before adding them into your newsletter.' (4) ANALYTICS → EVERYTHING: PostHog wired to see what each cohort does, with the planned PostHog→Twenty API integration closing the loop (signup lands in CRM automatically).

The ethics moment is deliberate pedagogy: he names the gray area, invites judgment, and lands on the distinction — harvesting public engagement for CONSENT-SEEKING outreach vs spamming: 'if you're using it in a non-spammy way... you are building relationships.' The motive is stated plainly too: domain authority, newsletter sponsors, brand.

Worked example · from the session

The tweet-to-CRM walk: the joke tweet shown, the engager list scrolled (a Mem0 employee among them), the Apify actor named as the harvest tool, the outreach script recited — the full pipeline, no curtain.

Why it matters

This is the operational answer to s23's 'become the source': authority needs an audience machine, and the machine's parts are all buildable.

People get this wrong

Audience building is content plus luck.

It's instrumented loops: every surface captures, every capture is attributed, every contact is consented — luck (virality) just feeds a machine that's already built.

I would always take permission before adding them into your newsletter or adding them into your ecosystem.

Building has become easy, but you have to talk about your idea, you have to bring it to the world.

For your projects

The Apify harvest is capability-knowledge for your KB; the consent framing is the part that transfers to your standards.

Go deeper

In one line: Capture loops: pop-up→Resend (source-attributed) · viral-content engagers→Apify→Twenty (tagged by lead source) · consent-first outreach converting engagement to subscription · PostHog behavioral overlay, with API wiring so signups land in the CRM. Boundary: permission before list addition; relationship framing over spam.

The pop-up outage caught BY teaching it — 'there are visitors on my site right now and they're not able to see the newsletter pop-up' → immediate agent hotfix ()

Apify as the X-engagement harvester; emails via the actor — capability stated without euphemism ()

Source attribution as the funnel's instrumentation: pop-up rising = the mechanism works ()

The prompt recipe for CRM wiring: Twenty API key → 'integrate this with my PostHog so newsletter signups are captured in my CRM' ()

Motive transparency: sponsors, domain authority, brand — 'I'm just showing you my entire playbook' ()

▶ Watch this taught:

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

What separates his engager-harvest from a spam operation, in his own framing?

The ask. Harvested contacts get a consent-seeking message referencing their own public engagement — the list membership is opt-in; only the OUTREACH uses the harvest. Whether that line holds is the gray area he names.

Newsletter craft: list health, open rates, and the value-add subject line

His own A/B result, read off his own dashboard: 'I went viral on X' underperformed. 'The Stripe for AI became Stripe' won. 'Nobody cares if I went viral on X.'

The operating numbers, from his live analytics: ~40% open rate on the best issue; his stated industry ladder — 30-35% = 'elite,' 25% good, 20% acceptable, below that 'remove the ones not reading and keep building.' LIST HEALTH is framed as domain protection: 'I don't want to burn my domain, because I spent a lot to acquire harshith.com' — engagement-pruning is deliverability insurance, not vanity. Cadence: Mon/Wed/Fri AI content + a Saturday personal note; email chosen as the channel that dodges 'being prey for the algorithm.'

The craft lesson is the subject-line autopsy: bragging ('I went viral on X') vs curiosity/value ('The Stripe for AI') — 'it always has to be a value add to the people.' Production doctrine: his system auto-drafts issues, he reviews and edits every one — 'I would recommend not to go full AI here... humanize it, pour in your thoughts.' Monetization target: sponsors, unlocked by list quality. Composition: React Email templates via Copy-for-AI prompts, brand-matched (the Velora newsletter one-shot on camera)."

Worked example · from the session

The two-issue comparison scrolled live: 254 opens vs 224-226, the winning subject's curiosity mechanics unpacked against his own flop — a practitioner grading his own homework in public.

Why it matters

It's the rare newsletter lesson with the author's real numbers attached — benchmarks, cadence, and the failure case, all verifiable on screen.

People get this wrong

A bigger list is a better list.

A healthier list is a better list — engagement rate protects the domain that delivers everything else, and pruning is how you keep it.

I don't want to burn my domain, because I spent a lot to acquire it... I have to make sure all of them are real people.

Nobody cares if I went viral on X. It always has to be a value add to the people.

For your projects

The 30-35%-elite benchmark and prune-below-20 rule are immediately quotable in your agency email work.

This idea elsewherebuilds onauthority's owned-audience arm

Go deeper

In one line: Doctrine: open-rate ladder (30-35% elite / 25% good / 20% floor; prune below); list health = domain protection; value-add/curiosity subjects over self-reference; AI-draft + mandatory human review; steady cadence (his: M/W/F + personal Saturday); email as algorithm-proof channel; sponsors as the monetization path; React Email + brand tokens for composition.

The pruning rule: sub-20% engagement segments get removed — deliverability compounds ()

'The best way to stay in touch with your audience without being a prey for the algorithm' — the channel thesis ()

Draft-review-send: automation writes, the human decides — stated as a floor, not a preference ()

Subject-line taxonomy from his own data: curiosity ('became Stripe') > utility > self-reference ()

His attribution dashboard tracks subscriber source per issue — craft measured, not felt ()

▶ Watch this taught:

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

Why does he treat low-engagement subscribers as a THREAT rather than a vanity cushion?

Mailbox providers score domains on engagement; dead weight drags the score and lands future issues in spam for everyone — the list's size is worth less than its pulse.

Every concept, three clicks deep

The same concepts as a quick reference: the closed row is the glance, open is the study card, and every timestamp jumps into the recording.

01The Marketing OS: sector-agnostic command center over swappable websitesMarketing OS = monorepo {apps/website (swappable, per-brand, interactive), apps/dashboard (permanent comman…›

Marketing OS = monorepo {apps/website (swappable, per-brand, interactive), apps/dashboard (permanent command center: analytics + CRM + newsletter pages integrating PostHog/Twenty/Resend)}; two Vercel deploys; integration-not-reimplementation; agent-readable by design.

Sector-agnostic split answered live to a cohort question — the OS ships once, websites rotate ()

Dashboard as 'one-stop shop... to look at everything and make informed decisions' — the mock-data version lands in minutes, integrations follow ()

Three.js/'3UI' as the consumer-site aesthetic lever ('SHADCN is for dashboards; 3UI is for interactive websites') ()

Deploy discipline: 'even with one Git repository, you deploy two apps on two separate Vercel projects' ()

The whole repo open-sourced to the cohort at session end — his standing pattern ()

02The handover document: firing an agent without losing the projectAgent-swap protocol: on misbehavior, (1) stop new work;›

Agent-swap protocol: on misbehavior, (1) stop new work; (2) demand a markdown handover document of all work done, written into the project folder; (3) verify it exists; (4) instruct the successor agent to read it first. State lives in files, not in the dying session. Consider debugging-in-place when time allows — swapping is triage, not doctrine.

The trigger symptoms named: unrequested tool use (Playwright), off-brief output, token burn without progress ()

The document must land IN the folder — 'one good thing it did was the handover document is inside the same folder' ()

Successor's first prompt: 'read the entire handover document' before any building ()

Same-family incident: the PostHog wizard vs Claude credentials — explicit do-not-touch instructions as guardrails ()

Model-market color: 'enterprises use a lot of Claude' (data security) vs his Codex speed preference; 'the first version of Fable was so good' ()

03The open-source marketing stack: PostHog, Twenty, React EmailStack: PostHog (product analytics + session replay;›

Stack: PostHog (product analytics + session replay; cloud free tier or VPS self-host; wizard/agent install; privacy disclosure required) · Twenty CRM (open source, self-host; infrastructure behind your own dashboard UI) · Resend (paid delivery — SMTP isn't self-hostable in practice) + React Email (open composition; Copy-for-AI templates). Heavy tools live on VPSes, not serverless.

PostHog per-source replay filters are the growth loop: watch exactly what ad-clickers or newsletter-readers do ()

Rage clicks as UX telemetry: 'the button has an issue... it could mean a lot of different things' ()

The n8n hosting analogy: PostHog and Twenty are VPS-class apps; Vercel won't hold them ()

Twenty's 7-day trial card-gate on cloud vs fully-free self-host — the recurring open-core shape ()

React Email's Copy-for-AI buttons: templates shipping their own integration prompts — a pattern to watch ()

PII posture: recordings mask emails/identities; declare analytics in the privacy page ('neither PostHog nor my visitors receive any raw IP address') ()

04The audience machine: capture loops and the permission lineCapture loops: pop-up→Resend (source-attributed) · viral-content engagers→Apify→Twenty (tagged by lead sour…›

Capture loops: pop-up→Resend (source-attributed) · viral-content engagers→Apify→Twenty (tagged by lead source) · consent-first outreach converting engagement to subscription · PostHog behavioral overlay, with API wiring so signups land in the CRM. Boundary: permission before list addition; relationship framing over spam.

The pop-up outage caught BY teaching it — 'there are visitors on my site right now and they're not able to see the newsletter pop-up' → immediate agent hotfix ()

Apify as the X-engagement harvester; emails via the actor — capability stated without euphemism ()

Source attribution as the funnel's instrumentation: pop-up rising = the mechanism works ()

The prompt recipe for CRM wiring: Twenty API key → 'integrate this with my PostHog so newsletter signups are captured in my CRM' ()

Motive transparency: sponsors, domain authority, brand — 'I'm just showing you my entire playbook' ()

Tools referenced

Tool	Coverage	Context
Codex	demonstrated	The recovering agent: reads the handover, one-shots the mini-site + Velora build, drives PostHog/dashboard integration
Claude Code	demonstrated	Opus 5 attempt — rogue Playwright, bad output, retired via handover document; enterprise-adoption color kept
Next.js	demonstrated	The monorepo substrate; two-app, two-Vercel-deploy shape
Vercel	demonstrated	Both apps deployed live (website + dashboard as separate projects)
GitHub	demonstrated	Repo pushed and flipped public for the cohort mid-session
Supabase	mentioned	Spotted configured in a cohort member's parallel build ('that's crazy that Supabase is also set up')
n8n	mentioned	The hosting analogy for PostHog-class self-hosted apps

Session materials

Archived locally on V: — click to open. Companion pages link to the LMS.

vtt

AICP C3 - Week 12 session 2 transcript.vtt

224 KB

Action items

Adopt the handover-document protocol as the standard agent-swap exit ()
Install PostHog (cloud free tier) on one property and watch ten session recordings — the fastest UX audit available ()
Evaluate Twenty CRM self-hosted as the agency CRM; keep its UI behind your own dashboard if needed ()
Stand up the minimal audience machine: pop-up → Resend → source attribution ()
Apply the open-rate ladder + prune rule to any list you touch ()
Steal the value-add subject test: would a stranger care, or is it about you? ()

Resources mentioned

Resources

docThe Marketing OS GitHub repo (made public for the cohort mid-session) + the Velora website/dashboard Vercel deploys
docThe stack links: posthog.com (+ its GitHub), twenty.com (+ GitHub, ~55k stars), resend.com, react.email templates, '3UI' component library (+ GitHub), Apify
docHis live infrastructure, shown as reference: insights.harshith.com (PostHog), crm.harshith.vc (Twenty), letter.harshith.com (newsletter dashboard), harshith.com/newsletter
docSession notes via Granola (his note-taker ran all session; notes promised into the resource pack)
docBusiness-idea asides: the $9 template-website marketplace; the deep-sea harness tease ('save that for some other day'); the caveman token-saving skill (100k stars, cohort-shared)

Extraction notes

This page was built from an auto-generated transcript, which garbles product and people's names. Those were corrected silently in everything above and logged here for transparency. The warnings flag claims that were true on the recording day but change fast.

Transcript corrections applied

The transcript says	The trainer actually means
3UI / 3JS / 3GS / Prejjs / pre-UI	Three.js and the 'threeui'-style component library (exact name unverified — check the shared notes)
20.com / 20CRM / 20 CRM	Twenty CRM (twenty.com)
postdoc / post hoc / posthoc / PostDog / postdog / Post Hog / host log	PostHog
Resend / recent (subscriber contexts)	Resend (resend.com)
React dot email	React Email (react.email)
crm.harshat.vc / insights.harshat.com / letter.harshat.com	*crm/insights/letter.harshith. (his subdomains as heard)**
Velora / Valora / Allura / well over	Velora (the demo skincare brand)
Appify / API (harvest contexts)	Apify (the scraping platform + its X-engagement actors)
Granola / granola / Canola	Granola (meeting note-taker)
Meatly	Meetily (open-source note-taker, as heard)
MemZero	Mem0
Ultra Code / Cloud log / Clark / plot	Claude Code (Opus 5; tier names as heard)
GPT image 2 / Image Gen 2 / GPT Image 1.5	OpenAI image models (the agent corrected him to the supported version live)
Subabase	Supabase
list monk	Listmonk (referenced on his privacy page as the newsletter service)
raj clicks / a raj clicks	rage clicks
Engle Light / Engle browser / EGO Browser Workflow	Ego Browser (his agentic browser; 'Ego Light' tier)
caveman skill (talking to a caveman, cut 65% of tokens)	(community token-compression skill, as heard — 100k stars claimed)
deep-sea hardness / deep-sea Cardness	DeepSeek harness (teased for a future session)
outbid.law	outbid.lol (carried over from s23)
viboded	vibe-coded
Fable 5 got banned / Claude has an attitude since Fable 5	(cohort chatter as heard — echoes the July Fable pull-forward episode in sprint-osp-d2)

True on recording day — verify before relying

The stack snapshot is launch-window fresh: Twenty's UI/pricing, PostHog's free-tier limits (~1M events), Resend tiers, and the '3UI' library will all drift.
The Claude-vs-Codex mood readings ('attitude since Fable 5', slower, pricier) are period sentiment, not benchmarks — but they triangulate with the sprint's subsidy-math session a day earlier.
The PostHog→Twenty CRM integration was prescribed as prompts, not demonstrated — verify the wiring before relying on it.
His open-rate ladder is stated practitioner lore, not a cited study.
The promised resource pack (notes, prompts, repo links) lands on the LMS 'tomorrow in the first half' — harvest it with the next scheduled run.