← AI Catalyst C3All programsHomeSearch
AI Catalyst C3·Core Sessions - Week 4·1:37:37

Office Hours IV: PostHog over Plausible, the Launch-Fast Loop, Scraping Ethics, Agent Swarms with Memory, and Why You Never Give Credits Away Free

Harshit Host - live demos of PostHog, scheduled agents vs repo agents, GPU hosting; the pricing doctrine · Niharika Outskill community - LMS logistics, closing announcement · Cody Participant - the persistent-intelligence system and the belief->outcome loop questions · Palak Participant - analytics, privacy policy, scraping ethics · Kabilan Participant - copycats, metering and pricing
▶no platform link on file
← PreviousSession 8: The Conversion Playbook
Lesson 3 of 3 in Core Sessions - Week 4
Next →Session 9: Introduction to Cursor & its principles

The short version

  1. Analytics: PostHog over Plausible - open source, self-hostable, session replay and heatmaps - run alongside Google Analytics and Search Console, and disclose it in a standard privacy policy; seven of 343 pages indexed after four days is normal (0:06-0:18).
  2. Cody's belief -> attention -> decision -> behaviour -> outcome loop: start from a niche problem you know, launch on Product Hunt even as a bare waitlist page, bug real users for feedback, instrument behaviour, iterate - only the market gives real signal (0:22-0:29). For his overloaded 'persistent intelligence' agent: split it into a swarm with Vercel's new Eve framework and give it mem0 as a memory layer - Obsidian is a knowledge graph, not agent memory (0:35-0:42).
  3. Scraping ethics: everyone scrapes - even Cloudflare, which blocks scrapers, sells a scraping API. The working compromise: public marketing pages fully crawlable for SEO and GEO, proprietary logged-in content locked down (0:50-0:59).
  4. Pricing doctrine for AI products: no free tier ('only hackers abuse free credits'), flat-plus-credits pricing, cost-plus markup on API + GPU + hosting, negotiate vendor deals early while VC money subsidises them, and think 'service as software' - an agency with proprietary tooling on retainer - not pure SaaS (1:05-1:21).
  5. Model landscape snapshot: GLM 5.2 (Z.ai) near parity on benchmarks, weak at coding, very cheap, use as a fallback; OpenRouter to compare; OpenCode to run any model in a Claude-Code-style harness; Codex + GPT-5.5 extra-high runs unattended for hours (0:29-1:04). Orchestration is boring GitHub Actions YAML plus a Slack 'chief of staff' channel (1:32-1:33).

At a glance, three clicks deep

Skim here first: the closed row is the glance, open is the study card with the key points and timestamps, and the ↓ link drops to that concept's full write-up below.

01PostHog, not Plausible - and disclose it in the privacy policyPostHog for product analytics + GA/Search Console for search signal + a standard privacy disclosure.›

PostHog for product analytics + GA/Search Console for search signal + a standard privacy disclosure.

PostHog: open source, self-host, replay, heatmaps (0:10-0:12)

Run GA + Search Console alongside (0:17-0:18)

Privacy-policy disclosure is standard and required (0:12-0:15)

↓ Full write-up of this concept

02Belief, attention, decision, behaviour, outcome: launch first, then measureNiche problem -> ship a minimal page -> Product Hunt -> user feedback -> behavioural analytics -> iterate.›

Niche problem -> ship a minimal page -> Product Hunt -> user feedback -> behavioural analytics -> iterate.

Start at niche x expertise (0:23-0:24)

Launch even a bare waitlist on Product Hunt (0:25-0:27)

Instrument behaviour; only the market gives signal (0:28-0:29)

↓ Full write-up of this concept

03Scraping ethics: crawlable marketing pages, locked proprietary contentPublic pages open to crawlers (SEO/GEO), private content behind auth;›

Public pages open to crawlers (SEO/GEO), private content behind auth; scraping others is a grey area everyone practises.

Even Cloudflare sells scraping (0:53-0:55)

Public marketing pages: crawlable (0:56-0:57)

Proprietary / logged-in content: locked (0:57-0:59)

↓ Full write-up of this concept

04Split the overloaded agent: Eve for the swarm, mem0 for memory, ponytail for the junkSwarm (Eve) + persistent cross-model memory (mem0) + code condensing (ponytail);›

Swarm (Eve) + persistent cross-model memory (mem0) + code condensing (ponytail); Obsidian is not agent memory.

Ponytail: paste the repo, 'install this for my repo' (0:33-0:34)

Eve: specialised sub-agents over one big one (0:37-0:40)

mem0 persists across GLM <-> Gemini switches (0:40-0:41)

Obsidian = knowledge graph, not memory (0:41)

↓ Full write-up of this concept

05Service as software: never free, flat plus credits, cost-plus, negotiate vendors earlyNo free tier;›

No free tier; flat or credit pricing at cost-plus; vendor deals early; position as an agency with tooling.

Moat = distribution + unscrapable internal workflow (1:07-1:10)

No free access; charge from day one (1:12-1:15)

Flat monthly or credits (~1,000 = $20) (1:13)

Vendor discounts now, VC-subsidised (1:18-1:21)

'Service as software' (1:16-1:17)

↓ Full write-up of this concept

06Prove ROI with your own endpoints and an LLM-drawn dashboardLog activity via your own endpoints -> LLM builds the client dashboard -> ROI is visible, not argued.›

Log activity via your own endpoints -> LLM builds the client dashboard -> ROI is visible, not argued.

Endpoints per activity, LLM-generated dashboards (1:23-1:26)

Endpoints via plain-English prompts to a coding agent (1:35-1:36)

Orchestration: GitHub Actions YAML + a Slack chief-of-staff channel (1:32-1:33)

↓ Full write-up of this concept

The concepts in full

01

PostHog, not Plausible - and disclose it in the privacy policy

A Lovable audit told Palak to install Plausible. Harshit screen-shared his own dashboard and showed her why not.

PostHog: fully open source, self-hostable, session replay, heatmaps, cross-device tracking. Keep Google Analytics and Search Console running too - the Google pixel helps ranking - and put a privacy-policy page up that says you track; GDPR requires it and 'every platform does it', so it does not scare users. Slow indexing (seven of 343 pages after four days) is normal and unrelated to your analytics choice.

Why it matters

The analytics stack recommended for every site in the cohort, including client sites.

This idea elsewhereextendsWhat to measure once Session 3's landing page is live
02

Belief, attention, decision, behaviour, outcome: launch first, then measure

Cody wanted a framework for turning any idea into a validated product. Harshit's is a loop, not a plan.

Pick a problem at the intersection of a niche and your own expertise. Ship immediately - a landing page and waitlist is enough - and put it on Product Hunt, which Harshit checks daily for inspiration. Bug real users for feedback, instrument what they actually do with PostHog (clicks, heatmaps, errors), and iterate. The belief and attention stages are marketing; decision, behaviour and outcome only exist once something is live.

Why it matters

The operating rhythm the host repeats in every office hour: nothing is validated until strangers use it.

This idea elsewherereappears inSession 25 turns 'launch before ready' into a playbookcontinuesOH V: when building gets easy, selling is what's left
03

Scraping ethics: crawlable marketing pages, locked proprietary content

Palak: LinkedIn bans scraping, I blocked scrapers on my own site, and you teach Firecrawl. Isn't that hypocritical?

Harshit's answer is that scraping is universal - Cloudflare blocks scrapers and sells a paid scraping API - and the law is a grey area everyone operates inside. The practical rule is two-tier: your public marketing pages must be fully crawlable, because that is how search engines and AI answer engines find you; your proprietary, logged-in content is what you protect. Blocking everything costs you SEO and GEO for no gain.

Why it matters

Resolves a tension every builder in the cohort hits, and sets the crawlability rule GEO depends on.

This idea elsewhereparallelSession 23: what crawlers cannot see, they cannot cite
04

Split the overloaded agent: Eve for the swarm, mem0 for memory, ponytail for the junk

Cody's single agent did everything - and forgot everything each time he switched from GLM to Gemini.

Three fixes. Ponytail (a GitHub repo you install by pasting the link) keeps the codebase crisp and cuts token bloat. Eve, Vercel's agent framework launched hours before the call, builds a swarm of specialised sub-agents instead of one overloaded one. And mem0 is an open-source memory layer that persists across sessions, agents and model switches - which Obsidian, being a knowledge graph rather than a memory store, does not. Then ship the landing page anyway.

Why it matters

The three-part architecture answer to 'my agent forgets and accumulates junk'.

This idea elsewherereappears inOH XIII: ponytail as a DIY code reviewer toocontrasts withOH III recommended Obsidian for chat backup - different job
05

Service as software: never free, flat plus credits, cost-plus, negotiate vendors early

Kabilan asked how to stop copycats, how to meter a photo-to-video feature, and whether 'first 100 free' is right. Three no's.

Ideas are copies of copies; the only moat is distribution (LinkedIn/Twitter presence, a specific ICP) and the internal workflow nobody can scrape. Meter by cost: API + GPU + hosting + markup. Never offer free access - 'only hackers abuse free credits, creating new accounts' - charge from day one with a flat monthly plan or a credit pack (1,000 credits ~ $20, adjustable). Negotiate discounts with vendors like ElevenLabs now, while VC funding subsidises their prices. And frame the product as service as software: an agency with proprietary tooling on retainer, not a pure SaaS.

Why it matters

The commercial doctrine behind the agency model; stated most completely here.

This idea elsewhereparallelSession 1's pricing stance, applied to a productparallelCohn's no-free-work rule for services
06

Prove ROI with your own endpoints and an LLM-drawn dashboard

Bernard: how do I prove to a client that the AI videos I post actually move their numbers?

Build small custom API endpoints (/post, /newsletter, /signup) that log every activity, feed the data to an LLM, and have it generate real-time dashboards and graphs for the client. Custom endpoints themselves are a plain-English request to Claude Code or Codex against your repo. The point is a live artefact the client can open, instead of an anecdote.

Why it matters

The retention tool for any recurring client engagement.

This idea elsewhereappliesMake Session 1's revenue / time / cost lever visible on a dashboard

Tools referenced

ToolCoverageMomentContext
PostHogdemonstratedLive on the host's own dashboard
Product HuntexplainedLaunch and daily inspiration
GLMexplainedGLM 5.2: cheap, near parity, weak at coding - fallback
PonytailexplainedCode condensing, install by repo link
EveexplainedVercel's agent-swarm framework, launched that day
mem0explainedPersistent cross-agent memory
Hugging FaceexplainedModel weights; training cost scale
PlausiblementionedNot recommended over PostHog
Google Search ConsolementionedIndexing takes time
OpenRoutermentionedCompare benchmarks, pricing, latency
ObsidianmentionedKnowledge graph, not agent memory
CodexmentionedGPT-5.5 extra-high, unattended for hours
OpenCodementionedRun any open model in a Claude-Code-style harness
OllamamentionedLocal small models, start with Gemma
RunPodmentionedGPU rental for self-hosted LLMs
InstantlymentionedPre-warmed outreach accounts
VAPImentionedVoice calling API
ApifymentionedLinkedIn scraper actor
GitHub ActionsmentionedAgent orchestration via YAML
ResendmentionedProgrammatic email for a SaaS

Action items

    Resources mentioned

    Resources
    • docCold-outreach infrastructure answers
    • docAnnouncements

    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 saysThe trainer actually means
    Harshish / Harshid / Arsha / ArjitHarshit
    GlorLovable (probable)
    resender.comResend
    Balak / Pollock / Dolphy / PalakkirPalak
    Kavilan / Babylon / Double onKabilan
    ActionAchin

    True on recording day — verify before relying

    ← PreviousSession 8: The Conversion PlaybookNext →Session 9: Introduction to Cursor & its principles