← All sessionsHomeSearch
AI Sprints (Live Weekend Programs)·AI Masterclass I·2:54:41

AI Masterclass: Exploring Antigravity 2 — Four Products, Four Slash Commands, and Two Live Builds

Prakash Gupta AI operations specialist — builds AI infrastructure for businesses and trains their teams; '25 businesses, 150+ agents and automations in 3 years, since ChatGPT-3 came out' · Arfa Program coordinator/host — audio-only ('network glitches'); frames this as the first AI Masterclass exclusively for Outskill alumni
▶local recording  ·  LMS lesson

Session map

THE LANDSCAPETHE MACHINERYTHE CRAFTOne brand, four productsIDE, CLI, SDK, and the chat-first Antig…Gemini 3.5 Flashpaying for speed, and the quota-tier re…/grillme, /goal, /schedule, /…the built-in loop kitSkills, hooks, and the 93-age…Claude Code's vocabulary, Google's impl…The business-dashboard arcplan review, the purple problem, and th…
The landscapeThe machineryThe craft
click a node — its card pops up (drag it anywhere, × to close)
Concept

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

The short version

  1. Antigravity is now FOUR products, and the naming confuses everyone: the IDE (what people used as 'Antigravity 1' — a VS Code fork, Cursor-like), the CLI (Claude Code's shape on Gemini; notable because it USED to be open source), the SDK, and ANTIGRAVITY 2 — 'not a code editor': chat-first, files hidden under layers, 'made for people who don't want to dive technical — they just want the output.' The industry read: both Google and Cursor are converging from file-based to chat-based surfaces.
  2. The four slash commands are the session's real payload: /GRILLME turns a one-line idea into a questionnaire that builds the context ('use it every time — if the AI doesn't know exactly what you want, it can't deliver it'); /GOAL keeps the agent working until a verified condition ('until the website is working'); /SCHEDULE registers CLOUD-side crons that run with your laptop closed — the quiet differentiator; /BROWSER drives Chrome via the DevTools MCP (logs in, clicks, screenshots, verifies the build's own UI).
  3. Loop machinery arrives productized: skills carry the front-matter/name/description pattern from Claude Code (create one by pasting a spec — or a YouTube transcript — into chat); HOOKS are deterministic code on events (pre/post tool use, stop) that spend NO tokens — including as MCP security screens; the agent swarm runs up to 93 sub-agents coordinated through a shared task.md.
  4. The dashboard build teaches the honest arc: grillme → plan review ('finalize the plan before implementing — after implementation, even a small change costs a lot of tokens') → goal → a working app whose first draft 'looks very AI — it is very purple' → rebrand, real transactions, AUTH, and the privacy move worth keeping: an import-CSV button so financial data populates the dashboard WITHOUT being given to the LLM.
  5. The founder-briefing agent closes the loop lineage: Gmail MCP + Tavily + a 9-AM scheduled task = the same morning-briefing pattern as the Hermes sprint, now on Google's cloud scheduler — ask for the PROMPT, then paste it into the scheduler yourself.
  6. Routing doctrine from the Q&A: plan with Claude (Opus), implement with Gemini 3.5 Flash for speed; Antigravity 2 'can replace Claude Code' but quality differs by task; his mockup-vs-production claim ('Lovable and Replit are only good for mockups') is a partisan take worth flagging, not a fact.

The concepts

01

One brand, four products: IDE, CLI, SDK, and the chat-first Antigravity 2

'Before, we used to have just one Antigravity. Now we have four — and every piece serves a different purpose.'

The map: the IDE is 'Antigravity 1' — a VS Code fork, 'very much similar to Cursor,' file tree and chat window, full code access. The CLI is 'exactly what Claude Code is, but on Antigravity and Gemini' — terminal-resident, edits files, builds projects; its differentiator was being open source ('a lot of people built apps on top of it, which Anthropic did not allow' — though 'now it is not' open). The SDK is for developers building on top. ANTIGRAVITY 2 is the star: 'this is NOT a code editor' — chat-first, minimal clutter, files hidden 'under a few layers,' aimed at 'people who do not want to dive too much technical — they just want to get the output.' One agent at a time in the IDE; multi-agent swarms in 2. Logos: black background = IDE, white = 2.

The positioning insight worth keeping: 'Both Google and Cursor are diving deeper into software focusing on the chat-based approach rather than the file-based approach' — the IDE generation and the chat generation are diverging products for diverging users, and the limitation is named honestly: 'you are losing control; the LLM makes most decisions on your behalf... but most people don't want that control.'

Worked example · from the session

The category boundaries drawn in Q&A: vs Hermes/OpenClaw ('those are autonomous agents living on a cloud; this is manual control with an AI helping hand'), vs models ('Opus is a model; this is a platform'), vs Claude Code ('competitors — you can't use one inside the other, but you can use each other's models').

Why it matters

It's the cleanest snapshot of Google's agent-surface strategy in the corpus — and the naming decoder every learner needed.

People get this wrong

Antigravity 2 is an upgraded version of the Antigravity IDE.

It's a different product for a different user — chat-first with hidden files vs a code editor; they coexist on one account.

One brand, four products IDE the editor CLI the terminal agent SDK build on it Chat the front door /grillme · /goal · /schedule · /browser — the built-in loop kit Antigravity is a family, not an app — know which product you're in before comparing it to anything
One brand, four products: IDE, CLI, SDK, chat, plus the slash-command loop kit
It is supposed to be made for people who do not want to dive too much technical. They just want to get the output, and they're happy with it.
For your projects

The 'losing control is the limitation — and most people want that' line is the honest framing for recommending chat-first tools to non-technical clients.

This idea elsewherecomplementsthe Catalyst track's Claude/Codex/Antigravity market read, from the Google side
Go deeper

In one line: Antigravity family = IDE (file-based VS Code fork), CLI (terminal agent, Gemini-powered), SDK (developer layer), Antigravity 2 (chat-first no-editor surface with swarms); one account spans all; chat-first trades control for accessibility.

The open-source contrast drawn explicitly: 'Anthropic is a very secretive company... Google wants to push open source' — with the CLI's own closure as the wrinkle ()

Code access in 2 isn't gone, just buried — 'hidden under a few layers... to keep things minimalistic' ()

Fifteen days old at recording: 'give it time to perform in the market' ()

▶ Watch this taught:

Check yourself

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

Who should stay on the IDE rather than move to 2?

Anyone who needs to touch code directly — the IDE keeps file-level control; 2 optimizes for people who evaluate outputs, not diffs.

02

/grillme, /goal, /schedule, /browser: the built-in loop kit

'I never thought about the tech stack, the features, or the UI. But to get the right application, I need to answer these questions.' /grillme asks them for you.

The four commands compose into a development loop. /GRILLME: give a one-liner ('I want to build a to-do application'), get a structured questionnaire back — framework, feature multi-selects, aesthetic, layout, interaction details — 'helping me finalize the right questions, giving Antigravity enough context to build what I really want.' His doctrine: use it EVERY time, with any agent. /GOAL: state the done-condition ('follow the implementation plan until the website is working'; 'the application must run on mobile') and the agent 'keeps working, keeps refining, without you manually asking, until the goal is verified' — the evaluator loop as a slash command. /SCHEDULE: natural-language crons ('at 8:21, say hi'; hourly gold prices via an MCP) that run ON ANTIGRAVITY'S CLOUD — 'it will execute even when your system is not working,' the differentiator over laptop-bound schedulers. /BROWSER: Chrome automation through the Chrome DevTools MCP (one-time remote-debugging enable), which logs into the freshly-built dashboard, searches, adds and completes tasks, takes screenshots, and attempts session recordings — 'I was not doing anything while this took place.'

The command the cohort votes most useful: split across schedule, grillme, and browser — correctly, because they're stages of one loop: context in, goal held, schedule sustained, browser verified.

Worked example · from the session

The to-do app demo: grillme's five questions (vanilla JS, all features, frosted-glass theme, dashboard layout, overlay editing) → implementation plan → /goal → a working localhost app with Kanban board, progress tracking, and category management.

Why it matters

This is the Hermes sprint's loop discipline shipped as product affordances — evaluator, interviewer, cron, and verifier, each one slash away.

People get this wrong

/goal is a stronger prompt.

It's a loop contract: a verifiable end-condition plus permission to iterate — the separate-evaluator pattern from loop engineering, productized.

I never thought of what tech stack, how many features, what the UI should be. But to get the right application, I need to answer these questions.
It will be scheduled on the cloud. It will be executed even when your system is not working.
For your projects

Your AskUserQuestion habit is /grillme from the other side of the table — the parallel is worth showing clients.

This idea elsewherecomplementsthe loop discipline this kit productizescomplementsthe same morning-cron pattern, cloud-hosted instead of laptop-bound
Go deeper

In one line: The Antigravity 2 loop kit = /grillme (context interview) → plan review → /goal (work-until-verified) → /schedule (cloud-side cron persistence) → /browser (DevTools-MCP verification); each command productizes a loop stage that other stacks hand-build.

Schedule-with-browser caveat: browser automation needs the machine on; pure schedules don't ()

/browser generalizes beyond your own app: 'check emails, upload a post to Instagram — any website,' Chrome-only, riding logged-in sessions ()

The 'always allow' decision framed honestly: 'you're letting Gemini run commands without your permission — depends on how much you trust the LLM' ()

Grill-first is his answer to 'best workflow from thought to app': right context → right app ()

▶ Watch this taught:

Check yourself

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

Which command replaces the plan-approval gate from the Hermes delegation loop?

None alone — grillme builds the context and the implementation plan appears for review, but YOU still hold the approve step before /goal starts the run; the gate is a habit, not a feature.

03

Skills, hooks, and the 93-agent swarm: Claude Code's vocabulary, Google's implementation

'Hooks are pieces of code that run on events — WITHOUT utilizing AI tokens.' The deterministic layer under the probabilistic one.

SKILLS carry the exact Claude Code anatomy — front-matter (name, description, tools) that tells the agent WHICH skill to invoke, body that says HOW — and are created by pasting a spec into chat ('create a new skill, name it SEO blog 2' → file written into the global skills folder, visible under customizations). His skill-creation technique travels well: find a long expert YouTube video on the topic, pull the transcript, and distill it into the skill — 'a professional overview of how the skill should perform.' HOOKS are the counterpoint: predefined CODE bound to events — pre-tool-use, post-tool-use, pre-compact, user-prompt, stop — 'not prompts; they don't have prompts.' Uses: auto-format after edits, Slack/email notifications on stop, and the security answer he gives twice: a pre-tool-use hook that screens MCP calls for leaked personal data. Zero tokens, deterministic, the guardrail layer.

The SWARM: up to 93 sub-agents under one main agent, each 'an individual Antigravity' with full features and its own context, coordinated through a shared TASK.MD that every agent updates — 'that's how they maintain consistency, and ideally multiple agents should not work on a single task.' The research→review→design→frontend→test decomposition mirrors the Hermes delegation loop natively.

Worked example · from the session

The hook taxonomy answered live against confusion: 'hooks are not skills (skills are prompts, hooks are code), not schedulers (schedulers are time-based, hooks are event-based), and they run on your machine.'

Why it matters

It's the fullest picture in the corpus of the three-layer agent stack: probabilistic skills, deterministic hooks, coordinated swarms.

People get this wrong

Everything an agent does should be a skill.

Skills for judgment, hooks for invariants, schedules for time — pushing invariants into skills makes them expensive and unreliable.

Hooks are supposed to do that without utilizing AI tokens... skills are prompts; hooks are pieces of code.
Whenever any agent makes a change, they update that task.md file — that's how they maintain consistency.
For your projects

The transcript-to-skill technique is literally this project's method — his version of what these session records do.

This idea elsewherecomplementsthe same orchestrator/sub-agent pattern, native with a 93-agent ceilingbuilds onthe Catalyst harness theory, visible in a second vendor's product
Go deeper

In one line: Antigravity 2's agent stack = skills (prompt capabilities with front-matter routing), hooks (token-free event-bound code for formatting, notification, and security screening), and sub-agent swarms (≤93, full-featured, task.md-coordinated, non-overlapping tasks).

The YouTube-transcript-to-skill pipeline is a general knowledge-capture trick, not an Antigravity feature ()

Hooks as MCP security is his repeated answer to 'how do I trust third-party MCPs' ()

Swarm consistency = a shared markdown file — the md-file doctrine again, now between machine peers ()

Mic input pitched as context economics: 150 wpm spoken vs 40-50 typed — more context per minute ()

▶ Watch this taught:

Check yourself

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

A check must NEVER be skipped and must cost nothing — skill or hook?

Hook. Skills are invoked probabilistically and burn tokens; hooks fire deterministically on the event, free.

04

The business-dashboard arc: plan review, the purple problem, and the CSV privacy pattern

'The UI looks very AI — it is very purple. Anyone can see this is created by AI. I don't want that.'

The build ritual, run end to end: /grillme the dashboard (finances, employees, projects; glassmorphism; Kanban; persistence) → REVIEW THE PLAN before proceeding ('very important... after implementing, even a small change can cause a lot of tokens as well as time') → /goal to execute → a working first draft in minutes — then the honest middle: the output 'looks like an AI-generated application,' changes silently fail ('I cannot see any changes... I think something is broken'), and recovery is part of the pedagogy: 'things may break, things may fail, and we're gonna recover — that's the ideology of the build session.' Iterations land the rebrand (Nebula → Outskill), a Google-flavored UI, admin AUTH (email/password, sign-in verified), and /browser as the QA pass — the agent logs into its own app and exercises it.

The keeper is the PRIVACY PATTERN: instead of pasting financials into chat, he builds an IMPORT-CSV button — 'the approach was NOT to give the Excel sheet or the financial data to an LLM. The import adds the data into the dashboard without giving it to the LLM. Your information is still protected.' Structure designed by the model; data flows around it. The second build (founder briefing: Gmail MCP + Tavily + 9-AM schedule) adds the finishing move — ask the agent for the PROMPT, then paste it into the scheduler yourself, 'leveraging schedule without Antigravity doing it itself.'

Worked example · from the session

The AI-tell checklist implied by his fixes: default purple, dummy 'Nebula' branding, mock data, no auth — the four things separating a demo from something a business would show.

Why it matters

It's the masterclass's practice core: the same grill→plan→goal→verify loop under real conditions, plus the one architectural idea (data around the model, not through it) with legs far beyond dashboards.

People get this wrong

Using an AI builder means your data goes through the AI.

The model can design the pipe without drinking from it — structured import paths keep data local while AI builds everything around it.

Finalize the plan before you implement. After implementing, even a small change can cost a lot of tokens as well as a lot of time.
The UI looks very AI — it is very purple. Anyone who sees it can see this is created by AI.
The approach was not to give the financial data to an LLM. The import-CSV button adds the data to the dashboard without giving it to the LLM.
Things may break. Things may fail. But we're gonna recover from that as well. That's the entire ideology behind a build session.
For your projects

The 'very purple' observation is a useful tell for your site-health audits — default-palette AI builds signal nobody did the de-AI pass.

This idea elsewherecomplementsthe same data-custody instinct, applied at the build-tool layercomplementsplan-with-Opus / build-with-Flash is the same judgment-vs-bulk split
Go deeper

In one line: Dashboard doctrine = grillme → plan review (cheapest change point) → goal-driven build → de-AI the surface (brand, palette, real data, auth) → browser-verified QA; sensitive data enters via structured import paths the LLM designs but never reads.

Plan review as cost control is stated in token terms, not process terms — the economics make the habit stick ()

Silent no-op changes are a real failure mode; verify with your own eyes or /browser, not the agent's claim ()

Claude-to-plan / Gemini-to-build routing offered as the standing division of labor ()

His 'Lovable and Replit are only good for mockups' is a partisan claim the corpus contradicts elsewhere — keep the CSV pattern, discount the verdict ()

▶ Watch this taught:

Check yourself

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

Where else does the import-CSV pattern apply verbatim?

Anywhere clients balk at pasting data into AI: the model builds the schema, parser, and UI; the sensitive rows flow through code paths it wrote but never sees — payroll, patient intake, donor lists.

05

Gemini 3.5 Flash: paying for speed, and the quota-tier reality

'People may confuse this — since it's a Flash model, it must be cheaper. It isn't. What it's good at is SPEED.'

The model positioning, as-heard: 3-3.5x faster than Claude Opus 4.7, twice as fast as Gemini 3; priced at ~$1.50/M input and $9/M output — '3x the price of the previous Flash, but still worth it... still cheaper and faster than Claude.' The thesis is 'speed WITH intelligence': the felt difference is response latency ('Claude Code spins for minutes; you notice immediately'). Access is quota-tiered: free = 20 requests/day; AI Pro ($20) and Ultra ($200) buy bandwidth, with per-model quotas and rolling refresh windows visible in settings — and Claude models are usable inside Antigravity WITHOUT your own Anthropic key, on the plan's quota. His routing rule completes it: Opus for planning and intelligence-bound judgment, Flash for implementation speed — 'compare cost, speed, and intelligence; make that decision.'

Worked example · from the session

The Q&A benchmark summary as-heard: 'Gemini 3.5 Flash gives better results than Sonnet; Opus 4.7 is still the smarter model' — a working hierarchy, not a lab result.

Why it matters

It's the third pricing lens in the sprint track: OSP priced WHERE to buy, Hermes D2 priced WHICH tier, this prices SPEED as its own axis.

People get this wrong

Flash tiers are the budget option.

This Flash is a premium SPEED product — costlier than its predecessor, cheaper than frontier, and chosen for latency, not price.

Since this is a Flash model, people assume it's cheaper. It isn't. What it is good at is the speed.
For your projects

Your own burn strategy already routes by judgment-density; this adds the third column (latency) for any future interactive tooling.

This idea elsewherecomplementsthe 80-percent-cost lens; this adds the latency lenscomplementsthe sprint track's pricing triptych completes
Go deeper

In one line: Flash economics = speed as a purchasable dimension distinct from intelligence and cost; quota-tiered plans (free/Pro/Ultra) with per-model windows; route planning to the smartest model and implementation to the fastest adequate one.

Speed changes behavior, not just bills — instant iterations keep the human in flow during builds ()

Bundled Claude access on Google's quota is a quiet cross-vendor subsidy worth noticing ()

All numbers are as-heard launch-week figures — decayed by definition ()

▶ Watch this taught:

Check yourself

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

When does the 3x-faster model lose to the smarter one even for implementation?

When rework dominates: if Flash's drafts need Opus-level correction passes, the latency win is eaten by iteration count — speed only pays on tasks the model clears in one or two passes.

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.

01One brand, four products: IDE, CLI, SDK, and the chat-first Antigravity 2Antigravity family = IDE (file-based VS Code fork), CLI (terminal agent, Gemini-powered), SDK (developer la…›

Antigravity family = IDE (file-based VS Code fork), CLI (terminal agent, Gemini-powered), SDK (developer layer), Antigravity 2 (chat-first no-editor surface with swarms); one account spans all; chat-first trades control for accessibility.

The open-source contrast drawn explicitly: 'Anthropic is a very secretive company... Google wants to push open source' — with the CLI's own closure as the wrinkle ()

Code access in 2 isn't gone, just buried — 'hidden under a few layers... to keep things minimalistic' ()

Fifteen days old at recording: 'give it time to perform in the market' ()

02/grillme, /goal, /schedule, /browser: the built-in loop kitThe Antigravity 2 loop kit = /grillme (context interview) → plan review → /goal (work-until-verified) → /sc…›

The Antigravity 2 loop kit = /grillme (context interview) → plan review → /goal (work-until-verified) → /schedule (cloud-side cron persistence) → /browser (DevTools-MCP verification); each command productizes a loop stage that other stacks hand-build.

Schedule-with-browser caveat: browser automation needs the machine on; pure schedules don't ()

/browser generalizes beyond your own app: 'check emails, upload a post to Instagram — any website,' Chrome-only, riding logged-in sessions ()

The 'always allow' decision framed honestly: 'you're letting Gemini run commands without your permission — depends on how much you trust the LLM' ()

Grill-first is his answer to 'best workflow from thought to app': right context → right app ()

03Skills, hooks, and the 93-agent swarm: Claude Code's vocabulary, Google's implementationAntigravity 2's agent stack = skills (prompt capabilities with front-matter routing), hooks (token-free eve…›

Antigravity 2's agent stack = skills (prompt capabilities with front-matter routing), hooks (token-free event-bound code for formatting, notification, and security screening), and sub-agent swarms (≤93, full-featured, task.md-coordinated, non-overlapping tasks).

The YouTube-transcript-to-skill pipeline is a general knowledge-capture trick, not an Antigravity feature ()

Hooks as MCP security is his repeated answer to 'how do I trust third-party MCPs' ()

Swarm consistency = a shared markdown file — the md-file doctrine again, now between machine peers ()

Mic input pitched as context economics: 150 wpm spoken vs 40-50 typed — more context per minute ()

04The business-dashboard arc: plan review, the purple problem, and the CSV privacy patternDashboard doctrine = grillme → plan review (cheapest change point) → goal-driven build → de-AI the surface…›

Dashboard doctrine = grillme → plan review (cheapest change point) → goal-driven build → de-AI the surface (brand, palette, real data, auth) → browser-verified QA; sensitive data enters via structured import paths the LLM designs but never reads.

Plan review as cost control is stated in token terms, not process terms — the economics make the habit stick ()

Silent no-op changes are a real failure mode; verify with your own eyes or /browser, not the agent's claim ()

Claude-to-plan / Gemini-to-build routing offered as the standing division of labor ()

His 'Lovable and Replit are only good for mockups' is a partisan claim the corpus contradicts elsewhere — keep the CSV pattern, discount the verdict ()

05Gemini 3.5 Flash: paying for speed, and the quota-tier realityFlash economics = speed as a purchasable dimension distinct from intelligence and cost;›

Flash economics = speed as a purchasable dimension distinct from intelligence and cost; quota-tiered plans (free/Pro/Ultra) with per-model windows; route planning to the smartest model and implementation to the fastest adequate one.

Speed changes behavior, not just bills — instant iterations keep the human in flow during builds ()

Bundled Claude access on Google's quota is a quiet cross-vendor subsidy worth noticing ()

All numbers are as-heard launch-week figures — decayed by definition ()

Tools referenced

ToolCoverageMomentContext
AntigravitydemonstratedAll four products toured; Antigravity 2 is the build surface: slash commands, skills, hooks, swarm, scheduler, browser
Geminidemonstrated3.5 Flash as the launch model — speed positioning, quota tiers, bundled Claude access
Claude CodedemonstratedThe constant comparison object — opened live in terminal; 'plan with Claude, build with Antigravity'
Chrome DevTools MCPdemonstratedThe /browser engine: remote debugging enabled once, then autonomous login/click/screenshot QA
TavilydemonstratedAs-heard 'Tabley' — news and gold-price fetching MCP feeding the founder briefing
NotionmentionedAdded as a connector via API key
Hermes agent (Nous Research)mentionedCategory boundary: 'totally autonomous, lives on a cloud — this is manual control with an AI helping hand'
OpenClawmentionedSame autonomous-agent category contrast; swarm comparison later
Hostinger / SiteGroundmentionedNamed with Vercel as the hosting answer for Antigravity-built apps
VercelmentionedThe other named hosting lane
Supabasementioned'Antigravity can handle Supabase — give it the information' (as-heard 'Superbase')
LovablementionedHis 'only good for mockups' claim — flagged as partisan; the corpus's Lovable projects say otherwise
ReplitmentionedBundled into the same mockup claim
n8nmentioned'If you want the exact deterministic approach, use n8n' — the routing rule again
GitHubmentionedThe team-collaboration answer: shared repo, since Antigravity has no teams mode

Action items

Resources mentioned

Resources
  • docAntigravity downloads + documentation
  • docGoogle Cloud Console OAuth flow
  • docSkills repository link
  • docTrainer's LinkedIn (Prakash Gupta)
  • docRecording location

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
anti gravity / handicraft video 0 / anti Clarity / AG 2 / a g 2.2 / 2.2Antigravity / Antigravity 2
cloud code / plot code / floor code / CloudPortal / cloud routineClaude Code (and once, likely Claude's scheduled 'routines' feature, as-heard)
Gemini 3.5 Flash / 3.53 0.5 flashGemini 3.5 Flash (as-heard version)
cloud opus 4.7 / OBS 4.8 / SONNET 4.6as-heard Claude model versions at recording
Tabley / Tableau / Family / travelingTavily (search MCP — as-heard; verify)
grill me / krill me / drill me / GrillMe/grillme command
Charge 3 / Charge g p d / Chargebee / charge GPTRChatGPT (3 / GPT wrappers)
Versus CodeVS Code
OpenClone / OpenCLOOpenClaw
Superbase / SuperBaseSupabase
Versal / wholesaleVercel
MCB / NCP / MCPLMCP
Nebula businessthe dashboard's AI-generated placeholder brand name
paperclip or anti gravity 2.2 orgarbled comparison question — unresolvable
task dot m dtask.md (the swarm's shared coordination file)
93 sub agentsas-heard swarm limit — verify against current docs
Arfa / Alpha / R 5 / ArafaArfa (host)
small group (startup name)as-heard demo startup name in the founder-briefing setup

True on recording day — verify before relying