Session 27: Lead Magnets as Proof Artifacts — the Mega-Prompt Build, Four Live Magnets, and the Skill That Packages the Workflow

Drawn live in Excalidraw: product or tool -> distribution channels (LinkedIn, Twitter/X, the email list, Product Hunt) -> lead magnet -> conversion to a paid or freemium product. The magnet is 'something of value, that helps your ICP' - 'you're just giving it out for free for your users, or for your network, so that they can just find value out of it' - and it is what turns a stored lead into someone who 'can potentially convert into a real user.' Not SaaS-only: a consulting guide, an SEO/GEO report of the reader's own site (his own product's magnet), a gated podcast, even an event invite qualifies. Cold email with React Email templates is the sibling channel he showed earlier in the programme.

Recommendation stated flat: 'it's just good enough to send a very clean PDF.' A vibe-coded URL dropped into LinkedIn, a network or a cold email 'might be marked as spam sometimes'; a PDF is safe everywhere. Format options: simple PDF, vibe-coded site, or a Google Sheet/Excel for research-style magnets. Jarvis's rule of thumb when asked to be crisp: PDF for polished, offline, structured learning; link for content that updates or embeds interactive tools.

Jarvis drafts, Harshit pastes into a fresh Codex chat. Six required elements: (1) a clear concept and title; (2) a structured outline with outcomes; (3) a concrete artifact and file plan - which files, where they live, how they are generated; (4) a capture-and-delivery flow - the steps a reader takes to receive it; (5) a quality checklist, the 'Did We Do It Right' pass (clarity, accessibility, links, safety); (6) a short demo script for presenting it. Standing constraints baked in every time: practical, plain English, 'non-stubby'; inspect the repo structure before writing; no unnecessary refactors; never deploy unless told; run checks after changes; publishing is 'planned' until a human explicitly approves. Reference the product as 'the above link' once pasted rather than repeating the URL.

The YT Growth Stack draft came back as 'how to use a lead magnet' - Harshit stops it: 'we want to go ahead and create a lead magnet for THIS product.' Jarvis restates the mental model for the class: 'the lead magnet is a proof artifact, not a generic guide... show the product doing its real job, from end to end on one niche.' For YT Growth Stack the proof chain modelled is competitor discovery -> outlier detection -> pattern extraction -> evidence-based idea generation -> production-ready brief, closing with a CTA. 'You're not describing value, you're demonstrating value.' Sahil's own version - built by actually running the product - is held up as the better artifact.

Three rules repeated in every prompt. No fabrication: 'we did not try to fabricate anything' - real data would need the product's own API keys, so unverified research gets 'neutral professional evidence placeholder wording', and a jokey placeholder is rejected as unprofessional. Asset pool: 'inspect the live website content and structure, and treat it as the primary source of truth... use the existing on-site thumbnails, images, testimonials, and pricing as the allowed asset pool, and do not invent new visuals or claims' - extended for ThumbifyX by folding in an existing blog post the agent found on the site; the owner confirms the patterns were genuinely his. Exclusions: internal references ('the SER reference is only our internal thing') and the pricing section - 'the lead magnet should only be a value add for our product', not a website overview.

Every finished magnet ends 'run the same research on your niche - use this link', paired with an incentive: '10% off... or a 3-day free trial'; the Role Compass prompt bakes in 'a reader-only 10% coupon as a placeholder, with clear instructions to replace with a real code later', marked pending until real. Sahil's version carried a line worth copying: 'public product offer and pricing are time-sensitive - confirm the live page before distributing.' Sizing comes from learner Praful: 'moderately sized... the optimum size that the customer will open, plus tracking the open rate' - Jarvis: 'big enough to feel valuable, not so long that people won't finish it', then measure. Planned but not built: a size rubric (short / moderate / longer) and a tracking checklist (delivery-platform analytics, link tracking, email tracking, privacy notes). Real bug: a 13-page PDF rendered as 19 through pagination overflow.

Building four magnets back-to-back in one thread leaked names and claims ('Rahul' credited where Praful belonged; Syed rendered three ways). Learner Pratik asks 'wouldn't you make multiple folders so that it doesn't overlap?' - first waved off ('it's a 1M context model'), then adopted once the mixing was visible: 'we're almost eating up the context - 57% full. That's why it's kind of hallucinating.' Fixes: an explicit isolation clause in every prompt ('product-specific... must not overlap with, reuse, or reference content for other products except generic process rules'); a fresh isolated chat for Home Diary; pass the exact thread/session ID (right-click -> copy working directory / deep link) so same-named sessions are not confused; switch Codex reasoning from Medium to Ultra when output degrades.

Caroline does lead generation for a property client (heard 'Prope Tech' / methodology.com) and asks how any of this applies. Answer: the magnet is any tangible upgrade of the client's existing offering - take a raw asset such as a property sketch and produce an upgraded static visual or one-pager as the free resource: 'if they just get a visual of it, now they're like, boom, hey, I love it, let me just go ahead and check it out.' A one-pager counts; it need not be a multi-page PDF. Cadence, answering Stephen: at launch, on any update worth attention, for new leads entering the funnel, and with every new campaign.

Jarvis is Harshit's personal voice agent built on the Hermes framework: phone/desktop voice in, Codex CLI and a WhatsApp bridge out; it greets the class, reports on a website-build Codex session and a newsletter send (~900 subscribers), drafts every mega prompt, updates its own memory with naming preferences ('Sir Syed Hassan'), and gives recaps. It runs on a Hostinger VPS ('the best one' for CLI support; Railway named by a learner) rather than locally, reached over Tailscale - a free VPN so the phone app only works on his own network; learner Lonnie asks how it is built and Shivani confirms a dedicated Hermes session 'for next week'. Codex is preferred over Claude Code for this loop because the agent integrates with it. Model as heard: 'Astra 6 / GPT 5.6', 'launched yesterday' - unverified.

Sequencing stated early and honoured: 'first, it's built out a strategy. Next, we're doing it for one product. Once we've done all of that, then we eventually just go ahead and say, hey, create a complete skill for this.' Mid-session, learner Prithul proposes the same ('summarize this as a skill') and is told 'do not generate the skill just now.' At the end Jarvis (1) opens a fresh Codex chat titled 'Skill Creation', (2) passes in the full Lead Magnet thread ID, (3) issues a mega prompt to generate a reusable skill from everything in that thread including the corrections, (4) publishes it open-source to Harshit's GitHub under a skills folder - 'Codex command line will take care of it, it has complete access' - and (5) hands it to the Resource Generator to sync into the shared Drive. Jarvis's two standardisation takeaways: consistent flow / verified sources / value-first / no fabricated claims / no internal naming; and asset reuse plus one folder per product.