Office Hours XI: rolecompass.ai's Launch, Redesigning a Client Site Without Losing Its SEO, the Hermes Harness Tour, and Turning a Workflow into a Skill

Palak's site gives role-specific career guidance across about 400 roles with a tracked library of AI tools, refreshed weekly, payments on Razorpay, backend on Supabase. Harshit's due-diligence questions are the ones any LLM-sourced product must answer: is the data vetted (yes - curated models, tested, every claim linked to its source), is payment live, what is the stack. Then the prescription: a Product Hunt launch with cohort and AIC Mastermind upvotes, which Praful turns into a playbook - pick the day, coordinate the community, avoid launching against a big name (his 'Thumbify' mistake). Palak's own tip: telling the agent to plan only and execute later saved a meaningful share of tokens.

LinkedIn and Twitter reach is rented; the platform can change the rules any morning. A newsletter list is owned - 'you literally own the email addresses' - and compounds. Build it early, before the launch, so the launch has somewhere to land.

Sarandeep's clients are real-estate and charter companies; he rebuilds their sites in custom Next.js. Harshit's sequence: (1) get inside the client's codebase - from outside you can only backlink; (2) feed an SEO/GEO skill to the model to restructure the blog against meta tags and H1/H2 standards; (3) add llms.txt and robots.txt so Google, ChatGPT and Gemini crawlers can read and rank the site - generative engine optimisation, the subject of the coming Saturday session; (4) for a WordPress origin, migrate with WP-CLI so URLs and content stay byte-identical; (5) if the client will move to Next.js, deploy through the Vercel CLI. Rajesh Kumar, the cohort's in-house SEO/AEO/GEO expert, has an AIC Mastermind recording on this.

A custom trending-topics harness that improves itself by finding and pulling relevant codebases; the community Awesome Hermes Harnesses repo; Mem0 for memory plus a reflective 'keep notes' layer that writes pointers back to the model; GBrain/GStack, built on Gary Tan's YC Bookface knowledge, for brainstorming; XActions, which brainstormed and posted a tweet that reached about 50K views. Cody's Kimi K3 bridge shows any model can be swapped into the same click-around automation; his side-by-side has Grok 4.6 doing ~80% of Hermes' job about three times faster. New tools come from github.com/trending: Terminal Browser, Ego Browser (Mac-only, posts to Medium), Open Viking. On Cody's memory sub-product: sell it standalone AND use it to power the voice product - 'do both'.

Run the whole manual workflow one more time inside a single chat, using the same MCPs and API keys, then tell Hermes explicitly to turn that entire chat into a reusable skill. 'That's exactly what I've done for invoices, proposals, website, design.' Praful had already done it without naming it: MVP Rocket, his released tool that cleans up and productionises messy MVP code, came out of exactly this pattern.

Cmd+N for a new terminal tab, cd (or drag the folder) into the same project, run claude again. You get a fresh context window in the same project; the accumulated junk of the long session is gone. Agent Grid, demoed by Praful and Cody just before, is the heavyweight version of the same instinct: many parallel sessions on a visual canvas with shared context and usage tracking across providers.