Office Hours VI: Paul's SEO/AEO/GEO Question - Moving a Lovable Site into GitHub and Claude Code, the GEO Skill, llms.txt, and CLIs as the Glue

Harshit's rule for Paul's agency work: Lovable and its peers are excellent at design and taste, so use one for the first version. But while the code lives in Lovable, Lovable owns it - you cannot deep-edit structure, metadata, files and routes the way SEO and GEO work requires, and pushing it out is deliberately awkward ('Lovable wants to own the core'). So: v1 in Lovable, then connect GitHub and move the codebase to your own repo, then do ongoing SEO, GEO and backend work in Claude Code, which 'does back end really well'. Keep Lovable only for occasional aesthetic passes - GitHub-syncing back is technically possible but burns credits and adds nothing. On WordPress: a site the client wants to edit themselves is 'definitely possible' through a WordPress CLI installed into Claude Code exactly like GitHub CLI.

On a mock client repo ('Iris Whale', an AI-consulting landing page) Harshit installs the GEO Content Optimizer skill from skills.sh and runs it read-only: 'don't try to commit anything - tell me all the improvements that can be done'. The skill reports; he then narrows its brief from on-page SEO to citation optimisation for AI search engines - ChatGPT, Perplexity, AI Overviews - and asks it to complete and report before pushing. Alongside it, llms.txt: a proposed standard file that tells LLMs how to use the website at inference time, so AI answer engines parse and rank the site. Both were promised to Paul in a written doc. The Saturday Session 23 that followed built the whole audit loop.

The most repeated move of the hour. Rather than learning each tool's dashboard, install its command-line interface and let Claude Code or Codex operate it conversationally: GitHub CLI to query commits and repos; Linear via its API key ('install the Linear MCP server for yourself and get fully integrated') so the agent reads and writes tickets and marks cards done; Railway CLI for deploys, which turns out to ship its own agent skills for SSH, databases and environments; and, for Paul, a WordPress CLI. A home-made Kanban works the same way once it exposes an API.

Sync across devices by pushing everything to GitHub (API keys excluded, re-set per environment); a skill installs by pasting its repo link. A single source of truth is PRD files plus a branch per workstream (a 'codex security review' branch merged to main) tracked in Linear, which Claude Code updates itself. Agent wake-up after a crash is a memory.md file - 'can you add this to our memory'. Modular repos per team versus one mono-repo is a founder's call; the mono-repo buys shared context at the cost of onboarding friction. Validate architecture early with Fable's /goal against the whole repository, giving it vision, timeline and inputs. Keep project memory in a Hermes agent and a visual Kanban pulled in over API, because Claude Code's own interface is text only.

Get the client's full requirements first. Ask Claude - using cheap Haiku for the estimate itself - to project token cost page by page against Anthropic's published pricing. Build the proposal, with markup, using Sonnet. Never show the client the underlying API and model costs. Then the trick: Pixpipe, an architecture (not a skill, but installed the same way) that converts context to an image before sending it to Fable, whose OCR is strong enough that the same information costs 60-70% fewer tokens - a cited benchmark claims ~38% fewer with comparable accuracy. It does nothing for Haiku-class models, whose OCR is not good enough.

Yes, it can. The protection is a complete PDF of what was delivered and what it cost, handed over with the work, and a clear position that changes the client makes afterwards are theirs - the car-modification analogy: alter the engine and the warranty is gone.

Fable is an underlying model; Lovable and Genspark are products that may run Fable (or something else) underneath and mark up the cost. Harshit's complaint: Lovable used to let you pick the model and now hides it - 'a black box' - so you cannot see what your money buys; Emergent still shows the choice. For app UI without a wrapper, install a front-end skill (taste skill, or Anthropic's own front-end design framework) from skills.sh into Claude Code. His closing thought to Sarah about Cursor applies here too: chasing the flashiest new tool is a mistake; 'sometimes sticking to basics always helps'.

Always a short .com - 'the most trusted, it indirectly helps with SEO'. Test mobile with the Mobile Simulator Chrome extension. Simplify a wrapping header; put the email-capture field at the very top of the page with no extra click, and allow a re-submit if someone mistypes. Run the GEO skill and install PostHog. For dashboards, use shadcn/ui blocks and Magic UI's free templates rather than hand-built layouts; mark required fields as required. Tailscale in passing: a free VPN so a Hermes agent appears to be on your home IP - 'a standard protection layer for your AI agents'. Hostinger for a VPS. And Vikram's warning: bypass-permissions mode once installed something he had to dig out by hand - 'scary'.