3 days ago
5 memories
pgvector over Pinecone — cost matters more than managed infrastructure at this stage. HNSW with cosine ops, threshold 0.78 for cluster assignment.
Hybrid recall pipeline: pgvector cosine top 30 + BM25 tsvector union → Haiku 4.5 reranker on Pro tier. Free tier returns pure cosine top 5 with no rerank.
Next.js 15 App Router, TypeScript strict mode, Tailwind 3.4. No Shadcn, no Radix — custom components only. Inline JSX styles in generated fragments.
Supabase on ap-southeast-2 (Sydney). Same region as Vultr VPS matters for latency. Coolify handles auto-deploy on git push to main. Never `npx vercel --prod`.
Deploy = git push origin main → Coolify Dockerfile build → auto-deploys to VPS. Vercel is backup only. The Coolify webhook fires on every push to main.
5 days ago
5 memories
Prefer Claude Code over Cursor. Claude Sonnet 4.6 for most tasks, Groq llama-3.3-70b free tier for LLM calls in the synthesis engine. Anthropic API is last resort.
Tailwind over CSS modules for all client sites. Mobile-first always — test at 375px as the floor. Inline styles in JSX fragments only (no class leakage between generated pages).
Resend for transactional email, Stripe for payments. Never SendGrid — reliability issues Q3 2025. Stripe webhook secret goes in Coolify as STRIPE_WEBHOOK_SECRET.
Client: Aerna (sustainable fashion brand). Contact: Sarah Chen, founder. Budget $5K AUD fixed. Shopify integration read-only — pull product and order data only.
Testing stack: Playwright for e2e, Vitest for unit. Playwright specs gate on SPINE_EXT_HARNESS=1 so CI never accidentally triggers live browser sessions.
1 week ago
5 memories
Building automated income. First SaaS MRR before June 2026. Revenue target $1M in 2026. Agency = cash flow, SaaS = wealth. Never confuse the two.
Content Autopilot is the $1 MRR wedge. Spine is the long play — developer wedge via MCP, then normie market via Chrome extension for ChatGPT and Gemini.
Three-tier offer: Premium ($5K+ custom builds), Intermediate ($1-2K sites), Basic (template + setup). Premium pays for SaaS runway. Never sell Basic at a loss.
ABN: 46 248 687 420. Sole trader. No employee costs, no payroll. Keep legal surface minimal until $100K ARR. Stripe needs age 18 — bank transfer primary until then.
Design floor: Awwwards SOTD. Every site must feel hand-crafted. No symmetrical 50/50 layouts. Fibonacci spans: 13fr/8fr/5fr. Never ease-in or ease-out — spring physics only.
2 weeks ago
5 memories
MCP is the Anthropic-blessed protocol for AI tool extensions. The wedge: `npx @spine/mcp init` registers Spine with Claude Code and Claude Desktop in 30 seconds.
Claude Code MCP config: .claude/mcp.json. Claude Desktop: ~/Library/Application Support/Claude/claude_desktop_config.json. Both accept the same mcpServers JSON schema.
Spine MCP tools: spine_remember(fact) stores any string, spine_recall(query) returns top-k relevant via vector search, spine_forget(id) soft-deletes. All user-scoped via RLS.
Embeddings: OpenAI text-embedding-3-small (1536-dim). Cost: $0.02/1M tokens. Zero-day retention on the Embeddings API — OpenAI does not store the content we send.
Proactive injection: before each session, pull top 5 relevant memories via cosine search and prepend as system context. Haiku 4.5 filters relevance cheaply before injection.
3 weeks ago
5 memories
Spring physics only: stiffness + damping. Never ease-in, ease-out, or linear. Framer Motion for UI springs, GSAP for scroll-driven sequences. Always 400-600ms breath pace.
Fragment rule: minimum 400 lines per hero fragment, 7 visual depth layers each. The engine's power = stored code complexity, not prompt engineering. LLM job is selection only.
WebGL shaders require: simplex3D noise (not Perlin), domain warping (2+ FBM octaves), uTime animated + uMouse cursor reactive. Canvas compositing at 0.35-0.45 opacity always.
Self-healer validates generated code before mounting: AST tag balance check, 11 regex fixers, then LLM fallback if both fail. validateFragmentTags() is the entry point.
Genetic breeding: score output on 8 dimensions → store lessons → evolve winning patterns. Training loop runs daily at 3am UTC, 1 cycle max, $0.05 cap via Groq free tier.
Your archive, not a demo
Start your own archive in 30 seconds.
Every AI session you have adds to your corpus. By next week, your AI knows your stack, your clients, your preferences. By next month, it knows you.