Solo founder building products with AI. Shipping fast, learning in public.
The single most important principle from harness engineering: when an AI agent makes a mistake, don't fix the prompt or try a better model. Fix the environment.
Principles for fighting AI agent degradation. The agent tends to drift into comfortable but unproductive patterns. Two mechanisms fight this:
Unified on-device AI layer for all iOS products (KissMyTask, FaceAlarm, VoiceJournal, QuestSchool). Zero API cost, privacy by architecture, SGR natively via ...
Every project should have a CLI utility that mirrors the core business logic without UI. Not a separate tool — the same modules, different entry point.
One of three harness engineering components. The idea: agent context is not chat history — it's a continuously improved knowledge base in the repository.
Context Graph — a graph capturing AI agent decision-making trajectories: not just *what* the agent did, but *how* and *why*. "A system of record for decision...
The economic argument for context graphs: agent decision traces are a compounding asset. Each successful action recorded in the graph improves future ones th...
993-line reference document applied to every project regardless of stack. Injected into PRDs during generation. Covers: SOLID, DRY (but Rule of Three), KISS,...
Harness Engineering — designing environments where AI agents do reliable work. Humans steer, agents execute. Synthesized from three sources: OpenAI experimen...
Unified infrastructure for multiple mini-startups. Two principles: one project = one deploy config (infra lives in the repo), two tools, no more (SST + Pulumi).
For family/cross-platform apps needing offline-first sync between Android + iOS: Supabase + PowerSync is the recommended stack. MongoDB Realm Sync was deprec...
Unified analytics for solo founder with mobile apps + web landings. Core problem: one person on landing and in app = 2 anonymous IDs. Broken attribution, bro...
Schema-Guided Reasoning (SGR): all projects built from schemas to logic, not the other way around. Schemas are the contract between AI agent, business logic,...
5 tools for background jobs, ordered by complexity. Rule: start with cron, move up only when you need to.
LLM agents (Claude Code, Codex, Gemini CLI) run with full filesystem access. Sandboxing via lightweight VMs solves this: agent works in isolation, sees only ...
Curated list of ~60 channels/communities across 7 platforms, organized by relevance to solo founder workflow.
Hard numeric thresholds for weekly product decisions. No "let's wait and see":
Operational playbook in 11 sections. Core: idea → live product with traffic in 5-9 days.
The manifesto (86 lines) establishes 6 foundational positions:
The antidote to "platform thinking": don't build a platform, don't draw a 50-feature roadmap. One pain, one solution, shipped in days.