Blog posts on building with AI agents
Memory alone is not enough for recurring AI behavior. Why the rule has to live in the skill file AND its launcher to actually stick.
workflowThe shift from a clever stack of prompts to a system with clear owners for repo policy, local adapters, skill runtime, and workstream state.
control-planeUseful agent workflows should prepare the exact action, corroborate inside a bounded lane, and pause before the side effect.
workflowThe AI PM OS stopped feeling like a morning trick once the workflow had a full lifecycle: start with a planner, sync visible state, and close with a written carry-forward.
ai-pm-osThe real leap isn't just a smarter model. It's durable context, ranked workstreams, startup routines, and an environment designed for re-entry.
ai-pm-osA technical walkthrough of start-day, the Chief of Staff planner, the public state layer, and resumable cmux workstreams.
Why I turned a growing pile of AI skills into a product-management operating system with durable workstream context and a start-of-day command center.
ai-pm-osThe auto-pilot skill routes requests to the right skill automatically. Routing tables, intent parsing, multi-skill chaining — how one meta-skill replaced a 97-option menu.
orchestrationI published my AGENTS.md — the single file that configures how my AI agent behaves across 76+ skills. Identity, memory, to-dos, boundaries, and tool routing in one markdown document.
agents-mdHow a team went from markdown-first specs to an agent skill that writes PRDs with auto-referenced internal sources — in two weeks.
product-managementA practical guide to building 35+ AI agent skills that automate morning triage, email, project management, and more — no engineering degree required.
workflowAfter distilling 2,718 observations, found ~90% were noise. Seven plugin changes cut volume by ~80% while preserving signal.
memory-systemA proactive to-do system that auto-captures from Slack, Gmail, Calendar, and Linear — prioritizes, syncs to calendar, and surfaces at every session.
workflowAfter a week of running amp-mem, context injection got noisy. Five targeted fixes brought the signal-to-noise ratio back.
memory-systemDebugging why amp-mem plugin tools returned empty results — the silent failure mode of Amp's experimental $ tagged template API.
debuggingFrom zero to cross-session intelligence — a ~500-line bash CLI backed by SQLite + FTS5, paired with a TypeScript plugin that hooks into Amp's event system.
memory-systemA "start of day" skill that triages Slack, Gmail, Calendar, and comments from Figma, Docs, Linear, and Notion — inbox zero in under 10 minutes.
workflow