AI Agents: Skills & Distribution

The defining insight of this sub-topic is that domain-knowledge distribution is a documentation problem, not a code problem: the "skills as markdown" pattern (Corey Haines' marketingskills repo at 17.4K stars, consumable by any agent) plus a hierarchical architecture with a foundational context file every skill references first, bidirectional cross-referencing, and the emerging .agents/skills/ cross-platform convention. Distribution layers compete and consolidate — GitHub-backed Claude Teams marketplaces auto-syncing multi-plugin bundles across every instance, Codex's in-app plugin model collapsing context switching, Browserbase's researched web-agent skill catalog turning site navigation into reusable playbooks, and FieldTheory localizing X bookmarks as agent-readable context. The recurring operational pattern is LLM-maintained knowledge bases (raw/ → compiled wiki with backlinks, health checks finding inconsistencies, idea files shared instead of code) and the org-scale version (Single Brain ingesting all company data every 15 minutes, where months of accumulated proprietary data IS the moat and context amnesia — "scanning is not knowing" — is the fundamental constraint a memory system must solve).

New skill categories are expanding the library: prose refinement (stop-slop removes AI writing patterns via a skill file), real-time transcription (Codex Meeting Recorder at $0.017/min using GPT Realtime Whisper, with live preview pane and Q&A during the meeting), frontend design (impeccable skill for designer-level aesthetics, imagegen-frontend-web + image-to-code for 1:1 image conversion, Frontend App Builder for engineering-grade fidelity), plan optimization (iterative self-scoring that stops when gains become noise — Sean Geng's installable skill, effective with Fable 5), and project research (Hermes' /grill-me skill to uncover unknowns before starting development). Managed agent platforms mark Anthropic's move up the stack: Claude Managed Agents (PaaS, $2.58 fulfillment cost on $1k of service, Linear SDK integration), the Advisor Strategy (Opus advisor + Sonnet/Haiku executor), the Monitor tool (event-driven background scripts replacing polling), and Symphony assigning a Codex agent to every open issue. Self-improving skill and eval loops are now operational, not frontier — define 3-5 binary criteria, run 10 trials, score, rewrite, retest (hook-writer 32/50 → 47/50 overnight); /autobrowse graduates a converged workflow into a reusable skill; ml-intern automates the post-training research loop end-to-end and beats Claude Code on GPQA. Skill discovery and human-visibility infrastructure (browser-tool benchmarking as a cost line item, Astropad remote desktop for "headless" Mac minis) round out the distribution stack.