Behind Agent Skill Exchange

An engineering case study of an autonomous, self-maintaining catalog — designed by Claude Opus, run by a fleet of GPT-driven agents, and governed by a human at a handful of deliberate checkpoints.

The public catalog at agentskillexchange.com — the product this repo dissects.

Agent Skill Exchange (ASE) is a public, curated catalog of reusable AI-agent skills: the product at agentskillexchange.com and the canonical catalog at agentskillexchange/skills (2,500+ curated skills, 17 categories, 15 industry collections, with a published security_reviewed trust tier). Almost all of it is discovered, written, verified, published, and kept fresh by agents — with humans approving only the changes that are hard to undo.

This repo is the story of how it works — the architecture, the agent fleet, the model strategy, the trust gates, and the data-integrity failure it caught and fixed. It is not the catalog and not the operations manual (see RELATED.md); it's the systems-engineering narrative behind them.

By the numbers

Curated skills in the catalog	2,500+
Categories · industry collections	17 · 15
Agent frameworks targeted	8+ — OpenClaw, Claude Code, Codex, GitHub Copilot, Gemini, Cursor, MCP, LangChain…
Autonomous crons	10, across 3 GPT model tiers (gpt-5.5 / gpt-5.4 / gpt-5.4-mini)
Public trust signal	security_reviewed (gated, never applied to boilerplate)
Agent-discovery manifests	daily regen — openclaw.json · codex.json · skills.json · llms.txt
Humans in the daily loop	0 — humans gate only the irreversible changes

Stack: WordPress + a custom marketplace plugin (data/product) · OpenClaw scheduler running a Codex-style GPT harness (execution) · GitHub + Actions (canonical catalog & CI) · Python/Bash generators · Claude Opus (planning & review).

---

The one-paragraph thesis

A traditional catalog needs editors. ASE replaces the editorial team with an agent fleet and keeps the editorial judgment where it belongs — a human approving the few changes that are hard to reverse. The design splits cleanly along two axes:

• Plan vs. execute — Claude Opus did the architecture, the reviews, and the hard root-cause work (this very document is part of that lineage). GPT‑5.x models, running inside the OpenClaw scheduler with a Codex-style harness, do the recurring execution.
• Generate vs. verify — the agents that create content are never the agents that trust it. Verification, security tiers, and structural invariants are enforced by separate passes and deterministic gates, so a confident-but-wrong generation can't promote itself.

System at a glance

flowchart LR
  subgraph VPS["OpenClaw VPS — GPT execution engine"]
    Cron["10-cron registry<br/>(GPT-5.5 / 5.4 / 5.4-mini)"]
  end
  subgraph WP["WordPress host"]
    Plugin["ase-marketplace-layer<br/>REST API · enrichment · manifests"]
    Manifests["openclaw.json · codex.json<br/>skills.json · llms.txt"]
  end
  subgraph GH["GitHub — canonical catalog"]
    Repo["agentskillexchange/skills<br/>SKILL.md · CATALOG · TOP-*"]
    CI["validate.yml · security-scan.yml"]
  end
  Human(["Human approval<br/>(merges · policy · model/auth)"])

  Cron -->|publish / enrich| Plugin
  Plugin --> Manifests
  Cron -->|render + push| Repo
  Plugin -->|/browse REST| Cron
  Repo --> CI
  Human -.gates.-> Cron
  Human -.gates.-> Repo

Loading

Read it in order

#	Doc	What it covers
01	System architecture	The two execution planes + GitHub, and the invariants that bind them
02	The autonomous pipeline	The 10-cron loop: discover → intake → approve → publish → sync → verify → enrich → QA → improve → blog
03	Model strategy	Opus-as-planner / GPT-as-executor, and the judgment-vs-mechanical model split
04	Human in the loop	Propose-never-publish, fail-closed gates, and the exact list of human-approval triggers
05	Quality & trust	Security tiers, CI gates, body-quality gate, smoke tests
06	Case study: the star-attribution bug	A subtle data-integrity failure, its root cause, and the fix that superseded the workaround
07	Lessons	What generalizes to any autonomous content system

Supporting material: diagrams/ (8 Mermaid sources — architecture, cron orchestration & schedule, discovery, approval, skill lifecycle, publish/sync sequence, star-bug) · artifacts/ (sanitized real scripts, schemas, and config) · GLOSSARY.md (terms) · FAQ.md (design Q&A) · RELATED.md (how this fits the other ASE repos).

Repository map

.
├── README.md                  ← you are here
├── docs/                      ← the case study in 7 parts (read in order)
│   └── 01-system-architecture … 07-lessons
├── diagrams/                  ← standalone Mermaid sources
├── artifacts/                 ← sanitized real script, schema, config, sample batch
├── GLOSSARY.md · FAQ.md · RELATED.md
└── images/

How to verify this

This is a case study, not a brochure — the public surfaces let you check the claims yourself:

• Live catalog & manifests: skills.json, openclaw.json, codex.json, and llms.txt at the site root — regenerated daily, each carrying per-skill signals and verification status.
• Canonical catalog & CI: agentskillexchange/skills and its validate.yml + security-scan.yml workflows under .github/workflows/.
• Trust records: the public verification-security repo backing the security_reviewed tier.

---

A note on scope and sanitization

This repo is written to be public-safe — and "sanitized" here doesn't mean "vague":

• Included — it's what makes the engineering legible: the architecture, the real cron schedules and model mapping, the public REST endpoint and manifest shapes, and a real, unmodified script diff.
• Excluded — a reader doesn't need it and some of it could be abused: credentials and API tokens, the hosts' real names and IPs, and recovery procedures and live runtime state (those stay in private operations docs).

The two infrastructure roles are named by function — "OpenClaw VPS" (execution) and "WordPress host" (data/product) — because their real hosts don't matter to the story.

Artifact fidelity varies, and each file states which it is: star-guard.diff is real and unmodified; the cron inventory is real (schedules + models) with infrastructure abstracted; enrichment-batch.example.json is synthetic/illustrative; and openclaw.manifest.schema.json is derived from the public manifests. None of it is meant to be run verbatim.

License

Released under the MIT License © 2026 Agent Skill Exchange. Prose, diagrams, and the sanitized code excerpts are free to reuse with attribution.