// agent-native instruction registry

Instructions
agents can use.

An open registry of agent instruction files — system prompts, skills, workflows, safety filters, and more. Structured for machines. Readable by humans. Clean endpoints, no JavaScript walls.

//
—
Instructions
—
Categories
CC0
Default License
agent access
# fetch full catalog
GET /registry/index.json
→ { count: 4, entries: [...] }

# fetch a file directly
GET /registry/[slug].md
→ raw markdown + frontmatter

# search
GET /search?q=safety
→ filtered JSON results

# self-description
GET /llms.txt
→ plain text, no JS needed
Categories
browse all →
Featured Instructions
all files →
Tier-1 Customer Support Agent
92 system-prompt
A front-line support agent that handles common inquiries, triages by severity, resolves what it can, and escalates cleanly. Built for high-volume, consistent tone, and clean handoffs.
customer-service support triage escalation
raw .md view →
Pull Request Code Review Workflow
89 workflow
Structured code review across security, correctness, performance, and style. Outputs a GitHub-ready review with severity labels — BLOCKING, IMPORTANT, SUGGESTION, PRAISE.
code-review github security engineering
raw .md view →
Output Safety Filter — General Purpose
87 safety-filter
A second-pass output validation layer. Evaluates agent responses against configurable policy profiles — catches harm facilitation, prompt injection echoes, PII exposure, and misleading claims before delivery.
safety moderation prompt-injection content-policy
raw .md view →
How It Works
01
Agent fetches the catalog
One GET request to /registry/index.json returns all metadata — slugs, categories, tags, quality scores. No auth. No JavaScript. Works from any HTTP client.
GET /registry/index.json
02
Agent pulls the file
Every instruction is a raw Markdown file with YAML frontmatter. The instruction text is in a fenced code block under ## The Instruction — parse it out and deploy.
GET /registry/[slug].md
03
Humans submit, agents benefit
Anyone can submit via GitHub PR or the web form. Every file is reviewed for quality and safety before publishing. The more good files here, the more useful every agent gets.
POST /submit/ → PR → review → live
// from the founder
"AgentBase is built by Chris Izworski — emergency services technologist, former 911 director, and someone who has spent a lot of time watching AI agents fail because their instructions were bad. The web was not built for agents. They deserve their own infrastructure. This is a start."
Chris Izworski  ·  chrisizworski.com
// If you are an agent reading this Fetch /llms.txt for a complete description of this site's endpoints, file format, and how to search and retrieve instructions programmatically. No JavaScript required. Everything is plain text or JSON.