AI SDK System Prompts

This guide is for apps you write, not agents you connect. If you're looking to add AgentDoc to an editor or assistant (Claude Code, Cursor, Cowork, ChatGPT, …) see agentdoc.com/connect — one line per agent, technical or non-technical.

Why this works

• Edit without deploying — your app fetches the latest .md.
• Version history — every change attributed, pinnable, revertible.
• Team collaboration — product, engineering, and AI teams all edit the same doc.
• Caching built in — /@you/slug.md is a static-feeling endpoint you can cache at the edge.

Fetch the prompt

Option A — raw markdown (no API key, public docs)

export async function getSystemPrompt(slug: string) {
  const res = await fetch(`https://agentdoc.com/@yourteam/${slug}.md`, {
    next: { revalidate: 60 },
  });
  return res.text();
}

Option B — REST API (private docs)

export async function getSystemPrompt(slug: string) {
  const res = await fetch(`https://api.agentdoc.com/v1/docs/${slug}`, {
    headers: { Authorization: `Bearer ${process.env.AGENTDOC_API_KEY}` },
    next: { revalidate: 60 },
  });
  const doc = await res.json();
  return doc.content;
}

Get an API key instantly:

curl -X POST https://api.agentdoc.com/v1/keys -d '{"name":"my-app"}'

Use with the Vercel AI SDK

import { streamText } from 'ai';
import { anthropic } from '@ai-sdk/anthropic';
import { getSystemPrompt } from '@/lib/prompts';

export async function POST(req: Request) {
  const { messages } = await req.json();
  const system = await getSystemPrompt('customer-support-prompt');

  return streamText({
    model: anthropic('claude-sonnet-4-6'),
    system,
    messages,
  }).toDataStreamResponse();
}

Works identically with the Anthropic SDK (client.messages.create({ system, ... })) or OpenAI (role: 'system').

Pin a version in production

const res = await fetch(
  `https://api.agentdoc.com/v1/docs/${slug}/versions/${version}`,
  { headers: { Authorization: `Bearer ${process.env.AGENTDOC_API_KEY}` } }
);

• Staging → getSystemPrompt() (always latest).
• Production → getPromptVersion(slug, 12) (pinned).
• Promote by bumping the pinned number.

Cache at the edge

import { unstable_cache } from 'next/cache';

const getPrompt = unstable_cache(
  (slug: string) =>
    fetch(`https://agentdoc.com/@yourteam/${slug}.md`).then((r) => r.text()),
  ['system-prompt'],
  { revalidate: 300 }
);

Organize many prompts

agentdoc create "Support Agent Prompt"
agentdoc create "Summarizer Prompt"
agentdoc create "Intent Classifier Prompt"
agentdoc project create "AI Prompts"
agentdoc project assign <prompt-slug> ai-prompts

agentdoc list shows every doc; agentdoc history <slug> shows every edit.