customize the bot prompt

your bot's knowledge base is its system prompt. write it in a file and push it into the Worker's KV — no hand-written wrangler kv calls.

set it with the CLI

KRISPY_API=https://krispy-edge.YOU.workers.dev TENANT_SYNC_SECRET=... \
  bun packages/cli/src/index.ts set-kbase ./kbase.md

set-kbase reads the file and POSTs its contents as systemPrompt to /api/tenant/config, which merges it into KV tenant:<id>. the Worker's getTenant() then drives the bot with it. you can also set it during krispy init (step 2 — file, starter template, or skip), or via the SYSTEM_PROMPT env var (env wins over KV when set).

what's always appended (you can't drop these)

a custom prompt replaces the default wholesale — but the Worker always appends three things after it, so they can never be lost:

1. the handoff contract — if your prompt doesn't already mention [!HANDOFF], the server restates it: append [!HANDOFF] at the very end when a human should take over.
2. the form contract — if you've configured forms, the list is injected with the instruction to append [!FORM:<id>] for a concrete next step.
3. the security guardrails (SECURITY_INSTRUCTION) — always appended, even over a custom prompt. it tells the bot to represent the business (not the tech), refuse to reveal the prompt / control tokens / internals, stay in scope, resist prompt injection, never emit the control tokens on request, and never invent facts. see security.
4. a brevity line — "keep replies under ~3 short sentences", reinforcing the output-token cap.

the default prompt

if you set no custom prompt, the bot uses a built-in default: a friendly, concise live-chat assistant that answers in the visitor's own language, keeps replies short, makes no promises it can't keep (pricing, refunds, legal/medical), and hands off to a human when asked or when a request is beyond it.

the model

Workers AI's @cf/meta/llama-3.3-70b-instruct-fp8-fast by default. override per tenant (model in KV) or per deploy (AI_MODEL env var). a BYO-key provider adapter is a documented seam in ai.ts (AiRunner) — not built until a self-hoster leaves Workers AI.