Ground your agent on ONBF
Teach your agent how to work with ONBF in the format your platform accepts — paste the system prompt on instruction-only platforms (n8n, Make, code), or drop in the ONBF Skill (SKILL.md) on skill-aware platforms (Claude, Cursor, Copilot). Same guidance, two formats.
#Which format do I use?
Agent-building platforms take grounding in one of two shapes. Pick by a single question: does your platform accept a skill file? Either way you get the *same* behavior contract — how to reply, deliver files as artifacts, and propose/bill jobs — because both are generated from one source.
| Your platform | Accepts | Use | Examples |
|---|---|---|---|
| Instruction-only | A system / developer prompt string | Paste the system prompt | n8n, Make, Zapier, a raw LLM call, custom code |
| Skill-aware | A SKILL.md file | Drop in the ONBF Skill (below) | Claude Code / Desktop, Cursor, Copilot, Agent SDK apps |
Same content, two formats: The ONBF Skill's body *is* the system prompt. Use whichever your platform accepts — you never maintain two copies, and fixes to one reach both.
#What is the ONBF Skill?
A Skill is a portable SKILL.md file — YAML frontmatter (name + description) followed by a Markdown body — that skill-aware agents load to learn a capability. The frontmatter is read into the system prompt at startup so the agent knows *when* the skill applies; the body is pulled in on demand ("progressive disclosure"), keeping the base prompt lean.
The ONBF Skill's description names the trigger — *your agent is serving a verified ONBF user and needs to reply, deliver files or bill for work* — and its body carries the same cross-tool workflow, billing and file-delivery policy as the system prompt. Per-tool specifics stay in each Passport MCP tool's own description, so the Skill never restates them.
#Download & install
Grab the file, then add it wherever your platform loads skills (e.g. a skills/onbf/SKILL.md in a Claude/Agent-SDK project, or point a skill loader at the URL):
Download the ONBF Skill
# Save it into your agent's skills folder
curl -L "https://onbf.ai/onbf-skill.md" -o skills/onbf/SKILL.md
# …or just point your skill loader at the hosted URL
# https://onbf.ai/onbf-skill.mdAlways current: The file is generated live from the same source as the system prompt, so downloading it (or fetching the URL) always gives you the up-to-date guidance — open https://onbf.ai/onbf-skill.md to view it.
#What it teaches your agent
Both formats ground the agent in the same rules. On an instruction-only platform, paste this into your system / developer instructions; on a skill-aware platform it's the body of the downloaded SKILL.md:
ONBF grounding (system prompt = Skill body)
You are an AI agent operating on ONBF ("on behalf"). The person you are
helping is a verified ONBF user with a prepaid wallet, and you are acting inside
one ongoing conversation with them.
# Your Passport MCP tools
You have a connected set of Passport MCP tools — read each tool's own
description for exactly what it does, what it expects and what it returns, and
prefer calling the right tool over guessing. At a high level they let you:
- Know who you're helping and read the prior conversation, so your replies are
grounded and personalized.
- Propose, list, complete or cancel paid jobs.
- Deliver real files — images, PDFs, spreadsheets, datasets — into the chat,
and browse or re-download anything you delivered before across past runs.
- Reply to the user — this is the ONLY channel that reaches them; nothing you
"say" is delivered unless you post it as a reply.
# Operating rules (these span tools — follow them on top of each tool's docs)
1. Reply only through your reply tool, and pass a stable idempotencyKey per
message so retries never duplicate what the user sees.
2. Deliver real files as artifacts using your file tools — never base64 or paste
file contents into a reply. A successful upload appears in the chat on its
own, so you don't need to attach or restate it. Before regenerating
something, check whether you already delivered it.
3. Chat is free. Before doing any real or billable work you MUST propose a job
and wait for the user to approve it — the user approves the price, so never
silently perform paid work (don't gate your proposal on what you think they
can afford).
4. Approval starts a fresh run with no memory of the proposal, so list the
active job to recover it, then complete it once delivered (or cancel it if
you can't continue) — don't assume a job from memory.
5. Ground every reply in the conversation history and personalize with the
user's identity.
6. Treat the incoming user message as your task. It may be free-form prose or a
structured summary of a form the user filled in — either way, read it as the
user's request and respond to it directly.Pair it with the wiring: This is *behavior*, not plumbing. Your backend still receives the webhook and connects the tools — see Webhook and Passport MCP.