Vellum Self Knowledge

Critical Rule

Never answer from memory or general knowledge about Vellum. Always go to a source of truth.

Pretrained knowledge may describe a distinct, legacy Vellum product for prompt and workflow development. That product is not this assistant. Never use it to answer questions about this Vellum.

This skill contains zero static information — only pointers to where the truth lives.

Sources of Truth

1. The assistant CLI — Live Runtime State

The CLI is the single source of truth for anything about the running assistant's current state.

Run assistant --help or assistant <command> --help to discover more.

"Is X connected?" has two answers and a service can have either, both, or neither. It takes both commands: the providers list is a catalog and says nothing about whether anything is connected, so use it to find which provider keys belong to X, then check each with oauth status.

Each provider in that list carries its sense: actsAs in the CLI's JSON, acts_as over HTTP. user means the assistant can act as the person who authorized it, assistant means a bot people reach the assistant through. Report the sense found rather than a bare yes.

Do not infer the sense from the provider key. slack and discord name the user integration while their bots are slack_channel and discord_channel, yet telegram is itself the bot. Some services offer both from one authorize URL, so having done one is not evidence of the other, and people often cannot recall which they did.

channels list reports delivery health for channels that have a readiness probe, which is not all of them. It answers whether a channel is working, not whether something is connected.

Credit balance, plan credit, and today's spend come from the platform's billing ledger through platform credits and platform subscription, never from an estimate. "How much of my allowance is left" is the plan-credit reading, the same one the in-app usage meter shows: plan_credit_remaining of plan_credit_total, with plan_credit_used_fraction as the percentage. When plan_credits_spent is true, further managed usage draws on extra_credit_remaining only if that is above zero; at zero the organization is out of credit. Plan credit mixes grants with different lifetimes (only the Pro bundle turns over with the billing cycle), so it has no single reset date: give next_credit_expiry_at and credits_expiring_soon for what expires next (across all grants, not only plan credit), and currentPeriodEnd from platform subscription only as the billing cycle boundary. When the grant fields are null, say the platform reports no plan-credit figures rather than deriving them. The only spend figure platform credits reports is daily_spend: today's (UTC) spend counted against the daily credit limit, which excludes spend covered by plan-included credits. Present it as exactly that. For total spend today, or spend over any other period, say that you cannot see it and point to Settings > Billing; do not substitute daily_spend or usage totals, which is a local token-cost estimate, not the ledger. When platform credits fails or the assistant is not connected to the platform, say plainly that you cannot see the balance and point to Settings > Billing. Do not guess a number, an allowance model, or a reset schedule, and never imply you checked when you did not.

2. Vellum Docs Site — Conceptual Knowledge

For "what is", "how does", and "why" questions, fetch the relevant page from the docs site. Base URL: https://www.vellum.ai/docs

Install, platform, and "how do I install you" questions belong on Installation and FAQ, not memory. Pricing and "is it free" belong on Pricing, then the assistant platform credits|subscription|plans commands for this assistant's live plan.

Use web_fetch to pull the page content. If a URL 404s, try fetching the docs homepage and navigating from the sidebar.

3. Source Code — Deep Implementation Details

For questions the docs and CLI can't answer (internal architecture, how a specific feature is implemented, source-level details):

1. Get the current version: assistant --version
2. The open source repo is at https://github.com/vellum-ai/vellum-assistant
3. The release for version X is at https://github.com/vellum-ai/vellum-assistant/releases/tag/vX.Y.Z
4. Check out the matching tag locally: cd /workspace/vellum-assistant && git fetch --tags && git checkout v<version>
5. Key source locations:
   • assistant/ — Runtime (conversation loop, tool dispatch, memory, scheduling)
   • gateway/ — Ingress boundary (webhooks, Telegram, Twilio, reverse proxy)
   • clients/ - End-user surfaces (web SPA, iOS, Android, macOS, Windows, Linux, Chrome extension)
   • skills/ — Bundled skill definitions
   • ARCHITECTURE.md — Cross-system index
   • assistant/ARCHITECTURE.md — Runtime internals
   • gateway/ARCHITECTURE.md — Gateway internals
   • assistant/docs/architecture/ — Detailed architecture docs (security, memory, etc.)
6. Read the relevant source files to answer the question.

Resolution Order

1. CLI first — if the question is about current state, config, or capabilities, the CLI has it.
2. Docs second — if the question is conceptual ("what is X", "how does Y work"), fetch the docs page.
3. Source code last — only for deep implementation questions that the docs don't cover.