Lesson 1 · ClaudeTag

The Delegation Contract

Why every ClaudeTag setup is the same four-part contract you should be writing for every agent.

~4 min · Mission: better Prism/agent prompts

You already delegate to a fleet of Claude agents daily. ClaudeTag's whole design is one idea made explicit, and it's the idea worth stealing: delegation is a contract with four clauses.

The four clauses

When an admin sets up a ClaudeTag channel, they fill in exactly four things.Field guide §6 Read them as a template for any agent you brief:

ClauseIn ClaudeTagIn your subagent prompt
ObjectiveThe @Claude task you tag inThe one-line goal at the top of the prompt
ToolsAccess bundle: credentials, repo grants, pluginsThe tool allowlist you give the agent
InstructionsAccess bundle's instructions field (per channel)The system prompt / standing rules
BudgetMonthly spend cap (over-cap = declined, not partial)Effort/model tier + "don't escalate without cause"

The insight: an Access bundle's instructions field is just a per-channel system prompt, and its credential grants are the same least-privilege scoping your routing policy already demands ("never delegate risk-control to coder directly"). ClaudeTag didn't invent the contract — it gave it a UI.

Apply it to Prism Next time you brief a subagent, write all four clauses explicitly. Most flaky agent runs are a missing clause: no budget (it over-escalates), or vague instructions (it guesses the baseline). A ClaudeTag channel that "declines entirely rather than partially executes" over budget is the behaviour you want everywhere near live money.

Check yourself

Recall, don't peek — retrieval is what makes it stick.

Which clause most directly maps to a subagent's system prompt?

The Access bundle's instructions are the per-channel standing rules Claude follows — the system prompt. Tools/repos map to the allowlist; the spend cap maps to budget. (Field guide §6)

A ClaudeTag task exceeds its channel spend cap. What happens?

Over-cap work is declined entirely, not partially executed — a clean failure, not a half-finished one. That's the safe default for risk-adjacent automation. (Field guide §6)

Your win

You can now state the delegation contract from memory: objective, tools, instructions, budget. That's a checklist you can run against any agent prompt you write today.

Read this next Primary source, ~5 min: Introducing Claude Tag — Anthropic Newsroom. Then skim the full field guide.
Ask your teacher Want to go further? Ask me to run Lesson 2, where we draft a real Access bundle for a "prism-ops" Slack channel — scoped MCP tools, instructions, and a spend cap — and convert one of your existing subagent prompts into a clean delegation contract.