/prime-fetch

Wrap a noisy data-fetching tool call in a delegated sub-agent so the raw dump lands in the sub-agent's context, not yours. The lead receives only the filtered, proof-of-work-verified digest.

Structured execution spec: protocol.yon. Read it for the canonical rules and step sequence; this file is explanation. The two must stay in sync — if you edit one, update the other and refresh the @STAMP date.

Caller Options. Before executing, run the Caller Options protocol (caller-options/SKILL.md).

Family invariant — /prime-* skills delegate. Every /prime-* skill uses sub-agents to keep the caller's context clean. /prime-fetch and /prime-sweep additionally require proof-of-work on every finding (source:line + verbatim quote ≥8 words OR a verbatim-token). /prime-expand is exempt — it produces intent (a question + candidate sources), not factual findings.

What's a verbatim-token? A discrete identifier where ≥8 prose words isn't applicable. Examples: version strings (v1.2.3), commit hashes (abc123def), function/class names (MyClass.handleRequest), numeric values (port 8080), enum values (STATUS_PENDING). Must be quoted exactly; the lead still grep-verifies findability.

Boundary

Use /prime-fetch for one read-only data-fetching tool call (Grep, Glob, Read, WebFetch, or an MCP search) regardless of expected result size — the family invariant is about keeping noise out of your context, not about whether the result happens to be small. Do NOT use it for write-side calls (Edit, Write, Bash with side effects), multi-source investigation (use /prime-sweep — threshold ≥3 distinct sources OR ≥~10k tokens), or vague gaps that need articulating first (use /prime-expand). Do NOT use it to gather facts for the user — that is /investigate.

Step 1 — Receive the tool call spec + objective

Inputs: