prime-fetch

Delegates noisy read-only tool calls to a sub-agent and returns verified digests.

14|1|Updated Jun 14, 2026
One-click install
npx skills add https://github.com/allemaar/open-skills --skill prime-fetch-allemaar
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: prime-fetch
Source: https://github.com/allemaar/open-skills/tree/main/skills/prime-fetch
Command: npx skills add https://github.com/allemaar/open-skills --skill prime-fetch-allemaar

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Running a noisy data-fetching tool call (Grep, Glob, WebFetch, MCP search) directly floods your context window with raw output, most of which is irrelevant. This Skill keeps the raw dump inside a delegated sub-agent's context so only a filtered, verified digest reaches the caller. ## Core Features & Use Cases - Delegated execution: A single sub-agent runs the read-only tool call in its own context window, keeping the lead's window clean. - Proof-of-work verification: Every finding and unfollowed pointer must carry a source:line plus a verbatim quote (≥8 words or a verbatim-token), which the lead re-verifies with deterministic Grep calls before emitting. - Structured digest output: Findings, grounded unfollowed pointers, discarded noise categories, and gaps are returned; the raw dump is never relayed. - Use Case: You need to know where 'session' is handled in a large TypeScript codebase. Instead of running Grep yourself and reading hundreds of matches, you invoke /prime-fetch and receive only the verified relevant findings with exact citations. ## Quick Start Invoke /prime-fetch with the exact read-only tool call you want run and the objective it serves, then review the verified digest it returns.

Frequently Asked Questions about prime-fetch

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I run a Grep search without flooding my context window?

Delegate the Grep call to a sub-agent that executes it in its own context and returns only filtered findings. Each finding must include a source:line citation and a verbatim quote of at least eight words, which the lead verifies before emitting.

What tool calls can be delegated to a sub-agent for filtering?

Read-only calls such as Grep, Glob, Read, WebFetch, and MCP search tools can be delegated. Write-side calls like Edit, Write, or Bash with side effects are refused, since the skill is strictly read-only.

When should I use multi-source investigation instead of a single delegated fetch?

Use a multi-source sweep when the gap spans three or more distinct sources or roughly 10k tokens of material. A single delegated fetch is designed for exactly one tool call; it never spawns multiple sub-agents or iterative waves.

Can a sub-agent verify web fetch or MCP search results?

Yes. The lead materializes the cited excerpt to a temp file with Write, then runs Grep against that file to confirm the quote exists. Any finding whose quote cannot be found is rejected and never reaches the caller.

What are the limitations of delegating a single tool call to a sub-agent?

It is capped at one sub-agent, one wave, and depth one, so the sub-agent cannot spawn further sub-agents. If signal is missed, you must re-invoke with a refined call or escalate to a multi-source investigation skill.