Gather

Gather: Structured Information Collection

You are running the gather primitive — collecting information on a topic into structured findings with sources and confidence levels. Topic: $ARGUMENTS

When to Use

• When starting research on a new topic or question
• Before distilling, ranking, or verifying — gather is the universal input generator
• When refining or expanding findings from a prior primitive (gather can compose with itself)
• When the user asks you to "find", "investigate", "research", or "collect" information

Process (default: serial)

1. Assess Source Breadth

Before searching, identify the distinct source types the topic spans:

• Codebase — implementation files, configs, tests
• Web/docs — external documentation, references, changelogs
• Settings/config — environment variables, deployment config, project-level settings

If the topic clearly spans 3 or more of these source types, consider switching to Parallel Collection Mode (see below) to gather concurrently. Otherwise, proceed serially.

2. Search Code First

Use Grep, Glob, and Read to search the codebase for relevant information. Code is faster and more reliable than web search.

3. Search Web Second

If codebase results are insufficient or the topic requires external context, use WebSearch and WebFetch to gather additional information.

4. Emit Structured Findings

Output in pipe format. Each finding must have a source: and confidence: per-item metadata.

Parallel Collection Mode

Trigger: When the topic clearly spans 3 or more distinct source types (codebase files, web/external docs, and config/settings) and the breadth warrants concurrent collection.

Dispatch per the fan-out-protocol rule (loaded alongside this skill). Gather-specific details:

Agents (dispatch 2-3, one per source type):

Agent	Scope	Prompt focus
Codebase	Implementation files, tests, inline docs	"Search the codebase for [topic]. Use Grep, Glob, Read."
Web/docs	External documentation, references, changelogs	"Search the web for [topic]. Use WebSearch, WebFetch."
Config	Env vars, deployment config, project settings	"Search for config related to [topic]. Look for .env, YAML/JSON, manifests."

All agent prompts must end with: "Return findings as a numbered list: title, detail, source (file:line or URL), confidence (CONFIRMED/LIKELY/POSSIBLE)."

Merge: Deduplicate overlapping results, higher-confidence items take precedence. Emit as a single pipe-format block with **Mode**: parallel (N agents) after the Pipeline line.

Modes

This primitive has one implicit secondary behavior:

Behavior	Description
Confidence labeling	Every finding gets a CONFIRMED/LIKELY/POSSIBLE label. This is a lightweight /assess operation embedded in gather's output.
Code-first priority (default)	Search order: codebase first, web second. This is a domain bias optimized for software engineering, not a user-configurable parameter.
Parallel collection (opt-in)	When 3+ distinct source types are identified, dispatch concurrent agents per source type and merge results.

Confidence labeling is embedded by design — separating it would require a /gather -> /assess chain for every collection, adding overhead without value. The code-first priority reflects the target domain; for non-code topics, web results will still be used when codebase results are insufficient.

Guidelines

• Code sources are more reliable than web sources — prioritize codebase over web
• If composing with upstream pipe-format output, refine or expand those findings
• Keep findings concise — the distill primitive can synthesize later
• If the topic is vague or broad, gather what you can and note scope limits in the summary
• In parallel mode, cap agent count at 3 to avoid API throttling — merge any additional source types into the closest existing agent