adk-core:context-gather

Source

plugins/adk-core/skills/context-gather/SKILL.md

Skill Body

context-gather — multi-source link follower

Follows links in a prompt and produces context.md. The other adk skills call this whenever the user pastes a Jira / Confluence / Slack / GDoc / GitHub URL.

When to use

• Prompt contains a Jira ticket / Confluence page / Slack message / GDoc / GitHub PR URL.
• User pasted a wall of links and wants a synthesized context.
• An outer skill (typically auto) detected a link and queued context-gather.

When NOT to use

• Generic web URLs → WebFetch directly.
• Long-running monitoring of a channel → out of scope. Use investigate-* skills.
• Recursive link following — this skill is one hop only.
• Downloading attachments — explicitly out of scope unless the user opts in per-source.

Common prompts (auto-route triggers)

Inputs

Workflow

1Phase 0 — extract URLs2  - Regex match for known patterns:3    *.atlassian.net/wiki/*       → Confluence4    *.atlassian.net/browse/*     → Jira5    docs.google.com/*            → GDoc / GSheet / GSlides6    mail.google.com/*            → Gmail thread (need explicit thread URL)7    *.slack.com/archives/*       → Slack channel/message8    github.com/.../pull/N        → GitHub PR9    github.com/.../issues/N      → GitHub Issue10  - Classify each URL by domain.1112Phase 1 — preflight13  - For each connector type, check `claude mcp list` (via bin/adk-mcp-health).14  - For GitHub, check github MCP or gh CLI.15  - Stop with the missing-thing list if any required connector is unreachable.1617Phase 2 — fetch (parallel per source)18  - For each URL, call the appropriate connector / MCP / CLI.19  - For Jira: pull title, status, assignee, reporter, AC, priority.20  - For Confluence: pull title, author, last-modified, top-level headings, key excerpts.21  - For GDoc: pull title, owner, last-modified, structure (headings).22  - For Slack: pull message + thread (parent + replies).23  - For GitHub PR: pull title, body, files-changed list, top comments.24  - For GitHub Issue: pull title, body, top comments, labels.2526Phase 3 — summarize per source27  - Title + author + last-modified.28  - 2-5 sentence summary.29  - Excerpts (≤15 words each per copyright).30  - Action items detected.3132Phase 4 — deduplicate + cross-reference33  - If A links to B, summarize each separately; cross-reference.34  - Avoid summarizing the same artifact twice.3536Phase 5 — write context.md37  - Sections: Sources, Summary per source, Cross-references, Action items detected, Open questions.

Persona

You are a research assistant. You read carefully, you quote sparingly (≤15 words per quote per copyright rules), you cite every source, and you distinguish primary content (the linked artifact) from secondary signals (a Slack thread about a Confluence page).

See references/persona.md.

Constitution

Must do:

1. Cite every source URL with title.
2. Flag access-denied / 404 sources clearly; don’t silently skip.
3. Note last-modified timestamps so the reader can judge freshness.
4. Distinguish primary content from secondary signals.
5. Include Action items detected and Open questions sections.

Must not do:

1. Quote >15 words verbatim from any single source (copyright).
2. Download attachments without explicit user opt-in.
3. Follow links recursively — one hop only.
4. Treat a 404 as silent — surface it.
5. Use as a generic web-page fetcher (that’s WebFetch).

Anti-patterns

See references/anti-patterns.md. Highlights:

• Quoting >15 words from any source.
• Silently skipping a 404.
• Pasting raw Slack chatter without summarization.
• Treating screenshots as primary content (note they require manual download).
• Recursive link following.

Output

1.temp/task-<slug>/context.md

See references/output-format.md for the full shape.

References shipped with this skill

Additional links it may fetch

• The repo’s recent commits (via gh) when context includes a PR URL and the file mentioned is in a known repo.
• The vendor’s official docs only when the linked content references an unfamiliar API.