AGENT SKILL

knowledge-grounding

Search the knowledge pool before answering, ground claims in retrieved sources and cite them, recognize ingest-worthy content and capture it with knowledge_create — don't fabricate when the answer is on record.

License 许可
Apache-2.0
最近更新
2026-11-05
Stars
★ 212
Forks
33

01 前言 Front Matter

name
knowledge-grounding
always
false
triggers
knowledge, search knowledge, what do we know, do we have, sources, cite, ingest, save to knowledge, bookmark, reference, look it up, according to
description
Search the knowledge pool before answering, ground claims in retrieved sources and cite them, recognize ingest-worthy content and capture it with knowledge_create — don't fabricate when the answer is on record.

03 SKILL.md 正文

Knowledge Grounding

Gideon has a knowledge pool — ingested documents, notes, bookmarks, and references the user has accumulated. When a question might be answered by what's already on record, search it first and ground your answer in what you find rather than answering from general assumptions or making something up.

Tools: knowledge_search (semantic/keyword search), knowledge_get (fetch a specific item by id), knowledge_create (add new content to the pool).

Search before answering

For any question that could plausibly be covered by the user's own material — their projects, decisions, docs, references, prior research — knowledge_search first. If relevant items come back, base your answer on them:

  1. knowledge_search with the user's question (or its key terms).
  2. knowledge_get the most relevant hits to read the actual content.
  3. Answer from the retrieved material, not from a guess.

This is the difference between a confident-but-wrong answer and a grounded one. If the user has the answer on record, use it.

Cite your sources

When your answer rests on knowledge items, say so — name the source (title/path/id) you drew from, so the user can trust and trace it. "Per your Q2-architecture note, …" beats an unattributed assertion. If you synthesized across several items, cite each. Citations also make it obvious when an answer is grounded vs. when you're reasoning beyond the record.

Don't fabricate when knowledge exists

If the pool plausibly contains the answer, searching is mandatory before you answer from memory. Never invent a fact, a number, a decision, or a citation when the real one is retrievable. And distinguish clearly:

  • Found it → answer from the source, cite it.
  • Searched, found nothing → say the knowledge pool doesn't cover it, then answer from general reasoning labelled as such (not presented as if it came from the user's records).

Don't dress up a guess as a recalled fact.

Recognize ingest-worthy content

Some content is worth keeping. When the user shares — or you produce together — something durable and reusable, capture it with knowledge_create:

  • text — a note, a decision record, a distilled summary, research findings, a snippet of reference material worth keeping.
  • bookmark — a URL/reference the user will want to find again.

Good candidates: "save this for later", a link the user clearly wants to keep, a conclusion reached after real research, reference material that answers a recurring question. Capture it so the next session can knowledge_search and find it.

Don't ingest: transient chatter, secrets, throwaway scratch, or content the user didn't signal is worth keeping. (For named, versioned generated UI/docs the user iterates on, use the artifacts skill instead; for durable user preferences and corrections, use memory-discipline. Knowledge is the searchable reference pool.)

04 同仓库更多 Skills

Sidiora-Labs
Sidiora-Labs/centra-gideon-agent

web-verify

Verify a built web page or app actually renders and works — drive it in a real browser, check the console and network, confirm the thing you changed is the thing being served, and capture evidence. Use before claiming a frontend change works.

★ 212
Sidiora-Labs
Sidiora-Labs/centra-gideon-agent

visual-output

Emit rich visual output in chat via <widget> tags — styled HTML cards, charts, tables, interactive tools, and drawn SVG/canvas diagrams, illustrations, and animations — theme-aware, centered, and responsive instead of clashing or cramped.

★ 212
Sidiora-Labs
Sidiora-Labs/centra-gideon-agent

task-and-project

Track work as Tasks (the /api/tasks CRUD + native task provider), and decompose larger projects into independent, verifiable units — fanning parallel investigation out to subagents while keeping one coherent plan; hand open-ended goal-driven work to a goal loop.

★ 212
Sidiora-Labs
Sidiora-Labs/centra-gideon-agent

research-campaign

Run a multi-cycle research campaign — grill the question, decompose it into answerable sub-questions, investigate each against sources, then synthesize with confidence and gaps stated. Use for open questions that need real investigation rather than one lookup.

★ 212
Sidiora-Labs
Sidiora-Labs/centra-gideon-agent

memory-discipline

When and how to persist durable memory — save genuine user preferences, facts, and corrections; NEVER persist transient environment failures (a failed command, a flaky network call, a one-off tool error) as lessons; recall before re-asking.

★ 212
Sidiora-Labs
Sidiora-Labs/centra-gideon-agent

loop-worker

Per-cycle protocol for an autonomous goal loop worker — the unified autonomous goal engine that runs one self-directed cycle per turn until its goal is met, then self-retires. Drives one cycle against a loop's file interface (status.json gate, brief.md goal, guidance.txt nudges, findings/cycle_NNN.json, FINDINGS.md, the type-appropriate deliverable, STOP sentinel). The worker produces work and reports evidence; a deterministic check or a separate judge — never the worker — decides done-ness.

★ 212