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:
knowledge_searchwith the user's question (or its key terms).knowledge_getthe most relevant hits to read the actual content.- 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
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.
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.
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.
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.
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.
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.