grounded-research
Skill0 starsChanged 6 months agoApache-2.0View on GitHub
Enforces source-grounded, citation-first research discipline. Use this skill whenever the user asks to research a topic, fact-check a claim, summarize recent findings, look up current information, explain what the evidence says about something, or produce any factual output where accuracy and verifiability matter. Trigger even on conversational phrasing like "what's the deal with X", "is it true that Y", "what do experts say about Z", or "can you check if..." — if the answer requires facts that could be wrong or outdated, this skill applies. When in doubt, use it.
5 other projects use a version of this.
Adopted by
Most recently taken up by
- PrisacariuRobert/sidematestheir own version, first seen Nov 8, 2026
- berkeley-humanoids/skillstheir own version, first seen Oct 9, 2026
- berkeley-humanoids/Skillstheir own version, first seen Oct 9, 2026
- Kaagemusha/agent-skillstheir own version, first seen Sep 27, 2026
- masbrokemanaaja/reethinktheir own version, first seen Sep 20, 2026
Other versions5 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 5
- PrisacariuRobert/sidemates★ 9same name
- berkeley-humanoids/skills★ 4same name
- berkeley-humanoids/Skills★ 4same name
- masbrokemanaaja/reethink★ 3same name
- Kaagemusha/agent-skills★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
The file
Show its rules and text
We split this skill into the rules it gives an agent
25 rules. 0 are common, 24 are distinctive, and 1 is very specific to this project.
0
Common
Most of the 5 edited versions we read keep these.
24
Distinctive
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 23
Before producing any output, mentally classify every claim the response will make:
- Distinctive: line 32
Never mix types silently. If a paragraph contains a Type C claim presented with Type A confidence, that's a hallucination risk.
- Distinctive: line 47
Only after verification is complete for all B and C claims, proceed to Phase 3.
- Distinctive: line 50
Do not state a statistic without a source, even if you are "pretty sure"
- Distinctive: line 51
Do not state someone's current job title, role, or status without searching
- Distinctive: line 52
Do not describe a product's current features or pricing without searching
- Distinctive: line 53
Do not summarize a study's findings without finding the actual study or a primary report
- Distinctive: line 65
Every Type B and C claim gets an inline source reference immediately after it.
- Distinctive: line 67
Never reproduce long quotes — paraphrase and attribute.
- Distinctive: line 76
Use these inline markers when precision matters:
- Distinctive: line 88
If a claim cannot be verified and is important to the answer, say so explicitly:
- Distinctive: line 98
After drafting the response, do a final pass:
- Distinctive: line 102
Check that no Type D (analytical) claim is presented as if it were Type A (established fact)
- Distinctive: line 103
Verify the response does not contradict itself across sections
- Very specific: line 112
Read references/source-quality-tiers.md for the full tier guide.
- Distinctive: line 120
When sources conflict, report the conflict.
- Distinctive: line 128
Add: "This situation may have developed since [date/time of source]"
- Distinctive: line 129
Do not speculate about what "probably" happened after your last source
- Distinctive: line 132
Prefer CVE databases, vendor advisories, NIST, and official security bulletins
- Distinctive: line 133
Do not paraphrase vulnerability details in ways that could introduce inaccuracy
- Distinctive: line 134
If the claim involves a specific version, configuration, or exploit, find the primary advisory
- Distinctive: line 137
Always source to peer-reviewed literature, official guidelines, or regulatory bodies
- Distinctive: line 138
Always add: "This is not medical/legal advice.
- Distinctive: line 139
Never state a treatment or legal position as settled if the primary literature shows debate
- Distinctive: line 166
Always prioritise precision over comprehensiveness.
“Common” is measured: kept by at least 6 in 10 of the 5 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Grounded Research Skill
This skill enforces a structured, citation-first discipline to produce outputs where every factual claim is either verified, explicitly uncertain, or openly declared unknown. The goal: no confident-sounding hallucinations. Ever.
Phase 1 — Claim Taxonomy (do this before writing anything)
Before producing any output, mentally classify every claim the response will make:
| Type | Description | Handling |
| A — Stable | Well-established, timeless facts (laws of physics, historical events pre-2020, definitions) | State directly, no search needed |
| B — Time-sensitive | Current status, prices, positions, policies, recent events, statistics that change | Must search before stating |
| C — Domain-specific | Technical claims, research findings, medical/legal/security facts, statistics | Must find a source before stating |
| D — Analytical | Opinions, comparisons, interpretations, synthesis | Label clearly as analysis/opinion |
Never mix types silently. If a paragraph contains a Type C claim presented with Type A confidence, that's a hallucination risk. Classify first, then write.
Phase 2 — Pre-Response Verification Loop
For every Type B and Type C claim identified in Phase 1:
- Search before writing — run a web search targeting that specific claim
- Extract the key fact — find the specific sentence or data point that supports it
- Note the source tier — see
references/source-quality-tiers.md - Flag if unverifiable — if no reliable source is found after 2 searches, the claim must be either dropped or explicitly marked as unverified
Only after verification is complete for all B and C claims, proceed to Phase 3.
Verification shortcuts to never take:
- Do not state a statistic without a source, even if you are "pretty sure"
- Do not state someone's current job title, role, or status without searching
- Do not describe a product's current features or pricing without searching
- Do not summarize a study's findings without finding the actual study or a primary report
Phase 3 — Structured Output
Structure all research responses as follows:
3a. Lead with the answer
Give the direct answer first. Do not bury it in preamble.
3b. Inline citations
Every Type B and C claim gets an inline source reference immediately after it. Format: state the claim in your own words, then note the source naturally in prose. Never reproduce long quotes — paraphrase and attribute.
Example (correct):
The ECB held rates steady in April 2026, according to its official press release.
Example (wrong):
The ECB held rates steady in April 2026. [This is stated confidently with no source]
3c. Confidence signals
Use these inline markers when precision matters:
[High confidence — primary source]— sourced from official/primary publication[Medium confidence — secondary source]— sourced from reputable but secondary coverage[Low confidence — unverified]— couldn't find a strong source; treat as uncertain[Analytical — my interpretation]— this is synthesis/opinion, not a sourced fact
You do not need to add these markers to every sentence — use them where the confidence level is non-obvious or where a reader might otherwise mistake analysis for established fact.
3d. "I don't know" declarations
If a claim cannot be verified and is important to the answer, say so explicitly:
I could not find a reliable source for [X]. This should be verified before acting on it.
This is always better than a confident-sounding guess.
Phase 4 — Confidence Audit
After drafting the response, do a final pass:
- Scan for any factual claim that lacks a source or uncertainty marker
- For each: either add attribution, add an uncertainty signal, or remove the claim
- Check that no Type D (analytical) claim is presented as if it were Type A (established fact)
- Verify the response does not contradict itself across sections
This audit step is not optional. It's the difference between a grounded response and a plausible-sounding one.
Phase 5 — Source Quality
Read references/source-quality-tiers.md for the full tier guide.
Quick reference:
- Tier 1 (strongest): Official publications, primary studies, government data, company IR
- Tier 2: Major wire services, peer-reviewed journals, established specialist press
- Tier 3: Quality general news, well-sourced blog posts from known experts
- Tier 4: Aggregators, wikis, forums — only as pointers to Tier 1–3, never as the final source
When sources conflict, report the conflict. Do not silently pick one.
Special Cases
Rapidly changing topics (breaking news, live events)
- State the timestamp of your most recent source explicitly
- Add: "This situation may have developed since [date/time of source]"
- Do not speculate about what "probably" happened after your last source
Technical / security claims
- Prefer CVE databases, vendor advisories, NIST, and official security bulletins
- Do not paraphrase vulnerability details in ways that could introduce inaccuracy
- If the claim involves a specific version, configuration, or exploit, find the primary advisory
Medical / legal claims
- Always source to peer-reviewed literature, official guidelines, or regulatory bodies
- Always add: "This is not medical/legal advice. Consult a qualified professional."
- Never state a treatment or legal position as settled if the primary literature shows debate
Historical research (pre-cutoff, no search needed)
- Type A claims can be stated without search
- But if the user is researching something where recent scholarship might have updated the consensus (e.g. archaeology, historical attribution, economic history), flag it
Anti-Patterns to Avoid
| Anti-pattern | Why it's a problem |
| "Studies show..." with no citation | Fabricated authority |
| "It is widely believed that..." | Vague appeal that hides a missing source |
| "As of my last update..." followed by a specific fact | Framing stale knowledge as verified |
| Summarising a document you haven't read | High hallucination risk on specific details |
| Quoting statistics without a source | Numbers are the most frequently hallucinated element |
| Confidently resolving a contested claim | Misrepresents genuine expert disagreement |
Output Length Guidance
- For quick fact checks: 2–4 sentences + source. Do not pad.
- For research summaries: structured prose, one section per sub-topic, sources inline.
- For contested topics: present multiple positions with their respective sources. Do not collapse to one view unless the evidence is overwhelming and sourced.
- Always prioritise precision over comprehensiveness. A shorter, accurate answer beats a longer, partly-hallucinated one.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

