research-with-sources
Findings traceable to something actually read, with "not found" reported as a result instead of filled in.
Save as .claude/skills/research-with-sources/SKILL.md in a repository, or under ~/.claude/skills/research-with-sources/SKILL.md to have it everywhere. The copy button takes the whole file, frontmatter included.
--- name: research-with-sources description: Research a question and report findings that are traceable to a source, with uncertainty marked rather than smoothed over. Use when asked to research a topic, compare options, check whether a claim is true, or gather background before a decision. --- # Research with sources The failure mode in research is a fluent summary that mixes what was found with what sounds right. It reads better than honest work and is worth less. Every factual claim in your output must be traceable to something you actually read, or marked as not verified. No exceptions. ## Before searching Write down the question in one sentence, and what kind of answer would settle it. "Is X faster than Y" is settled by a benchmark with stated conditions, not by three blog posts asserting it. If the question is ambiguous, resolve that first. Researching the wrong question thoroughly is the most expensive outcome available. ## While gathering - **Prefer primary sources.** Documentation over a tutorial about the documentation. The specification over a summary. The original study over the article reporting it. - **Check the date on everything.** In fast moving areas, a confident article from three years ago is often wrong now. Record the date beside each source. - **Note who is speaking.** A vendor comparing itself to a competitor is evidence of a claim, not evidence of a fact. - **Look for the disagreement.** If every source agrees, you may be reading copies of one source. Actively look for the counter argument. If you cannot find one, say so. - **Follow the citation.** When an article cites a number, find the thing it cites. Numbers mutate as they are repeated. ## Marking confidence Label every finding: - **Confirmed.** Stated directly by a primary source you read. Cite it. - **Reported.** Secondary sources agree, no primary source found. - **Contested.** Credible sources disagree. Give both positions. - **Not found.** You looked and could not establish it. Say where you looked. "Not found" is a real result and must be reported as one. Replacing it with a plausible answer is the single most damaging thing you can do here, because the reader cannot tell the difference. ## Never - Never invent or approximate a citation, a URL, a title, a date or an author. - Never present a number without its source and its date. - Never let a paraphrase become more certain than the original. If the source says "may reduce", do not write "reduces". - Never fill a gap with what is probably true. Mark the gap. ## When comparing options Build the comparison on criteria that come from the decision, not from whichever features happen to be documented. State the criteria first, then fill them in, and leave a cell empty rather than guessing. Note what each option is bad at. A comparison where one option wins every row is usually a sales page in disguise. ## Report back 1. The answer in one or two sentences, with its confidence. 2. The findings, each with a source and a date. 3. What is contested, and who holds which position. 4. What you could not establish, and where you looked. 5. What would settle the open points.
what it needs
Nothing. This skill is instructions only, so it works in any agent that reads SKILL.md with nothing installed and no account. Save the file and it is live in the next session.
the instructions, in full
Research with sources
The failure mode in research is a fluent summary that mixes what was found with what sounds right. It reads better than honest work and is worth less.
Every factual claim in your output must be traceable to something you actually read, or marked as not verified. No exceptions.
Before searching
Write down the question in one sentence, and what kind of answer would settle it. "Is X faster than Y" is settled by a benchmark with stated conditions, not by three blog posts asserting it.
If the question is ambiguous, resolve that first. Researching the wrong question thoroughly is the most expensive outcome available.
While gathering
- Prefer primary sources. Documentation over a tutorial about the documentation. The specification over a summary. The original study over the article reporting it.
- Check the date on everything. In fast moving areas, a confident article from three years ago is often wrong now. Record the date beside each source.
- Note who is speaking. A vendor comparing itself to a competitor is evidence of a claim, not evidence of a fact.
- Look for the disagreement. If every source agrees, you may be reading copies of one source. Actively look for the counter argument. If you cannot find one, say so.
- Follow the citation. When an article cites a number, find the thing it cites. Numbers mutate as they are repeated.
Marking confidence
Label every finding:
- Confirmed. Stated directly by a primary source you read. Cite it.
- Reported. Secondary sources agree, no primary source found.
- Contested. Credible sources disagree. Give both positions.
- Not found. You looked and could not establish it. Say where you looked.
"Not found" is a real result and must be reported as one. Replacing it with a plausible answer is the single most damaging thing you can do here, because the reader cannot tell the difference.
Never
- Never invent or approximate a citation, a URL, a title, a date or an author.
- Never present a number without its source and its date.
- Never let a paraphrase become more certain than the original. If the source says "may reduce", do not write "reduces".
- Never fill a gap with what is probably true. Mark the gap.
When comparing options
Build the comparison on criteria that come from the decision, not from whichever features happen to be documented. State the criteria first, then fill them in, and leave a cell empty rather than guessing.
Note what each option is bad at. A comparison where one option wins every row is usually a sales page in disguise.
Report back
- The answer in one or two sentences, with its confidence.
- The findings, each with a source and a date.
- What is contested, and who holds which position.
- What you could not establish, and where you looked.
- What would settle the open points.