web-researchSKILL.md · 4.3 KB
Research you can check.
Give it a topic and it comes back with a report where every claim carries the link it came from — at the point the claim is made rather than in a pile at the end, so checking one is a click rather than a search.
It stops searching when new searches stop changing the report, and then it writes down what it could not find. The second half is what the first half is worth anything on.
cp -r web-research ~/.claude/skills/What it does
Most of the work is deciding what not to write down.
Sources are weighed before they are used: whether the original is reachable behind the article describing it, how old the page is against how fast the topic moves, who benefits from the claim, and whether it shows any evidence at all. Content farms restating a press release are dropped rather than cited — the MacBook report says so and names what it used instead. Searching stops when the results are pages already read, not when a quota is met.
Then it cites as it writes, and leaves the disagreements standing. The 2026 survey puts Git's share of version control at ~85% from tech-detection aggregates and ~94% from developer surveys, keeps both, and says why neither is a measurement: the aggregators infer usage from job postings and site fingerprints, which over-weights companies with a public web presence. Where the best source is paywalled or years out of date it is cited and labelled as such rather than paraphrased from memory.
Every report then ends with the section most reports leave out. In the Docker one it is a table grading each claim area — Kubernetes release facts strong, taken off kubernetes.io directly; container-versus-VM performance percentages weak, because no controlled benchmark exists; cost and waste statistics weak, because every figure traces back to a vendor selling cost optimisation; Docker's pricing tiers unverified, because the live pricing page was never checked. The MacBook report names the fetches that failed instead: Tom's Hardware truncated, two benchmark sites returned 403, so one table rests on a single aggregator and says so.
Example reports
Three real runs, unedited — a little over 12,000 words between them, and every link in them is one the skill put there.