Recency-Grounded Research

Answering a question about a volatile subject by fanning sub-agents across the places people actually talk — Reddit, X, YouTube, Hacker News, forums, prediction markets — inside a bounded recency window, and returning ranked evidence clusters with verbatim quotes and the exact URL each came from. The name comes from the pattern's canonical instance, a /last30days-style skill; the window is the whole idea. What distinguishes this from ordinary web search is the output contract: not a fluent summary but a scored, ranked set of clusters, each anchored to a quote and a link, written to a durable markdown artifact.

The concept earns its place because of what it makes visible in an A/B. Run the same question with and without the skill, and the un-skilled model produces something confident, well-organised, and unfalsifiable — aggregate ratings, "the overall vibe is more conditional than it used to be," a recommendation. The skilled run produces "anyone generating PDFs from Zapier without it being a nightmare" with a score and a link. These are not a worse and a better answer to the same question. They are different artifacts: one is recall, the other is evidence. On a subject that moved after the model's cutoff — or moved last week — recall is not a degraded form of evidence, it is a category error, and fluency is exactly what makes it dangerous.

This is the demand side of Evergreen vs Volatile Context. That page says: don't ingest volatile data, keep it accessible and fetch on demand. This one says what "fetch on demand" has to look like to be worth anything — bounded window, named sources, verbatim quotes, links, ranking — and it is the same discipline firehose applies to itself: attribute, don't adjudicate; cite, don't launder.

Claims

On volatile subjects, a fluent un-sourced answer and a cited, dated one are different artifacts, not better and worse versions of the same one. principle — durable: the model's parametric recall cannot contain what happened after training, and its fluency is uncorrelated with whether it does. The failure mode is not "less accurate," it is "confidently unfalsifiable."
Grounding is what differentiates output once models converge — the delta in the A/B is the method, not the intelligence. principle — the same model produced both answers. See Execution Commoditization.
Bound the search by a recency window and make the window part of the request. best practice — context: subjects whose truth changes on a weekly-to-monthly cadence (product sentiment, tooling opinion, market speculation, anything post-cutoff). On evergreen questions the window is noise and costs you the best sources, which are usually old.
Search where the population complains, not where it reviews. best practice — context: sentiment and pain-point research, where the goal is unvarnished signal. The source's own comparison shows the un-skilled model reaching for aggregate review sites (G2, Capterra) while the skilled run surfaces forum complaints with quotes. Aggregate ratings are the right source when you want a population estimate, and the wrong one when you want a product idea.
Make the output contract citation-bearing: rank clusters, keep the verbatim wording, attach the URL, write it to a durable file. best practice — context: research feeding a decision or an artifact someone else will check. Ranking without quotes is opinion; quotes without links are unverifiable; both without a file are lost. The cost is a slower, uglier answer, which is the trade.
Pain points harvested this way are directly convertible into product and content decisions. observation — the source's demonstration: a top-scoring cluster ("people want a human approval layer for AI agents in Zapier workflows") reads as a roadmap item, and a third-ranked one ("generating PDFs without it being a nightmare") as a demo to build. What makes them actionable is that each is a real sentence someone wrote, not a synthesis.
The canonical instance fans out across Reddit, X, YouTube, TikTok, Instagram, Hacker News, Polymarket and the web, and reports in a few minutes. observation — the source states ~5 minutes for a sentiment sweep; the skill repo describes itself as researching "any topic across Reddit, X, YouTube, HN, Polymarket, and the web — then synthesizes a grounded summary." Point-in-time, groundable.

Evergreen vs Volatile Context — the ingest-side rule this concept completes. That page says keep volatile data out of the brain but reachable; this one specifies what reaching for it must produce.
Dynamic Retrieval — the same inversion (the system goes and gets it rather than the human remembering) applied to an external corpus rather than a personal store.
Contextual Retrieval — the retrieval-quality machinery a grounded answer depends on, one layer down.
Execution Commoditization — why the grounded/un-grounded delta is where value now sits.
Evidence-Gated Completion — the same "prove it, with something real" discipline, applied to research output rather than to an agent's claim of doneness. Ranked clusters with links are the receipt.
Frontier Scouting — recency-grounded sweeps are a scouting instrument: they tell you what changed, which is where new asks come from.
Public Skill Adoption — how this pattern actually reaches an operator: as a cloned repo, not a rebuilt workflow.
LLM-as-Judge — the scoring of evidence clusters is a judge in the loop; the score's meaning is only as good as the rubric behind it, which the source does not show.
Distillate: 160,000+ Cloned These 3 FREE AI Employees: Here's How (GitHub Claude Skills)

Recency-Grounded Research

Claims

Related

Linked from