/last30days skill for Claude Code, tested: can it read Reddit and Polymarket?
I compared 54 Claude Code reports with and without the /last30days skill. It doubles the hits on Reddit and Hacker News and reaches Polymarket, but uses 3.7 times more and gets fewer news facts right.

“Reddit: the search tool blocks the domain, so I have no Reddit threads and cannot say what is being said there.” Claude Code wrote that sentence in one of its reports (the report was in Spanish; the translation is mine), after I had asked it in writing to search Reddit, Hacker News, X and Polymarket as well as the press.
That gap is what /last30days promises to fill. It is a skill with 70,273 stars on GitHub on the day I tested it. Its README describes it as an agent-led search engine “scored by upvotes, likes, and real money - not editors”, and adds that Claude has none of that out of the box. This is the second episode of Useful or futile?, the series where I measure these promises instead of trusting the stars. The first one was Everything Claude Code.
A research skill needs different tasks
The five programming tasks from the first episode were no use here, because the skill would not even activate. I prepared six research assignments about news from 8 to 18 November, one for each use case the README advertises: a person, a company, a product comparison, a security incident, a rumor and an event.
To score without relying on my impression, before running anything I fixed a list of eight verifiable facts per topic: about five from the news, about two from the community (Hacker News threads) and one from a prediction market when there was one. I added three false but plausible claims per topic, to catch reports that make things up. A report passes if it includes five or more facts and none of the false claims. The grading is done blind by another model, Opus: it receives the report without the skill’s branding or footer and has to quote the exact sentence that supports each fact it accepts.
I compared three ways of doing the same assignment, always with Claude Sonnet in Claude Code and with web search enabled:
- Claude on its own.
- Claude with one sentence added to the prompt asking for those sources explicitly.
- Claude with
/last30days, installed without keys or cookies, which is the “zero configuration” mode in the README.
Each assignment was repeated three times per configuration, 54 reports in total.
What each configuration finds
| Claude on its own | Asking for the sources | /last30days | |
|---|---|---|---|
| Reports passed | 12/18 | 13/18 | 15/18 |
| News facts | 90% | 86% | 78% |
| Community facts | 28% | 36% | 62% |
| Market facts | 0/15 | 3/15 | 7/15 |
| Average usage per report | $0.17 | $0.22 | $0.63 |
| Average time | 73 s | 65 s | 358 s |
Usage is what Claude Code calculates at API prices. I ran everything on the subscription, so what matters is the ratio: 3.7 times more with the skill.
On what it promises, the skill delivers. It gets more than twice as many community facts right and it is the only configuration that brings back Polymarket odds taken from the market itself. Without it, Claude did not cite a single market in 18 reports. The market facts come from five markets repeated three times, so the figure says “it gets there or it does not” more than “how often”.
The difference is largest on topics that start in the community. In the rumor assignment, about whether a model had got worse after an update, the skill passed all three times and always found the original post and the Hacker News thread. Claude on its own passed once.
Asking for it in the prompt does not fix it
The middle configuration exists to answer the obvious objection: maybe asking Claude for those sources is enough. It is not. With the added sentence, the reports cited more links, but across all 18 there is not one reddit.com URL, and the agent did not open Reddit or Polymarket a single time. The odds it gave came from websites that comment on them. In the assignment about an IPO, all three reports repeated the same snapshot from a blog, four days older than the test.
The explanation lies in access, and I want to separate what I saw from what I infer. What I saw is that Claude’s search tool returns no Reddit results, according to its own reports, and that while I was preparing the list of facts Reddit answered direct requests with a 403. What I infer is that this is the limit, and that the wording of the prompt is not. The skill gets around it because it ships a Python engine that queries the public Reddit, Hacker News and Polymarket APIs by itself. I did not test whether a prompt with curl and the addresses of those APIs would match the result.
What gets lost along the way
The skill does not replace a news search, and the data make that clear. It gets 78% of the news facts right, against 90% for Claude on its own. In the assignment about a person it passed once out of three, when Claude on its own passed all three: the skill’s reports left out two central facts of the timeline and spent that space on forum reactions.
My interpretation is that the format causes it. The skill’s protocol requires quoting sentences from the community and passing along a footer with statistics, and that pushes the sequence of events out. It is an inference, because I have not tested the skill with a different output format.
Reviewing by hand, I also found an error the grader was not counting: one report attributed to an official document a date the document does not give. It was one case in 18, and Claude on its own made no similar mistake, but it is the kind of failure that means asking for a date and a primary source for every important piece of data.
The other cost is time. A report with the skill takes about six minutes and processes 8.5 times more input tokens, because before searching it reads four instruction files that add up to 143 KB. Having it installed, on the other hand, weighs nothing: the context at startup is the same with it as without it. Here it is the opposite of ECC, which doubled the context before starting.
Before installing it
I read the code before running it, as I recommend in what to check before installing a skill. I found no telemetry and no data sent to third parties. There are four things worth knowing, though:
- If you give it permission, it reads the X and Reddit session cookies from the browser, and for Chrome it asks the macOS Keychain for the key. That permission is requested by the model following its instructions, not by the code.
- It declares
allowed-tools: Bash, so while it is active it runs commands without asking. In my test there were 92. - The setup wizard installs
yt-dlpand an npm package by the author that is not in the repository. - With the default installation, every report is saved in
~/Documents/Last30Days.
I tested it without cookies, without keys and without that wizard. With them it can find more, especially on X, and the risks change too.
Verdict: useful, for what it promises
To find out what people think about something, or what probability a market gives it, /last30days provides what Claude on its own cannot get, and I have not found a way to replace it with a prompt. I would use it for that, knowing that each query uses almost four times more and takes six minutes.
For news I would not use it. Claude with its web search gets more right, takes a minute and misses fewer basic facts. And if the report is going to inform a decision, I would ask for the two things separately: the facts with their primary source on one side and the community reaction on the other.
The limits of the test are the usual ones in this series, plus a few of its own. Three repetitions per assignment is a small sample, and three of the six assignments ended in a tie. The list of facts includes community and market facts on purpose, because that is what the tool promises, and it has none from Reddit because I could not get access while preparing it. The grader is a model and I only reviewed the doubtful cases by hand. And all the topics were about technology and in English.


