Most AI prompts are stale before they start. They rely on the model's training data, your memory, or a few links you remembered to paste.
Matt Van Horn's Last 30 Days skill is interesting because it attacks that exact problem. It gives Claude Code and other agent hosts a way to research what people have been saying recently across Reddit, X, YouTube, Hacker News, Polymarket, GitHub, and the web.
Source note
Credit for the interview goes to Greg Isenberg and Idea Browser. The guest is Matt Van Horn, creator of the mvanhorn/last30days-skill repo.
This article uses the video and transcript as the workflow source, then checks the current GitHub repo so the setup notes reflect the latest public install path. The repo is moving quickly, so always read its README and runtime skill spec before installing.
Repo research update
I also checked the live GitHub repo during implementation on June 27, 2026. At that point the project had crossed roughly 47k stars and 3.9k forks, and the latest public release listed on GitHub was v3.8.3, published on June 25, 2026.
That matters because the video explains the concept, but the repo is the operational source of truth. The README says the current v3 pipeline is tracked there, while the runtime behavior lives in skills/last30days/SKILL.md.
| Source | Useful signal | JQ AI SYSTEMS caution |
|---|---|---|
| Unfiltered comments, objections, upvotes, niche pain language. | Strong for voice of customer, but noisy and sometimes wrong. | |
| X / Twitter | Fast reactions, expert threads, launch chatter, hooks people repeat. | Engagement is not truth. Watch for hype loops. |
| YouTube | Long-form demos, transcripts, creator explanations, practical examples. | Great for workflow learning; verify product facts elsewhere. |
| Hacker News | Developer consensus, objections, technical pushback. | Useful for engineering products, less representative of mainstream buyers. |
| Polymarket | Market odds and prediction signals backed by money. | Not a substitute for research or legal/financial judgment. |
| GitHub | Repo velocity, stars, releases, issues, pull requests, developer adoption. | Stars can be hype; check commits, issues, and install friction. |
The important research update: Last 30 Days is no longer just a one-off skill from a video demo. It is a fast-moving public repo with releases, install paths, update commands, and a defined skill spec. That makes it more credible, but it also means readers should check the latest README before copying old setup steps from any article or video.
What Last 30 Days is
Last 30 Days is a reusable agent skill. Instead of asking Claude Code to "research this topic" from scratch, you ask it to use a structured recent-research workflow.
The current repo describes it as an AI-agent-led search engine scored by signals from people: upvotes, likes, comments, transcripts, odds, GitHub activity, and recent public web content.
That is useful because a lot of practical AI work depends on live language:
- What objections are prospects using right now?
- Which cold email frameworks are being shared this month?
- What landing pages are designers praising this week?
- What product category is trending on X or Reddit?
- What competitor use cases are customers actually discussing?
The model still reasons, writes, and builds. Last 30 Days gives it fresher raw material.
Why current context matters
A stale model can still write a decent first draft. But it will often miss the thing that makes work commercially useful: the current mood of the market.
Matt's core point in the episode is simple: online conversations move too fast to track manually. If you want Claude Code to help with cold email, audience growth, product research, or landing-page design, it needs the recent examples people are already reacting to.
That is the difference between:
- "write me cold emails for this product"; and
- "research cold email frameworks people praised in the last 30 days, identify the patterns, then write a version for this exact audience."
The second prompt has a better chance because it starts from current examples, not generic advice.
Setup reality
In the video, Matt talks through the API setup needed for richer data sources such as X and Reddit-style research. The current GitHub README has a slightly cleaner message: some sources work immediately, and the setup wizard unlocks more.
Current Claude Code install path from the repo:
/plugin marketplace add mvanhorn/last30days-skill
/plugin install last30daysCurrent Agent Skills install path for Codex, Cursor, Copilot, Gemini CLI, and other hosts:
npx skills add mvanhorn/last30days-skill -gCurrent update commands from the repo:
npx skills update last30days -g
npx skills update -gCurrent list/remove commands:
npx skills list -g
npx skills remove last30days -gThe repo says Reddit, Hacker News, Polymarket, and GitHub work immediately. X, YouTube, TikTok, Instagram Reels, and additional sources may need login, local tooling, APIs, or setup wizard steps depending on your environment.
What Matt demos
Popular rap songs
Matt starts with a simple demo: ask Last 30 Days to find what rap songs are popular right now. The point is not the music result. The point is that Claude Code can ask the internet a current question and synthesize recent signals.
Cold email frameworks
The cold email section is the most business-useful part of the episode. Matt uses the skill to research current high-performing cold email frameworks, then turns the patterns into outreach copy.
The examples include AIDA, Praise-Picture-Push, and intent-data-trigger style angles. The better version of this workflow is not "copy a viral template." It is: research what is working, extract the pattern, adapt it to your offer, then review it like a human.
Growing an X following
Matt also uses recent X data to study what content is getting traction. This is where Last 30 Days becomes a distribution tool: it can look at recent examples before drafting hooks, content angles, and positioning.
Researching Moltbot to build a competitor
The bigger demo is product research. Matt researches Moltbot, then uses that context to help plan and build a competitor-like product. This is the workflow builders should pay attention to: research first, plan second, code third.
Web design inspiration
Matt asks what website designs are getting praised right now, then turns the research into a Figma-ready prompt and AI image/design direction. That is a strong use case because design taste is context-heavy. The agent needs examples before it invents.
The research-to-build loop
The useful loop looks like this:
- Ask Last 30 Days for recent signals around a topic.
- Have Claude Code summarize patterns, objections, examples, and language.
- Turn the research into a spec, prompt, landing page, email, or product plan.
- Use a builder workflow, such as Claude Code or Compound Engineering-style planning, to create the first implementation.
- Review the output against sources, not vibes.
That is the real unlock. The skill is not only for learning. It can become the first step in product, content, outreach, and design workflows.
Best practices
If you install Last 30 Days, I would use it with these habits:
- Ask for citations and source grouping. Separate Reddit, X, YouTube, HN, Polymarket, GitHub, and web results.
- Ask for patterns, not just summaries. You want repeated language, objections, hooks, features, and claims.
- Ask what changed recently. The value is recency, so make the agent compare old assumptions with new signals.
- Give it a business target. A research brief for a cold email is different from a research brief for a SaaS landing page.
- Keep the build step separate. Do not let the agent jump from raw research into code without a spec.
- Verify before publishing. Recent does not mean true.
My favorite prompt shape is:
Use Last 30 Days to research [topic] for [audience].
Separate findings by source.
Extract repeated language, objections, examples, and current trends.
Then propose 5 practical angles I could use for [email / landing page / product spec / content].Why this matters for non-engineers
Matt says something important in the episode: he is not approaching this like a traditional engineer. His workflow is closer to a builder-operator workflow.
He uses Claude Code in the terminal, screenshots errors, pastes them into ChatGPT for debugging help, and keeps iterating. That sounds messy, but it is also how a lot of non-engineers are starting to ship useful software.
The key is not pretending the agent is magic. The key is learning a repeatable rhythm:
- research with current context;
- turn research into a plan;
- build a narrow version;
- debug with screenshots and logs;
- ship only after review.
Last 30 Days helps most at the first step. It makes the plan less generic.
Caveats
There are a few places to be careful:
- API keys: X, OpenAI, xAI, browser tools, or other sources may require keys or login. Store them safely.
- Platform terms: If a workflow touches X, Reddit, YouTube, TikTok, or Instagram, review the relevant terms and rate limits.
- Source bias: Social engagement is not truth. It is attention.
- Privacy: Do not run client data, private strategy, or confidential lists through a toolchain you have not reviewed.
- Overbuilding: Research can become procrastination. Use it to build a sharper first version, not a 90-page plan.
This is not a reason to avoid the skill. It is a reason to use it like a serious workflow component.
JQ AI SYSTEMS checklist
I would test Last 30 Days with one of these narrow jobs:
- Research the last 30 days of buyer objections for your offer.
- Find current cold email patterns for one specific niche.
- Study landing pages people praised this month and extract design patterns.
- Research a competitor and turn it into a feature map.
- Find what people are complaining about in a niche and turn it into a micro-SaaS idea list.
- Build a weekly research brief for your business using only public sources.
The simplest first test: ask it to research a market you already understand. If the output teaches you something new and cites where it came from, it earns a place in your stack.
CTA: Before asking Claude Code to build, ask it to research what people cared about in the last 30 days. Current context is often the difference between a generic build and a useful one.