Source: raw/9_Free_AI_Agent_Skills_You_NEED_to_Install_Now.md. Repo names, stars and licences were checked against the GitHub API on 2026-10-27, because the transcript doesn’t give the links.
Creator: Futurepedia (YouTube @futurepedia_io) URL: https://www.youtube.com/watch?v=3fdb_giOrLo Duration: not in the transcript header Platform: YouTube Published: 2026-09-28
Futurepedia’s pick of nine free skills, each demonstrated on a real task: research, video creation, a security check before installing, de-slopping writing, watching video, and three skill packs for building software and better-looking sites. He installs them across ChatGPT Work, Claude Code, Cowork and Codex to show that one skill format runs in all of them. The useful parts are the install-by-prompt habit, scanning a skill before you install it, and treating a downloaded skill as a starting point to customise. Most of these skills already have their own wiki articles, linked below. The one the wiki lacked is humanizer.
Key Takeaways
- Install any skill by prompt. Copy the GitHub repo link and say “install this skill from GitHub using its setup instructions.” Futurepedia uses the same sentence in ChatGPT Work, Claude Code, Cowork and Codex. You don’t need a slash command or
@mention to invoke a skill. Naming the tool in the prompt (“use Remotion to…”) is enough, and some skills fire from context alone. - Scan before you install. He runs NVIDIA’s skill scanner, which the captions render as “Skill Inspector”, as a skill of its own: call it with the next skill’s repo link and read the report. His calibration is that it sometimes flags minor things he installs anyway, and that a clean report “isn’t an absolute guarantee.” The verified repo is
NVIDIA/SkillSpector(see SkillSpector). - A downloaded skill is a starting point. He ran last30days, had Codex pour the leftover findings into a site he’d built (use cases, unconventional takes, embedded original posts), and then asked it to “package everything up as a new expanded skill.” Customising a skill to your own setup is the step he says most people skip.
- Remotion and HyperFrames are good at different jobs. He gave both the same two prompts. Remotion’s icon-expansion motion graphic was the one he’d “use exactly as is.” For a 15-second Utah road-trip map, HyperFrames gave the smoother zoom and better camera angle, even though Remotion has a dedicated
remotion-mapsskill (the skill exists inremotion-dev/skills). Remotion renders with React and HyperFrames with HTML and web animation. His advice is to install both and test each on your own use case. - Humanizer fixes AI tells in writing. Three default Claude tweets about Jev came out with “not X, it’s Y” lines and slogan closers. “Fix those up using humanizer” rewrote them and listed what changed. He says it pays off more on long-form writing than on tweets.
- The video skill handles links a plain chat can’t. Given an Instagram camping-meal link, the skill produced a full recipe, including timings that only appeared on screen. ChatGPT without the skill couldn’t open the Instagram link at all. With the uploaded file it read the frames but skipped the transcript. The skill also downloads a video itself when a site blocks viewing.
- Superpowers and gstack overlap but don’t replace each other. Both are skill packs for building software. Superpowers covers brainstorming, plans, implementation, testing and debugging, and he often just sees it fire mid-task. gstack adds product-idea challenges, plan review, design, code review, browser QA, bug fixes and releases. He runs both.
- Impeccable is for AI-looking design. On a one-prompt “Utah National Parks Passport” site,
bolderadded stronger type, deep forest-green sections and larger photos.delightadded a stamp-press animation when you mark a park as visited.critiquereturned a full design, usability and responsiveness review. He counts 24 commands at recording. The wiki’s v3 article lists 23.
The nine, verified
Star counts from the GitHub API on 2026-10-27. The video gives no counts.
| # | Skill | Repo | Stars | Licence | What Futurepedia used it for | Wiki article |
|---|---|---|---|---|---|---|
| 1 | last30days | mvanhorn/last30days-skill | 65.7k | MIT | Five-minute research brief on Jev, drawn from 14 Reddit threads, 13 Hacker News stories and GitHub. X and YouTube are optional sources you can add at setup | last30days-skill |
| 2 | Remotion skills | remotion-dev/skills | 5.8k | not stated in the API response | Icon-expansion motion graphic; remotion-maps road-trip animation | Remotion Motion Graphics |
| 3 | HyperFrames | heygen-com/hyperframes | 73.9k | Apache-2.0 | Same two prompts, for comparison | HyperFrames |
| 4 | SkillSpector (“Skill Inspector” in the captions) | NVIDIA/SkillSpector | 22.7k | Apache-2.0 | Scanning humanizer before installing it, in Cowork | SkillSpector |
| 5 | humanizer | blader/humanizer | 61.5k | MIT | Rewriting three AI-sounding tweets | this article, below |
| 6 | claude-video (/watch) | bradautomates/claude-video | 19.4k | MIT | A recipe from an Instagram cooking video, run in Codex | claude-video |
| 7 | Superpowers | obra/superpowers | 306k | MIT | A development playbook that fires automatically | Superpowers |
| 8 | gstack | garrytan/gstack | 138k | MIT | Idea challenge, plan review, QA, releases | gstack |
| 9 | Impeccable | pbakaus/impeccable | 90.5k | Apache-2.0 | bolder, delight and critique on a demo site | Impeccable v3 |
Humanizer (blader/humanizer)
The wiki had no article for this skill. Details below come from the repo README, read 2026-10-27.
- What it does: rewrites AI-written text so it reads as if a person wrote it, without changing what it says. It is built on Wikipedia’s “Signs of AI writing” guide and checks 26 patterns. The README names five as the strongest tells, each enough to justify an edit on its own: “not X but Y”, one-line closers, sayings that sound deep, a staged run-up (“Here’s the thing.”), and arguing with no one.
- It doesn’t invent facts. Names, numbers, dates and quotes must come from the source. If a sentence needs a missing detail, it asks for it. On a file it changes only the prose and leaves code, data, frontmatter and link targets alone.
- Voice matching: paste two or three paragraphs of your own writing before the text and it follows their rhythm, word choice and punctuation.
- Not a detector bypass. The README says beating AI detectors isn’t the goal, and detectors still flag most of its output. Its blind-test claim (16 of 16 judges preferred the rewrite) is the maintainer’s, from issue #229, and hasn’t been checked independently.
Try It
- Vet, then install. Install SkillSpector, run it on the next skill you plan to add, and read the report before installing. For a private skill, use the offline
--no-llmscan described in SkillSpector. - De-slop a real draft. Install humanizer (
/plugin marketplace add blader/humanizer, then/plugin install humanizer@humanizerin Claude Code). Run/humanizeron a long draft with a writing sample of yours pasted above it, and compare the critique it shows with the final version. If you already have an anti-slop skill, merge the rules into it rather than running two, as Plugins and Marketplaces recommends. - A/B the two video engines. Send the same 15-second brief to Remotion and to HyperFrames, one motion graphic and one map, and keep a note of which engine wins which job.
- Extend one skill. Run last30days on a topic you track, then ask the agent to send the parts the summary drops into a format you’ll reread, and to “package everything up as a new expanded skill.”
- Run a design pass. On any AI-built page, run Impeccable’s
critique, fix what it flags, then trybolderordelightonce to see how far each pushes the design.
Implementation
Tool/Service: Any agent that supports skills. In the video: ChatGPT Work, Claude Code, Cowork and Codex.
Setup: Paste the repo link with “install this skill from GitHub using its setup instructions.” Skill-specific: claude-video asks during install for an optional API key, used to transcribe videos that have no captions. Futurepedia used an OpenAI key. The wiki’s claude-video article also lists Groq as a backend.
Cost: All nine skills are free. The optional transcription key is the only paid add-on shown. SkillSpector can also use a model key, which is paid, for sharper scans.
Integration notes: Futurepedia’s free “12 skills for regular chats” resource and his ChatGPT sites video are promotion and were not ingested. Last30days output can feed a custom site. Futurepedia built his with Codex. Compare Codex Sites.
Open Questions
- Impeccable’s command count: Futurepedia says 24 in September 2026, and the wiki’s v3 article says 23. The repo’s current skill list wasn’t counted for this article.
- Futurepedia calls the skills compatible with “Grok bot, Hermes, OpenClaw and others.” Only the Claude, Codex and ChatGPT Work installs are shown on screen.
- Neither the video nor the README shows how humanizer interacts with an existing brand-voice skill when both are installed.
Related
- Agent Skills Overview — skill format and the security checklist to apply before installing
- Shopping for Skills and Plugins — how to choose among third-party skills
- Anti-AI Slop Guide — the writing-tells problem humanizer targets
- SkillSpector — the scanner, in depth
- Superpowers — the development-playbook skill pack
- gstack — Garry Tan’s 23-tool setup
- HyperFrames — the HTML video engine compared against Remotion
- Impeccable v3 — the design-command toolkit