9 AI Agent Skills Worth Trying for Research, Video, and Code

A useful agent skill is a repeatable workflow, not a magic prompt. In this Futurepedia walkthrough, nine projects cover recent-topic research, video creation, writing, video analysis, software delivery, and interface design. The interesting part is seeing the same idea run across ChatGPT Work, Claude Code, Claude Cowork, and Codex. The important qualification is that each host has different access to files, commands, websites, and installed tools.

Watch the Nine-Skill Walkthrough

Credit: Futurepedia's original video, published 28 September 2026, and the supplied transcript. The demonstrations and preferences are the creator's tests; the repository and compatibility notes below were checked against the maintainers' current pages.

Choose a Starting Point

• Find what people are saying now: Last30Days.
• Make coded motion: Remotion for a React production workflow; HyperFrames for HTML-based video and agent-led composition.
• Review a skill before adding it: SkillSpector, followed by your own inspection.
• Improve existing writing or understand a video: Humanizer or Watch.
• Build and review software: Superpowers, gstack, or Impeccable, depending on whether your gap is process, team roles, or visual quality.

1. Last30Days: Research a Live Conversation

Last30Days collects and scores recent material rather than asking a model to summarize its older training data. In the video, Futurepedia uses it to research Jev, then customizes the output into a site that preserves source posts, use cases, and unconventional takes. That second step is the best lesson: use a public skill as a starting point, then define the report shape you actually need.

First test: research one product you already know, inspect the cited posts, and note what the summary missed. The maintainer says Reddit, Hacker News, Polymarket, and GitHub work without extra setup; sources such as X and YouTube need their own setup. Do not assume that a research brief verifies every social claim it repeats.

2. Remotion: Video as React Code

Remotion lets an agent build programmatic video with React. Futurepedia asks for an agent-skills motion graphic, then a national-parks route animation. A supplied logo and a specific follow-up improve the first result. The framework and its official agent skills are separate repositories, so use the skill documentation when setting up an agent.

First test: make a 10-second graphic with supplied brand assets, inspect frame timing and legibility, then request one focused correction. Check Remotion's license before client or company use; an accessible source repository does not mean every commercial use is free.

3. HyperFrames: HTML-Based Video Production

HyperFrames turns HTML, CSS, media, and seekable animation into rendered video. Futurepedia runs a similar prompt through both video tools and prefers HyperFrames' camera movement and composition on the map example. That is a creator preference from two demos, not a general benchmark. Its official skills guide recommends the core entry skill and adds specialized workflows only as needed.

First test: use the same 10-second brief in both tools. Compare accepted output, revision time, render cost, and whether you can maintain the source. A polished preview still needs a real render and a human viewing pass.

4. SkillSpector: Check the Trust Boundary

NVIDIA SkillSpector scans skill files, scripts, repositories, and archives for suspicious behavior such as prompt injection, exfiltration patterns, and excessive permissions. Futurepedia runs it before installing Humanizer. It is a security tool, not an ordinary writing or coding skill, and its report is one input to the decision.

First test: scan a downloaded skill before installing it, read high-severity findings, then inspect the files and required network calls yourself. NVIDIA explicitly says the scanner does not sandbox a skill. A clean report cannot make unknown code safe to run with production credentials.

5. Humanizer: Edit Recognizable AI Prose

Humanizer identifies repetitive AI-writing patterns. In the video, it revises three short Jev posts and explains what it changed. The result is a useful editing pass, but it cannot supply a writer's lived examples, correct unsupported facts, or replace an editorial point of view.

First test: give it a draft you wrote or approved, ask it to mark the patterns before revising, and compare the new version against your voice. Keep the facts and citations under human review.

6. Watch: Turn Video into Usable Evidence

Watch extracts timed frames and transcript evidence from a video URL or local file. Futurepedia uses an Instagram cooking clip because the essential timings are shown on-screen rather than in the caption. It is a good example of why a transcript alone may not answer a video question.

First test: ask for a step-by-step summary with timestamps, then spot-check frames and speech against the source. The maintainer currently documents Claude Code and Codex support but says this Watch project does not run in Claude Chat or Cowork. Python, FFmpeg, and yt-dlp may be needed, and optional transcription routes can carry setup or API costs. Respect the source site's access and media rights.

7. Superpowers: A Coding Process in Skills

Superpowers packages brainstorming, planning, implementation, testing, and debugging practices into skills that can activate during coding work. Futurepedia's point is that the agent can use the playbook without you remembering a slash command for every step.

First test: take a small existing issue and check whether the agent writes a plan, proves the failure, makes a bounded change, and runs the relevant tests. Inspect the plan and diff; the skill's process should improve your team's review, not replace it. Installation varies by agent host, as its README notes.

8. gstack: Opinionated Roles for Shipping Software

gstack offers a larger set of product, design, engineering review, QA, and release workflows. Futurepedia pairs it with Superpowers, but they overlap. The maintainer's quick start uses roles such as /office-hours, /review, and /qa; its setup has host-specific dependencies rather than a universal paste-a-link install.

First test: run a plan review on one feature, then QA one staging build. Keep the recommendations you can verify and reject process that adds ceremony without reducing defects. Review browser, deployment, and credential permissions before using its broader automation paths.

9. Impeccable: Give Design Feedback a Vocabulary

Impeccable gives coding agents design commands for hierarchy, layout, motion, accessibility, and polish. Futurepedia applies stronger styling, a passport-stamp interaction, and critique to a parks site. The concrete takeaway is to make the agent inspect the interface as a user would, not merely produce attractive code.

First test: ask for a critique of one real screen, fix the top two issues, and verify both desktop and mobile. The current repository lists commands such as critique, bolder, and animate; command names and installation flow can change, so follow current docs rather than copying a spoken command from the video.

Install One Skill Safely

For a local coding-agent environment, these are three separate maintainer-documented starting commands. Pick the one matching your job; do not run the block as a bundle. The first example installs globally, while the other two default to the current project:

npx skills add mvanhorn/last30days-skill -g
npx skills add remotion-dev/skills
npx skills add heygen-com/hyperframes

Other projects use different routes: Superpowers documents host-specific installation, Watch separates Claude Code from Codex setup, and Impeccable publishes its own installer. Check the current instructions before running any command from a video or article.

1. Choose a real task: select one skill whose output you can check against your current workflow.
2. Read the source: open the maintainer's README, SKILL.md, scripts, license, dependencies, and host-specific instructions. A GitHub URL in a prompt is a request for your agent to inspect and run code, not a trust signal.
3. Limit access: use a disposable or low-risk project first. Keep credentials out of the skill's files and approve network calls, installs, and edits deliberately.
4. Verify installation: start a fresh agent session if required, confirm the skill is discoverable, and run its smallest supported example.
5. Keep the evidence: compare the output with the source material, tests, rendered video, or UI screenshot. Customize only after the base workflow works.

Not every project needs all nine. A small collection of trusted, well-tested skills is more useful than a crowded agent environment whose instructions conflict or whose tools you never inspect.

Video Chapters

• 00:00 - Introduction to agent skills
• 00:27 - What a skill is
• 01:01 - Last30Days research
• 02:55 - Customize a skill
• 04:58 - Remotion and HyperFrames
• 08:16 - SkillSpector
• 09:26 - Humanizer
• 10:51 - Watch videos
• 12:36 - Superpowers
• 13:12 - gstack
• 13:44 - Impeccable

Official Skills and Project Links

• Futurepedia: 9 Free AI Agent Skills You NEED to Install Now (primary video and supplied transcript)
• Last30Days - recent-topic research
• Remotion framework and Remotion agent skills - React video production
• HyperFrames and skills guide - HTML-based video production
• SkillSpector - pre-installation security scanning
• Humanizer - AI-writing pattern review
• Watch - video frames and transcripts
• Superpowers - engineering process skills
• gstack - product-to-release skill collection
• Impeccable and source repository - design critique and UI polish