Direct Answer
These five repositories are useful because each turns a fuzzy AI capability into an inspectable workflow: edit a draft, maintain relationships, cut a video, vet an agent skill, or operate a phone. The value is not the promise that open source automatically makes money. It is the ability to inspect the code, run one narrow test, and decide whether the workflow should become a product or remain private leverage.
The collection reviewed by Greg Isenberg spans very different risk levels. No AI Slop can start with a copied draft. Comp AI CRM touches identity, email, calendar, and customer data. Video Use invokes a local media toolchain and a transcription API. SkillSpector evaluates packages that may later receive agent permissions. Phone Harness can act inside real mobile apps.
Watch the Episode
Credit and scope: this article follows Greg Isenberg's solo episode, published on 2 September 2026, and checks the claims against each project's official repository. “Go viral” and “make money” are possibilities discussed by the creator, not guaranteed outcomes.
The Five-Repo Map
| Repository | Job | Small first test | Main boundary |
|---|---|---|---|
| No AI Slop | Edit repetitive AI-style prose | Review one draft without changing its argument | Authorship and brand voice |
| Comp AI CRM | Maintain a relationship workspace for agents | Import a small set of warm leads | Customer data and outbound actions |
| Video Use | Turn editing choices into a repeatable pipeline | Cut one short talking-head clip | Media dependencies and quality control |
| SkillSpector | Inspect agent skills before installation | Scan one candidate skill | A scan is not a safety guarantee |
| Phone Harness | Let an agent operate a real phone | Open a harmless app and read one value | Real-world account actions |
1. No AI Slop: Human Ideas, Cleaner Delivery
Peter Yang's No AI Slop skill looks for recurring patterns that make AI-assisted writing feel generic: false contrasts, unnecessary scene-setting, abstract claims, over-polished symmetry, and familiar filler language. Its stated goal is to remove those patterns without sanding away the writer's actual voice.
That distinction matters. Founders write landing pages, launch posts, emails, product updates, onboarding copy, and investor notes. A grammar tool can make every sentence correct while making the author less recognizable. This skill works best as an editor after the human has supplied the argument, evidence, examples, and point of view.
The official repository documents detection and editing modes and a global installation through the Skills CLI. Treat it as a reusable style review, not a detector-evasion tool or proof that a person wrote the text. Disclosure still matters where a platform, client, school, or employer requires it.
2. Comp AI CRM: Give the Agent a Relationship Workspace
Comp AI's open-source CRM reverses the usual CRM model. Instead of asking people to keep every field current, it gives an agent a workspace for researching contacts, enriching companies, preparing follow-ups, and maintaining relationship context.
The repository is unusually explicit about evidence. Agent proposals pass through an evidence ledger, facts should not be guessed, and weak evidence is held for human resolution. That is a better foundation than silently writing uncertain information into customer records.
The documented stack includes Bun, Next.js, NestJS, Prisma, PostgreSQL, Google or Microsoft authentication, email, and calendar access. Local development exposes the interface on port 3000 and the API on 3001. The software is open source, but the operational surface is not trivial.
| Start with | Keep human-owned | Measure |
|---|---|---|
| A small warm-lead list | Identity merges and deletions | Records corrected by humans |
| Read-only enrichment | Outbound messages | Useful facts with sources |
| Suggested follow-up dates | Calendar creation | Missed follow-ups recovered |
A sensible first pipeline is everyone who replied, asked for pricing, booked a call, or promised to return later. Let the agent gather evidence and propose the next step, then require approval before it writes back to the CRM, sends email, or changes a calendar.
3. Video Use: Turn Editing Taste Into a System
Video Use lets a coding agent transform raw footage into a rendered file. Its documented pipeline transcribes sources, packs them into a compact text representation, lets the model reason over an edit decision list, renders with FFmpeg, and checks cut boundaries before presenting the result.
The repository can remove filler words and dead space, apply color treatment, add audio fades, burn styled subtitles, and generate overlays through tools such as HyperFrames, Remotion, Manim, or PIL. Session decisions persist in a project file, so recurring work can become a documented editing system instead of living only in an editor's muscle memory.
This is not a zero-cost button. The official setup requires FFmpeg and documents an ElevenLabs API key for transcription. Visual self-evaluation helps, but it does not replace watching the complete render with sound.
4. SkillSpector: Scan the Tool Before You Trust It
NVIDIA SkillSpector scans agent skills for prompt injection, data exfiltration, suspicious dependencies, hidden instructions, and supply-chain risks before installation. That matters because a skill is not merely a prompt. It may include scripts, packages, network behavior, tool calls, and instructions that change what an agent is willing or able to do.
The scanner supports skills used with Claude Code, Codex, and MCP-based workflows. Static inspection can catch recognizable hazards, while model-assisted analysis can reason about combinations of instructions that are harder to express as a simple rule. The transcript also highlights a no-LLM option for teams that want a local static pass without sending skill content to a model.
A clean report is not proof that a package is harmless. Novel attacks, runtime downloads, compromised dependencies, excessive permissions, and subtle business-logic abuse can escape a pre-install scan.
5. Phone Harness: Agents That Operate a Real Phone
Phone Harness connects Claude Code, Codex, or another shell-capable agent to an iPhone through macOS iPhone Mirroring or to Android through ADB. The agent can see the screen, find text, tap, type, swipe, open apps, and read the resulting state.
On iPhone, the project captures the mirrored window, uses Apple's Vision framework for OCR, and converts coordinates into input events. On Android, ADB provides screenshots, accessibility-tree text, and input commands over USB or Wi-Fi. The repository documents limits: the user must connect the phone, iPhone unlocking pauses mirroring, OCR does not understand every icon, and camera, Face ID, multi-touch, and DRM video flows are not supported.
The opportunity is clear: mobile QA, repetitive creator tasks, cross-app workflows, and automation where no API exists. The risk is equally clear. A phone contains messages, authentication prompts, payment apps, photos, contacts, and security settings.
| Safe first steps | Require approval | Avoid initially |
|---|---|---|
| Open a test app | Send or publish | Primary personal account |
| Read a non-sensitive value | Purchase or subscribe | Banking and password apps |
| Run a scripted QA path | Delete or change settings | Unbounded background control |
Use a spare device or emulator, a dedicated account, a narrow app allowlist, and an action log. Let the agent navigate and collect evidence; keep irreversible or identity-bearing actions under human confirmation.
Install, Make One Workflow Work, Then Decide
The episode closes with a useful three-step method:
- Install the repository deliberately. Read the README and license, inspect the setup script, list external services, and identify every permission before running it.
- Make one small workflow succeed. Define the input, output, acceptance test, time limit, cost ceiling, and actions that require approval.
- Choose the destination. Productize the workflow only when the problem repeats, the buyer is clear, and operating costs leave margin. Otherwise, keep it as internal leverage.
This is a better monthly habit than cloning every trending project. The goal is not a larger tools folder. It is one additional capability that works reliably, has a known owner, and produces evidence you can review.
Which Repository Should You Try First?
| Your bottleneck | Start here | Success after one week |
|---|---|---|
| Generic AI-assisted writing | No AI Slop | Fewer edits without losing your voice |
| Leads disappear after conversations | Comp AI CRM | One clean, source-backed follow-up queue |
| Raw footage waits too long | Video Use | One approved short edit with a reusable project file |
| Your team installs third-party skills | SkillSpector | A documented scan and review gate |
| You need mobile QA or API-free automation | Phone Harness | One logged test-device workflow with approval controls |
Video Chapters
| Time | Topic | Time | Topic |
|---|---|---|---|
| 00:00 | Introduction | 15:15 | Repo 4: SkillSpector |
| 01:44 | Repo 1: No AI Slop | 18:49 | Repo 5: Phone Harness |
| 05:20 | Repo 2: agent-first CRM | 22:25 | Closing thoughts |
| 10:52 | Repo 3: Video Use |
Verdict
These repositories do not share a product category. They share a design pattern: turn work that normally depends on memory, taste, or manual navigation into a process an agent can read and execute.
No AI Slop is the lightest experiment. The CRM and Video Use have larger setup and data surfaces. SkillSpector belongs near the entrance to any shared skill library. Phone Harness is the most operationally powerful and deserves the strictest permissions.
Install less, learn more. Pick the repository closest to a recurring bottleneck, make one bounded workflow reliable, and keep the evidence. That is how an interesting GitHub project becomes either a defensible internal tool or the beginning of a real product.
Sources and Links
- Startup Ideas Podcast: 5 GitHub Repos: Kill AI Slop, Go Viral, Make Money
- Peter Yang: No AI Slop
- Comp AI: agent-first CRM
- Browser Use: Video Use
- NVIDIA: SkillSpector
- Shawn Pana: Phone Harness
This article uses the primary video's official YouTube publication date of 2 September 2026 and was researched and published on 6 September 2026. Repository capabilities, setup requirements, and dependencies can change, so check the current README before installation.