AI Video Editing Moves Into Production, While Agent Loops and Solo SaaS Show New Playbooks

Key Takeaways

1. A creator breaks down where AI video editing actually works today: video essays and tutorials, not most other formats.
2. A detailed AI-assisted video pipeline shows Codex, Final Cut, Remotion, and MCP tools handling script, cut, B-roll, and sound.
3. "The Gauntlet Loop" proposes blind evaluators and real benchmarks to stop agents from declaring victory too early.
4. A non-technical founder built a $29K MRR investor database, with half of revenue still coming from programmatic SEO.
5. AppSumo automated most of its Google Ads workflow, cutting agency workload and cost by roughly 90% while keeping human review.
6. A solo CRM added a Granola integration that syncs meeting notes into a context graph and daily briefings.
7. A local MLX port of a Jev-like system is fast but weak at Tetris, and a prior benchmark was found to be misleading.

1. AI Video Editing: Where It Works and Where It Doesn't

1. Alex Lieberman, who says he was once the biggest skeptic of AI video editing, relayed creator @cinkotweets' breakdown that AI currently fits mainly video essays and tutorials, because both have clear outlines, fixed scripts, and footage that can be recorded once and then edited by AI; other formats still depend heavily on craft. 1
2. The same breakdown describes an AI-assisted pipeline: AI helps find topic opportunities and mine material from tools and team discussions, but final judgment and sign-off stay with humans; Ahrefs can be connected to Claude and Notion to check weekly AI content opportunities and whether a video has entered the top five, or skills can scrape Slack, Notion, and meetings to surface real discussions and generate a personalized daily briefing, with Anthony building a content planning app on Codex and daily briefings costing about 8 cents on Opus. 1
3. For scripts and rough cuts, outlines are still built by humans in a three-column dialogue, visual, and sound format; a Scriptwriter skill uses interview-style questions so creators express ideas in their own words, then converts each beat into talking points for live reaction, with the reaction becoming the real script and visual and sound cues prompted alongside. Editing is staged: a radio cut for dialogue rhythm first, then VFX, archival material, and B-roll, and audio last; Codex with Final Cut produced a smooth cut without awkward pauses after one additional correction prompt, two prompts consumed about 86% of a 5-hour context window while roughly 53% of weekly Codex usage remained, and Codex beat Descript on the radio cut. 1
4. For post and measurement, an "archival finder" agent finds and downloads B-roll, archival material, and VFX in a given style with references, archives them, places them on the timeline, and generates paper-texture backgrounds; Remotion plus agents handles motion graphics, producing a plan first and then iterating, reaching the Final Cut timeline in about 9 minutes 10 seconds; Epidemic Sound is connected through MCP for track and sound-effect selection, iterating against creator references and using its trimming features to fit music to the edit with the mix sitting under the voice. 1

2. Agent Loops, Benchmarks, and Local Model Tests

1. Alex Lieberman introduced @mattshumer_'s "The Gauntlet Loop" for the problem of agents doing a task once, declaring completion, and scoring themselves too highly: first set a verifiable real benchmark such as "latest Call of Duty" quality, split the task among specialized sub-agents, loop block by block until the output looks like real work rather than "good for AI," then have blind evaluators compare output against references until they pick their own version; blind evaluators need fresh context so they do not inherit the builder's reasoning and rubber-stamp it. The method is also used for writing and websites, and Lieberman says the Something Big newsletter has used it for growth strategy and conversion copy with subscriber CAC results far above industry standards; on cost, demos and toys do not need the full loop, important work can cost hundreds of dollars or hit subscription limits, and he stopped Claude of Duty while it was still improving because of cost and because it was "already impressive," believing a few more days would get closer to real CoD quality. 1
2. Tony Dinh tested Laya-mlx playing Tetris on a MacBook: it was indeed fast at about 84ms, but clearly not smart enough, losing all three games to Jev. The quoted post describes Laya as an open-source, Jev-like text-output probability classification system ported to MLX with performance optimization, using at most 1G of memory, making 60 decisions per second playing Snake locally on an M3 Max, and claimed to be 50 times faster than Jev. Dinh said he looks forward to these models being fast and smart enough to perform a "tuck," and admitted his earlier test was flawed: the game framework did a lot of work for Jev/Haiku by preparing the best options based on game state for it to choose from; if the framework only provides game rules and state and lets Jev decide only the control button (left/right/rotate/drop), it basically becomes useless. 1 2 3

3. Solo SaaS, Programmatic SEO, and Automation

1. Rashid, a non-programmer from a finance background who once maxed out credit cards to raise $100,000 for a fintech company, manually built a database of 40,000 investors because finding investor lists was too painful, and launched it on Product Hunt, reaching $4,000 in revenue in the first month. That product, AngelMatch, is now at $29,000 MRR, 360 subscribers, a 33% trial conversion rate, and 800–1,000 clicks per day, priced from $59 per month up to thousands of dollars. He replicated the model: InvestorHunt makes $2,800 per month purely through SEO with zero marketing, and JournalistHunt is a database of 200,000 journalists priced at $49–$99. The turning point came when he saw 60 clicks per day in Google Analytics and decided to scale 10x: he hired 6 content writers, launched free tools, and did 8 months of programmatic SEO, growing MRR from $3,000 to $20,000 with a peak of $43,000; half of revenue still comes from programmatic SEO. His 2026 playbook: pick a database that solves a B2B pain point, collect or scrape the data manually, launch quickly to get traffic, then go all in on programmatic SEO; directions he would pursue today include an influencer database sorted by niche and audience size, or a newsletter sponsorship database. His conclusion is that "boring beats trendy," B2B pain points plus SEO compound for years, while AI apps go to zero. 1
2. Noah Kagan said AppSumo spends over $57,000 per month on Google Ads and the team has automated nearly the entire process; it previously outsourced to a Google agency costing over $10,000 and a separate Facebook agency, and the new setup is said to cut agency workload and cost by about 90%. The bot suggests which campaigns to cut, which to increase, which new campaigns to create, and how overall performance looks. Kagan said the bot still makes mistakes, so every result is reviewed by a human while adjustments continue; the system is built with Claude Code, n8n, Metabase API, Meta MCP, Google API, and GA4. He asked others how they manage ad spend. 1
3. Dharmesh said he spent the weekend building a Granola integration for YouSpotHub: after a user pastes a Granola personal API key, YouSpot syncs notes, extracts entities such as company mentions from the notes, links them in a context graph, and stores them in Second Brain. After the integration, that content can be accessed through chat, MCP, and cloud agents, and the daily briefing email becomes smarter by including meeting notes. He positions YouSpot as an AI-native Solo CRM for "companies of one," including independent side hustlers and creators building a personal brand, priced at $10 per month with a free tier. 1

4. Market Signals and Founder Economics

1. Acquire.com listed several new businesses: a SaaS content generation platform for publishing brand-aligned social media content with $11.8K TTM revenue asking $53K; a content B2B fashion authority brand with a lead list and registered trademark at $33K TTM revenue asking $39K; a subscription dog supplement brand in e-commerce with a stable customer base and recurring revenue at $740K TTM revenue asking $1.3M; and a SaaS options trading tool for serious traders with broker auto-sync and analytics at $17K TTM profit and $19.5K TTM revenue asking $68K. 1 2 3 4
2. Acquire.com also relayed the view that a founder earning $400K per year and owning 100% may be wealthier than a founder running a $50M-valuation VC-backed company; the former has growing cash, while the latter's 10% stake may be worth millions or may go to zero, and valuation and wealth are two different things. The account said another batch of founders completed acquisitions and promoted a free valuation service based on real acquisition data. 1 2
3. Nick Huber argued that the core problem for AI companies is that user switching costs are nearly zero, making return on investment hard to achieve, and predicted the market will be hit hard. He relayed a view that Meta Muse AI will take significant consumer ChatGPT market share; the quoted person said they use Chat personally and Claude for work, but after a day with Muse believed Chat will be replaced and said its distribution should be watched. 1
4. @levelsio argued that building a product outside AI and applying AI to that product is now more competitive than building another AI product, and that industries with the fewest technical people have the least competition. A reposted item said a class of SaaS whose moat is "developers don't have time to build it themselves" is being eroded. 1 2 3