Three numbers from last month: 6,906 installs. $0 in software cost. $0 in marketing spend. That is the first 30 days of kinocut, the video-editing pipeline we built for our own short-form work and then released as free, open-source software. This post is the business case - what it replaced, what it costs to run, and why giving it away was the cheaper decision.
The before state: two bad options
Any business that publishes short-form video on a schedule faces the same fork. Pay a human editor on retainer - US rates commonly run $500 to $2,500 per month for recurring clip work (market estimate, not a PuenteWorks quote) - or cut the clips in-house, where the cost stops being money and becomes the owner's evenings.
The obvious third option, letting an AI agent do the cutting, has a trust problem. Agents edit video the way they write code: confidently, and sometimes destructively. A wrong cut is recoverable. A silent overwrite of the master file is not. Every operator who has tried agent-driven editing has a version of that story, which is why the tooling never survived the pilot.
The intervention: guardrails, not intelligence
kinocut takes the position that the failure was never model intelligence - it was tool design. It exposes video editing to AI agents as a guardrailed pipeline: every operation validates its inputs before touching a file, nothing overwrites silently, real failure modes surface as errors instead of confident nonsense, and every output lands on disk with a receipt you can check.
It runs on local FFmpeg. No API keys, no accounts, no footage leaving your building. One long video becomes a week of platform-correct clips - cuts, captions, aspect ratios - with the agent planning and the pipeline rendering.
Software cost: zero. It is free and open source, and it stays that way.
The after state, quantified
Two measurements matter more than installs.
Small models become usable. We benchmarked it. A raw 3-billion-parameter model writing ffmpeg commands directly produced zero working commands out of twelve attempts. The same model working through kinocut's typed surface produced four out of twelve - at lower latency than a 35-billion-parameter model writing raw commands. Translation for a budget-holder: commodity local hardware plus a deterministic guardrail layer beats paying for frontier-model capacity on this task class.
Demand arrived before marketing did. Before we posted a single announcement, the tool passed 6,906 installs in 30 days on word of mouth alone, and an external team integrating it into production filed fifteen real issues in a single morning - the kind of usage signal you cannot buy. As of this week the repository carries 170 stars.
Inside our own shop, the change is simple: per-clip editing cost is now review time. The editor judgment did not go away - it moved up the stack, from executing cuts to approving them.
Recommendation
If your business publishes recurring short-form video, run a two-week pilot: one source video per week through the pipeline, one person reviewing output instead of producing it. Expected impact, based on our own workflow rather than a formal study: the production cost of a weekly clip cadence drops from retainer-plus-revision cycles to review time only, and nothing irreversible happens to your source files while you find out whether it holds. The pilot costs nothing but that time - the tool is free, local, and installs in one command.
The repo is at github.com/kyanitelabs/kinocut. Built by Kyanite Labs, our software side.
I spent six weeks building a business plan for a Panamanian coffee cafe in Long Beach. Not a template - a real plan with a financial model that generates itself, a Monte Carlo simulation, a six-seat AI committee that adversarially audited every number, and a bank package a CDFI lender would actually read.
Along the way I discovered something that shouldn't have surprised me but did: nobody has built AI tools for the hardest part of starting a business.
The gap
Anthropic shipped their "Claude for Small Business" pack in May 2026. Thirty-one workflows covering invoices, payroll, month-end close, tax organizers. They partnered with CDFIs - Accion Opportunity Fund, CRF, Pacific Community Ventures. Gave them credits and support.
And shipped zero tools for the borrowers those CDFIs serve.
Every one of their finance skills requires QuickBooks history. But if you're standing in front of a CDFI loan officer, you don't have QuickBooks history. You have an idea, a bank statement, and a spreadsheet you made at midnight. Their tools literally cannot help you until after you don't need them anymore.
OpenAI shipped a small-business program in July. Training academies with Shopify and Intuit. No curated skills, no lending tools, no planning capability. Google and Microsoft? Features and education, not ownable skills.
There's one restaurant vertical pack on GitHub (fourteen skills, well-built, food-only). There are 263 SaaS go-to-market agents. There are 334 marketing skills in the directory with exactly one local-SEO skill.
Nobody covers the journey from "I have an idea" to "I opened my doors."
What I built
Eleven skills. Each one is a SKILL.md file - a structured prompt that Claude Code, Cursor, ZCode, or any agent that reads skill files can execute. They cover the full lifecycle:
| Skill | What it does |
|---|---|
smb-router | Knows your stage and what's blocking you |
validate-opportunity | Stress-tests the idea before you spend money |
plan-business | Builds a 13-section lender-ready plan |
fund-business | Structures SBA/CDFI/family capital correctly |
launch-business | Sequences entity -> permits -> insurance -> opening |
control-money | Cash flow, invoices, close - without QuickBooks |
run-operations | Capacity, inventory, SOPs, compliance |
lead-people | Hiring, scheduling, founder-dependence scoring |
win-keep-customers | Positioning, local SEO, reviews, retention |
steer-growth | Weekly pulse, kill/scale/park decisions |
transition-business | Succession, sale, closure |
What makes them different
Three things, all learned the hard way:
1. Self-contained. Every money skill reads from a bank CSV export. Photos of receipts. Pasted numbers from a statement. No QuickBooks, no Stripe connection, no SaaS dependency. Day one, no setup. When you eventually connect an accounting tool, it's an accelerator, not a prerequisite.
2. Deterministic money. The AI never does arithmetic. All calculations run in scripts - Python or R - and the model explains the results. I learned this after a financial model silently computed a working-capital table at one ticket price while stating another in the prose. The sign of the year-one cash position flipped. Nobody caught it for three audits. Now: assertions that fail the build if prose and computed values diverge.
3. Evidence-gated decisions. The funding skill won't recommend debt before counted revenue. The growth skill won't recommend expansion without a kill/scale/park receipt. Every capital decision is tied to a gate with a pre-registered metric and a kill date. This came from a Shark-Tank-style investor who told me: "The 91% is not a probability of success. The founder has not entered the arena." He was right.
The comparison that matters
| Anthropic's pack | This pack | |
|---|---|---|
| Works pre-revenue | No (needs QuickBooks) | Yes (bank CSV) |
| Business plan builder | No | Yes |
| SBA/CDFI borrower tools | No | Yes |
| Launch sequencer | No | Yes |
| Inventory management | No | Yes |
| Succession planning | No | Yes |
| Cost | Free (Claude subscription) | Free (MIT, any agent) |
I'm not competing with Anthropic's pack. I'm completing it. Their back-office workflows are excellent - once you have books. Mine get you to the point where you have books.
Why open source
Because the people who need these skills the most - the cafe owner, the food truck operator, the woman opening a salon - can't afford a consultant. They're the ones standing in front of a CDFI loan officer with an idea and no books. They need tools that work from a shoebox of receipts.
MIT license. Fork it. Build a business on top of it. Just keep the license.
Get it
github.com/simongonzalezdc/small-business-operator-pack
Clone it, copy the skills into your agent's skills directory, and start with smb-router. It'll ask you where you are in the journey and route you to the right skill.
If you're a CDFI, SBDC, or small-business advisor and want to see how these could serve your borrowers, get in touch. That's the partnership Anthropic talked about but didn't build.