You are using Claude Fable 5 wrong

An Anthropic employee used Claude 5 to edit and launch a professional video that would previously have cost hundreds of thousands of dollars. The pipeline ran 17 takes across 4 scenes, with Claude 5 sub-agents selecting the best takes and encoding edits in a JSON file executed by FFmpeg. Supporting tools included 11 Labs for transcription, Whisper for filler-word removal, FFmpeg for stitching and color grading, and Remotion for final assembly. Claude 5 performed color grading from a plain-language instruction about muted colors, with no manual configuration.

The most counterintuitive finding from early Claude 5 use is that low effort is the alpha. The claim is that Claude 5 on a low effort setting outperforms Claude Opus on a high effort setting, meaning high effort mode should not be the default. Starting around June 22nd, Claude 5 API pricing will not be included in Max accounts, making token management a real cost consideration. Factory.ai's Droid product can orchestrate when to use Claude 5 versus cheaper models based on task complexity to control those costs.

A copywriting tournament framework has Claude 5 write a landing page eight different ways, then five judges score each version, producing 40 scores and one winner. The judges include a skeptical CFO, a distracted founder, a competitor, an ideal customer, and a conversion copywriter. The CFO judge flagged that a claim of two thousand dollars a month must be immediately defended or it reads like a Zillow estimate. The midnight founder judge noted that one hundred thousand dollars on a guess is a gut punch that stops the scroll. Claude 5 is described as unusually effective at simulating how real expert personas evaluate content.

The recommended startup-building workflow is to have Claude 5 interview the user in the style of Mark Zuckerberg, Sam Altman, or Brian Chesky, asking up to 15 questions, then write a full spec and list three ways it fails before building V1. Claude 5 pushed back on vague answers in a way prior models had not, telling the speaker that life got in the way is a horoscope answer and demanding a specific scene. That interview process changed the product significantly, producing a phone number that can see the user's calendar and text with receipts and a consequence, rather than a widget or habit platform. Claude then built the app and spec document including architecture recommendations of Python, FastAPI, and APScheduler along with pricing and V1 scope.

A use case called hire it to kill your company involves feeding Claude 5 a P&L, pricing page, churn data, and support tickets and asking it to build the competitor that destroys the business. The key differentiator in the resulting threat ranking is which attacks could execute themselves with no employees. Claude 5 is described as particularly strong with large data sets including churn data and support tickets. The speaker identifies an opportunity to buy SaaS products cheaply, use Claude 5 to refactor code and analyze support tickets, then flip or hold for cash flow.

Claude 5 has a one million token context window, which the speaker says can hold years of notes and decision documents simultaneously. Feeding two years of notes, decision docs, and postmortems to map decision patterns and produce a one-page operating manual is described as a use case very few people are currently doing. Claude 5 is also noted to explore before it answers, making it effective at identifying what is absent rather than what is wrong, and it can read tables, footnotes, and charts inside documents.

Three service business models are outlined using Claude 5. A synthetic focus group startup charges approximately three thousand dollars per launch at a token cost of roughly one hundred dollars, targeting DTC brands that spend fifty thousand dollars on media before learning an ad does not work. A custom software firm sells internal tools for five thousand dollars flat built in two days versus a traditional cost of eighty thousand dollars per year, with token costs estimated at a couple of hundred dollars maximum. A contract refund firm takes 25 percent of savings found with no upfront charge, targeting auto renewals in the next 90 days, mismatched invoices, unused seats, and price escalators ratcheting 8 percent per year. The speaker notes most mid-sized companies hold hundreds of vendor contracts not reviewed since signing. The recommended contract review prompt asks Claude 5 to identify non-obvious costs, obligations arising in 18 months, missing protections, three changes to request, and items requiring a real lawyer, with the explicit note that AI should not replace lawyers but should be used alongside them.