Automating a business with Claude: audit your work, eval each task ten times, turn the winners into skills, then schedule them

Step 1: audit your work by showing, not telling

• He had Claude build a tracker that screenshots his screen about every five seconds, compresses each image to the smallest legible size, describes what he was working on and stores it in a database. He ran it for about a week.
• On top of that, Claude asked him 10 to 15 short questions a day: what are you working on, why, what systems does this need.
• The reason: you often can't list everything you do or the context behind it, so it is more reliable to let the agent watch.
• The output is a full list of deliverables and processes (for him: lead lists, email enrichment, cold email copy, domain warm-up, replies, booking, proposals and more).

He says the analysis used a full session of usage on the top model with maximum thinking and that he had to upgrade his plan. A screenshot logger captures everything on screen, including client data and passwords, so decide where that database lives and delete it when you are done.

Step 2: plan the automations with Claude

Give the list to a fresh Claude session with a goal along these lines: "Here's everything I do on a daily basis. My goal is automating the vast majority of my non-customer-facing work. Help me build systems for each. If you don't have enough context, ask me." Then answer its questions by voice dictation; he says about an hour of talking gave it enough context.

Step 3: eval every task until it scores about 8/10

1. Run each task 10 times with the same prompt.
2. Score the 10 outputs yourself. Single runs mislead: the same job can come out well once and badly the next time.
3. Below roughly 8 out of 10, tell Claude what is wrong ("a lot of outputs look like this, I want them to look like that") and have it improve the skill or system.
4. Repeat until it passes. Supply templates and examples where needed, such as high-performing cold emails, qualified lead lists and good proposals.

Time saver: batch three or four changes and run their tests in parallel instead of fixing one thing and re-running the whole suite. He says that turned about a week of iteration into a day. This was the most hands-on part of the whole process.

Step 4: turn passing workflows into skills, then decide what to schedule

• Anything that scores well enough becomes a skill. In Claude Code a project skill is a folder at .claude/skills/<name>/SKILL.md with optional supporting files such as templates (Claude Code skills docs).
• He says that on a Claude Teams plan, skills can be shared with non-technical colleagues who use Cowork or Claude chat rather than Claude Code. Check your plan's skill-sharing options before relying on that.
• For each skill, ask Claude whether it makes sense to run it on a schedule. A daily client update draft is a good fit, though he still reads it before sending. Building a deliverable or writing a full campaign keeps a human in the loop and is run on demand.
• His result: five or six tasks a day left for him, out of 80 to 90.

Gotchas & Caveats

• The revenue figure and the share of work automated are his own claims.
• He recommends the most capable model on maximum thinking for the planning steps, since everything later depends on them, even if that means waiting for your usage limit to reset.
• The 8/10 threshold is his judgement. Set yours by how costly a bad output is.

Key Takeaways

• Let the agent observe your work for a week before asking it to automate anything.
• Ten runs per task, scored by you, iterated to about 8/10, with tests batched in parallel.
• Only passing workflows become skills, and only low-risk skills get scheduled.