Every developer has them. The deploy sequence you could do blindfolded. The PR review → merge → tag → release dance you run three times a week. The data-export pipeline that's five steps across four tools.
You know these should be automated. But writing a custom script for each one is its own project — and half the workflow is judgment calls a bash script can't make.
The old way vs. the new way
Traditional automation (scripts, Zapier, CI/CD rules) requires you to manually define every trigger, step, and branch. That works for simple pipelines, but it breaks down when workflows involve tool-switching, visual checks, or context-dependent decisions.
AI-based automation flips the model. Instead of defining the automation upfront, you record yourself doing the work. The AI watches, extracts the pattern, and proposes the automation for you.
The difference matters: you don't need to articulate your workflow — you just do it.
Five steps to get started
1. Identify the candidates. List tasks you repeat 3+ times per week that follow roughly the same shape each time. Deploy sequences, bug triage flows, data transforms, review checklists.
2. Record an execution. Use a tool that captures system events alongside video — file changes, terminal commands, app switches, not just pixels. The AI needs structured data.
3. Let the AI extract the pattern. A good tool produces: the step sequence, decision points, variables (what changes between runs), and constants (what stays the same).
4. Review and refine. First-pass accuracy is typically 80–90%. Confirm the sequence, flag true variables vs. incidental variation, add guardrails for edge cases.
5. Run in supervised mode, then let go. Each execution teaches the AI about edge cases. Accuracy compounds.
Three mistakes to avoid
Over-automating too early. Start with workflows you've done 10+ times. AI needs repetitions to find patterns.
Ignoring context switches. A workflow spanning IDE → browser → terminal → Slack is harder than one staying in a single tool. Start contained.
Expecting perfection from one recording. Record 3–5 executions of the same workflow. The AI needs variance to separate signal from noise.
What we're building toward
Distill captures your screen and system events in real time, then extracts structured, reusable skills from each session. After multiple recordings, it identifies patterns you might not notice yourself and proposes rules for your approval.
The compounding part is the point: your AI gets measurably better every week because it learns how you specifically work, not generic best practices.
The gap between "I could automate this" and "I have automated this" shouldn't require a weekend project. It should require pressing record.