Most “AI productivity” content is a list of prompts with no context on how they fit an actual workday. These are five workflows we run ourselves at O.N.E Agency, with the specific setup for each — not hypotheticals.

1. Meeting notes that turn into tasks automatically

Tools: a meeting recorder piped into Claude via n8n.

The recording gets transcribed, then a prompt extracts action items with owners and deadlines, formatted for direct paste into our task tool. The time saved isn’t in writing notes — it’s the 15 minutes after every call where someone would otherwise reconstruct “wait, who owns the client follow-up?”

2. First-draft client proposals from a call transcript

Tools: call transcript + Claude with our proposal template as a system prompt.

We feed the sales call transcript and our standard proposal structure into Claude, and it drafts a tailored first version referencing what the prospect actually said, not generic boilerplate. A human still edits it — but editing a relevant draft takes 20 minutes instead of writing one from scratch takes two hours.

3. Support and client ticket triage

Tools: our helpdesk’s API + a script that classifies and drafts.

Incoming requests get auto-tagged by category and urgency, and a draft reply is generated for review, not auto-send. This is the difference between “AI replacing the team” (which frustrates clients) and “AI removing the blank-page problem” (which actually works).

4. Weekly content repurposing

Tools: Claude + our existing long-form content.

One long piece — a blog post, a client case study — gets turned into a week’s worth of social posts and a newsletter blurb, each adapted for the platform rather than copy-pasted. This only works well if you feed the model real examples of your brand voice first; generic prompts produce generic output.

5. Code review pre-pass before a human reviewer

Tools: Claude Code.

Before a pull request goes to a human reviewer, an AI pass catches the obvious stuff — unused imports, missing error handling, inconsistent naming — so the reviewer’s time goes to architecture and logic decisions instead of style nits. We’ve seen review turnaround drop by roughly a third since adopting this.

What these have in common

None of these replace judgment on the final output. Each removes a specific, recurring chunk of manual work — transcribing, first-drafting, triaging, reformatting, nit-picking — and leaves the decision-making to a person. That’s the pattern worth copying: automate the mechanical part, not the part that requires judgment.