Benchmark your assistants on your own top tasks

How this helps you

You decide which assistant handles which job based on your own tasks, not someone else's leaderboard.

✦ Try it

I'm comparing assistants on my real work. From our history, list the top 10 tasks you do for me, most frequent first, with one line on what a great result looks like for each. Then list the connectors and skills you used for them, and anything you couldn't do because you lacked access.

Fill in: No placeholders. Send it to each assistant you're comparing, then build your task list from the overlap.

What they did

Claire Vo of How I AI ran a personal benchmark of four agents, including Grok Bot, on nine of her real tasks. She started by asking each agent for the top 10 tasks it does for her, then picked the tasks most common across them plus a few that test different skills: drafting a Slack reply in her voice, cutting a short video, fixing a bug, making a thumbnail, triaging her inbox for to-dos, a family newspaper, explaining a homework question, shoe shopping and a launch video. Before running it she checked that every agent had roughly the same connectors and skills. Prompts were one line, written the way she'd really ask, with a run sheet saying what to answer if an agent asked a question. She timed each run and scored the results herself. In her test Grok Bot came out on top overall, winning the thumbnail, to-do and shopping tasks, while it was weaker at cutting video and coding. Her takeaways: pick the agent for the job, and access matters, so give them the same tools before comparing.

Who it's for: People paying for more than one assistant, or running several Bots, who want to decide what goes where.

Before you start

Two or more assistants or Bots you want to compare.
About an hour to run and score the tasks.

How to try it

Ask each assistant: 'What are the top 10 tasks you do for me?' and collect the lists.
Pick 5 to 9 tasks that appear on more than one list, plus one or two that test something different (a design, a purchase search, a document).
Check each assistant has the same connectors and skills for those tasks, so you're comparing the assistant and not its access.
Write a one-line prompt per task, the way you'd normally ask, and a note on what you'll say if it asks a question. Run each prompt and note the time.
Score each result (did it finish, quality, time) and decide which assistant gets which kind of job.

What to expect

A small table: 7 tasks down the side, assistants across the top, each cell with done or not, minutes taken and a 1 to 5 score, plus a line on which assistant you'll use for what.

Common mistakes

Comparing assistants with different access. Claire forgot to connect one agent to her business tools and it couldn't do some tasks.
Writing long AI-style prompts. She used one-line prompts in her own voice to test real use.
Judging on one run. Treat a single test as a hint, and rerun the task you care most about.
Scoring with names showing. Viewers suggested a blind test, since knowing which agent made what adds bias.

Related tips

Three questions before you create another Bot: Only make a new Bot if the job needs its own memory, different permissions or its own schedule. Otherwise give it to your chief of staff.
Plan every Bot job with four blocks: Corey Ganim runs 6 Bots and 14 weekly routines across 11 apps. Each job is built from the same four blocks: Bot, routine, skill and connector.
One job per Bot: Give each Bot one job and a list of things it never does.