← CV

Project archive

Helm

A unit for the human work of steering AI agents, weighted by how hard the steering was, not by hours.

My role
Creator and sole author, inside Meter
When
October 2026 — present, at Prosper Digital
Original
meter.pike.network
Archived

A copy kept here in case the original moves, goes offline, or I lose access to it.

An agent can run for hours on one prompt, so hours say little about the person behind it. Helm is the experimental unit I built into Meter to estimate that person's part: the time they spent steering, weighted by how demanding each piece of steering was.

The helm is where you steer from, and a nod to Helmholtz. The agent is the boat; the helm counts only your hands on the wheel. Meter draws the unit as a ship's wheel after the number: 12 helm.

The unit

One minute of attended time at moderate difficulty, 3 on a scale of 5, is 1 helm. Each scored stretch of work adds its attended minutes × difficulty ÷ 3, so six attended minutes at difficulty 4 make 8 helm.

  • Attended time is the person's measured share of an agent session: reading what came back, checking it and writing the next prompt. It is measured, never estimated, and is a floor, never added to tracked time.
  • Steering moments are the person's recorded interventions: instructions, decisions, reviews, corrections and verifications.
  • Difficulty comes from a model that groups linked prompts into stretches and rates each from 1 (routine) to 5 (very hard), with its pattern: clarification, reasoning, diagnosis or verification. Code does all the timing and arithmetic; low-confidence readings are marked uncertain.

What it shows

  • Total, Deliberating, Engaged and Peak gust tiles on every entry: the helm, the share of engaged time spent attending, attended time and agent turns together, and the most helm in any fifteen-minute window.
  • A Helm tab on each entry, charting helm per minute with each streak's pattern.
  • A ticket's lifetime as one timeline, coloured per person and stacked when two people steer at once, with a running burn of helm per day.
  • A nightly reference that places each ticket against past tickets of similar size: a description of past work, not an estimate for new work.

Privacy

The plugin sends each prompt's time and word count and when the agent stopped: never the prompt's text, a hash of it, or the agent's replies. The difficulty model sees no names, tickets or absolute times. Helm is an opt-in experiment, visible only to the people an admin selects, and absent from Meter's public API.

It is an unvalidated estimate, not a measure of anyone's mental effort or ability, and Meter says so wherever it shows the number.

Built with

Next.js inside Meter, Vercel AI Gateway for the difficulty model, and a nightly cron for the reference. 16 commits from October 2026.