machine-real-time (MRT)

Claude: "this project will probably take 3 months to complete"

Me: 🫸✋

A SKILL.md that stops AI models from giving human timelines for work they are going to execute at machine speed. Estimates come back in MRT — machine real time.

The problem

Model training data is decades of human project histories — sprints, standups, review cycles, coordination overhead. Ask a model "how long will this take?" and it pattern-matches to that history instead of to its own execution speed. You get "2–3 weeks" for a task the model finishes in the same session.

What this skill does

• Makes the model estimate as the executor, in MRT (minutes/hours of autonomous loops), not calendar-project time.
• Splits every estimate into agent execution time vs genuine external waits (DNS, CI, app review, human sign-off) — and bans padding beyond the named waits.
• Turns plans into execution queues: if every step is agent-executable, the model runs them instead of proposing phases.
• Parallel by default: independent tasks become concurrent sub-agents, not a sequential roadmap.
• Makes the model own its contribution — "I built and tested it," not "one would then implement." This is a team effort now.

Built for mixed human + agent teams

MRT matters most when agents stop being tools and become teammates. Buzz is exactly that: a self-hostable workspace by Block where humans and AI agents share the same channels — agents open repos, send patches, review code, run workflows, and huddle, each with their own keys and their own audit trail.

On a team like that, timeline language is load-bearing:

• One agent quoting "2–3 weeks" poisons the plan. Humans schedule around estimates. An agent teammate that pattern-matches to sprint culture makes the whole room plan human-slow for machine-fast work.
• MRT's split is the team's split. Agent execution (minutes/hours — run it now) vs external waits (human review, sign-off reactions, CI) is exactly the boundary a human + agent channel plans around. In Buzz, the human 👍 on a release is a named external wait — everything before it is machine real time.
• "Own the contribution" becomes literal. Rule 6 says speak as the teammate who did the work. In Buzz the agent is a member — its "I built it, patch in thread" comes with signed events and receipts.

Drop SKILL.md into your agents' instructions (Buzz's ACP harness runs Goose, Codex, and Claude Code) and estimates in your channels come back in MRT.

Install

Claude Code — as a skill:

mkdir -p ~/.claude/skills/machine-real-time
curl -fsSL https://raw.githubusercontent.com/DannyDesert/machine-real-time/main/SKILL.md \
  -o ~/.claude/skills/machine-real-time/SKILL.md

Or drop it into any project at .claude/skills/machine-real-time/SKILL.md.

Grok — same file, different folder:

mkdir -p ~/.grok/skills/machine-real-time
curl -fsSL https://raw.githubusercontent.com/DannyDesert/machine-real-time/main/SKILL.md \
  -o ~/.grok/skills/machine-real-time/SKILL.md

Anything else — paste SKILL.md into your system prompt / agent instructions. It's plain markdown.

Always-on (recommended): @-include it in your global instructions (e.g. ~/.claude/CLAUDE.md):

@~/.claude/skills/machine-real-time/SKILL.md

License

MIT