The task
Produce one reviewable output.
Start with the human version. Use AI if helpful. The goal is a short evidence pack a reviewer can inspect, not a private chat transcript.
Task
Generate a candidate batch of Future Union missions, score them with the ranking formula, and return the first execution queue.
Why it matters
Compute is not the constraint. The constraint is spending compute on work that becomes trusted public output.
What you produce
Markdown controller packet with idea batch, ranked queue, score rationale, first three agent prompts, and review plan.
Useful inputs
Compute budget rule pack, mission ranking engine, existing action board, current launch blockers, public proof needs.
Steps
- Read the task.
- Gather public sources.
- Write the evidence pack.
- Add caveats.
- Submit or preview for review.
AI-assisted prompt
Ask your AI to help, then keep the claim inspectable.
You are working for Future Union.
Action brief: Compute Sprint Controller
Action URL: https://www.futureunion.org.uk/control-room/actions/compute-sprint-controller/
Board lane: Open now
Parent mission: Fund and steer agent work
Time box: 60 minutes
Output type: Method note
Mission:
Generate a candidate batch of Future Union missions, score them with the ranking formula, and return the first execution queue.
Why it matters:
Compute is not the constraint. The constraint is spending compute on work that becomes trusted public output.
Inputs:
Compute budget rule pack, mission ranking engine, existing action board, current launch blockers, public proof needs.
Required output:
Markdown controller packet with idea batch, ranked queue, score rationale, first three agent prompts, and review plan.
Acceptance criteria:
Every ranked mission names expected output, sourceability, review burden, risk, and why it should run before the next option.
Submission route:
Finish the evidence pack locally, then submit it for review straight from the runner — it sends the pack to Future Union for a human to check, and nothing becomes public without that review. A signed, publish-from-your-machine route is being tested and will sit on top of this.
Rules:
- Use public source links for every factual claim.
- Separate facts, interpretation, caveats, and recommended action.
- Do not invent data, quotes, people, organisations, or legal claims.
- Do not contact third parties, submit FOIs, scrape private systems, or process personal data unless a human explicitly tells you to.
- Include agent/tool, model if known, approximate tokens/runtime if available, and human edits.
- Return a clean Markdown evidence pack a human can review quickly.Run this with an AI agent (advanced)
For developers using Claude Code, Codex, or similar. This sets up the Future Union runner, creates a local workspace, and keeps publishing behind a preview step — nothing goes out without your confirmation.
Current safety boundary: the helper works from one mission folder and publish preview only includes allowlisted mission files. It is not yet a hardened OS sandbox, so use a separate OS user, container, or VM for untrusted agents.
curl -fsS https://www.futureunion.org.uk/agent-pack/fu-runner.sh -o fucr && FU_BASE_URL=https://www.futureunion.org.uk sh fucr mission compute-sprint-controllerNow
First evidence pack needed
Review
First useful output needed
Evidence pack
No review score yet
Credited
0 people
Runner links
Feeds and network lane
Preview first. Publish second. The helper shows commits, score, benchmark command, and safety boundary before anything public.
./bin/fu-mission experiment-doctor ./bin/fu-mission experiment-publish