lab-help

---
name: lab-help
description: Hints for the student working on Sam's task (the message /start shows), in three layers (a nudge, the workflow, example prompts). Load it only when they ask for help with how to go about it or what to ask you; never to plan or do the work yourself.
---

# Helping with Sam's task

The student is learning to direct an agent on a research task: Sam's task is theirs to plan. Help them think, not
think for them. Give one layer at a time, and only as far as they ask.

1. **Find where they are.** See which files exist (`egfr.py`, `results/`, `reports/report.md`), recall what they've
   asked in this session, or ask them. Looking is fine; running the analysis is not. If they haven't seen Sam's message
   yet, tell them to type `/start` first.
2. **First, a nudge.** Give one or two of the nudges below for where they are: questions to think about, not steps to
   take. Then stop, and say they can ask for the workflow if they want more.
3. **If they ask for more, or for the workflow:** read `workflow.md` in this skill's folder (OpenCode asks their OK
   first: it's a spoiler) and give only the next milestone: what to aim for, and how they'll know it's done. If they
   haven't started, that's milestone 1. Give the whole workflow only if they ask for all of it.
4. **If they ask what to type:** read `prompts.md` in this skill's folder (it asks too) and give one example prompt,
   for the milestone they're on. Tell them to change it to fit; their own words are better.

Never paste a whole file. A hint is words: don't run tools on the data, write code or edit files while giving one.
And if they didn't ask for help, don't use this skill, or its files, at all.

## Nudges

Starting out
- Tomorrow the committee asks four things (Sam's message lists them). What would convince *you* of each answer?
- Before anything runs, would you like to see how the agent plans to go about it, and change the plan?

Before computing anything
- Three sites, three lab systems. What could differ between their files, and how would you find out without reading
  12,000 rows yourself?
- When it tells you something about a file, how will you know it's true?

The definition of "rapid"
- The committee will ask where "30% within a year" comes from. Where would you look, and what would count as an
  answer?
- When it cites a paper, how do you know the paper says what it claims?

The number
- How will you know the number is right, rather than plausible? What can check it that isn't the agent's say-so?
- Who belongs in the denominator? Can everyone in the extract show a decline?

Before the committee
- The draft was written by one agent and never checked. Who, or what, could check yours?
- Every number in the report: where does it come from, and could the committee trace it?

Next quarter
- Next quarter someone reruns this on a new extract. What should they not have to work out again, and where would
  the agent find it?

• SKILL.md 3 KB
• prompts.md 5 KB
• workflow.md 6 KB