Use planning when the choice of method matters, and a persistent goal when the work needs several stages. State what counts as done so OpenScience can evaluate progress against a concrete result.
Write a useful objective
Include the research question, inputs, deliverables, validation, and resource limits:
Produce a reproducible comparison of the two assays in data/. Deliver a
cleaning script, an effect-size table, one figure, and a methods note.
Preserve original data, account for repeated samples, and ask before any
paid compute. Finish by rerunning the script and listing limitations.
A broad instruction such as “analyze everything” leaves the output and stopping point ambiguous. Narrowing it makes the work easier to review.
Start in Plan mode
/plan Review the input files and propose the analysis before executing it.
Identify the unit of observation, missing information, candidate methods,
and the checks that would make the result trustworthy.
Plan mode can inspect material and prepare its plan. Review the assumptions, dependencies, costs, and intended file changes, then return to Research and authorize the agreed execution.
Set a goal
/goal Complete the agreed assay comparison with the four deliverables and
a successful rerun from the preserved inputs.
The plan panel and session header show current progress. Keep the task's acceptance criteria in a project file when the goal is long enough to span compaction or handoff.
Persistent goals still require the application and necessary resources to be running. They do not schedule future work while the machine is off.
Split work into reviewable stages
| Stage | A useful checkpoint |
|---|---|
| Inspect | Input inventory, data dictionary, missing prerequisites. |
| Plan | Chosen method, alternatives, assumptions, and budget. |
| Pilot | A small end-to-end run with inspected outputs. |
| Execute | Full run with recorded parameters and progress. |
| Validate | Reproduction command, diagnostic checks, unresolved issues. |
| Package | Final report, code, source references, and saved Results. |
Delegate independent questions when helpful, such as reviewing statistical assumptions while surveying relevant methods. Choose Off, Auto, or High delegation through Research controls.
Steer or stop work
Queue a follow-up when the current response can finish first. Stop it when the direction needs to change immediately. Inspect existing files and job status before retrying: a stopped response may already have written useful outputs or launched a compute job.
Stop expanding the experiment. Summarize completed work and current jobs,
save a checkpoint, and propose the smallest remaining validation step.
Use Compute jobs to manage an external job itself. Use Context and handoffs before moving the task to another conversation.