Skip to content
neuroflow0.2.31 · docs

← neuroflow

/neuroflow:pipeline

Define and run a multi-step research pipeline across any sequence of neuroflow phases.

/pipeline lets you plan and execute a sequence of neuroflow commands in order, one step per invocation — either interactively (asks for your approval before each step) or in brutal mode (--executor, no questions). The plan is saved, so each new /pipeline picks up the next step.

When to use it

  • You want to run several phases in a planned order — one /pipeline per step, without remembering which command comes next
  • You want a plan-first approach: see the full pipeline before executing any step
  • You need to resume a pipeline that was interrupted mid-run
  • You want automated end-to-end execution with error logging

Modes

Claude asks for confirmation before each step, runs it, and ends by telling you the next step — run /neuroflow:pipeline again to do it. You can skip individual steps, adjust the plan, or stop at any point. Claude also asks you where you want to stop upfront, so you can run the pipeline in stages.

/neuroflow:pipeline

You confirm the plan once; after that, every /neuroflow:pipeline --executor runs the next pending step without asking. Errors are logged and the next invocation moves on — unless the next step depends on the failed one, or two steps in a row failed. The stop-point question is skipped — all pending steps run. To leave it running unattended, wrap it in Claude Code's /loop.

/neuroflow:pipeline --executor
/loop /neuroflow:pipeline --executor

How it works

  1. Read project state — Claude checks flow.md and project_config.md to understand which phases have been worked on
  2. Draft a pipeline plan — a pipeline-plan.md is written to .neuroflow/pipeline/ listing all steps with status ([pending], [done], [skipped], [deferred])
  3. Choose your stop point — Claude asks how far you want to go this session (a specific phase, or all the way through). Steps beyond the stop point are saved as [deferred] and picked up next time.
  4. Confirm before running — in interactive mode, Claude shows the plan and waits for your go-ahead
  5. Run one step — the next pending step, following the same lifecycle as running the command directly; then Claude updates the plan, names the next step, and stops (a /compact between long steps keeps the context lean)
  6. Continue — each new /neuroflow:pipeline reads pipeline-plan.md and runs the next step, including any previously deferred steps once you extend the stop point. When the last step is done, Claude suggests its natural follow-up (after /paper: /paper --submit)

Stop-point selection

After confirming the plan, Claude presents your pending steps as numbered options:

How far would you like to run the pipeline?

  1. Stop after ideation
  2. Stop after experiment
  3. Stop after data-analyze
  4. All the way through → paper

Pick a number (or type a custom stop point). Steps beyond your chosen point are marked [deferred] in the plan — run /neuroflow:pipeline again to continue from where you left off.

If you choose all the way through from ideation to paper, expect a surprise. 🎉

Files read and written

Direction Files
Reads .neuroflow/project_config.md, .neuroflow/flow.md, .neuroflow/pipeline/pipeline-plan.md
Writes .neuroflow/pipeline/pipeline-plan.md, .neuroflow/pipeline/flow.md, .neuroflow/sessions/YYYY-MM-DD.md, .neuroflow/reasoning/pipeline.jsonl, phase subfolders (via the commands it invokes)
  • /phase — check or switch the active phase before planning a pipeline
  • /neuroflow — get a full project status overview
  • /sentinel — audit .neuroflow/ for consistency after a full pipeline run