Engine skill — Claude Code `/workflow`

[oesteban]

Objective

Deliver an MVP of Claude Code operating as a drop-in replacement for a workflow engine: a grounded skill for the Claude Code /workflow harness so the agent orchestrates a reconstruction from grounded knowledge (the KB + this skill) rather than parametric recall of the harness's syntax. This is the first engine-skill milestone and the one that establishes the reference workflow the others validate against. Victory: using the wiki — by this point a few operations are covered — the agent implements a simple anatomical workflow (brain extraction → intensity-inhomogeneity (INU) correction → brain-tissue segmentation → spatial normalization), with every step grounded in the KB and execution run locally, managed by Claude Code (the harness's agent() / parallel() / pipeline() / phase() primitives) — no external workflow engine involved.

Rationale

The transplant thesis holds only if a reconstruction is grounded: if the agent wrote harness code from parametric memory the result would be unverifiable, reproducing exactly the "did the model just memorize it" failure mode the project rejects. Claude Code is the lowest-friction execution target — it is the agent's native environment — so proving it can stand in for a workflow engine on a real (if simple) anatomical chain is the fastest route to a working end-to-end grounded reconstruction. It also fixes the reference workflow every other engine skill (bash, CWL, Nextflow, and the secondary targets) is subsequently measured against.

Methodological approach

Important

This is a separate skill family from niskiller (nipreps/skills-comm#6): the Claude Code Workflow harness is not a neuroimaging tool, and the boundary is structural — engine skills live in agentic-files, neuroimaging-tool skills in skills-comm. Grounding the engine as a skill is precisely what makes the reconstruction a product of knowledge, not parametric recall.

1. Author the skill. A lean decision surface plus a grounded reference of the harness's execution model — the agent() / parallel() / pipeline() / phase() primitives and their control-flow semantics — each construct tied to an authoritative source (official docs / reference) with provenance (version, retrieval date).
2. Ground the lowering. How a KB operation lowers to harness constructs, so emission is a grounded transformation, and assemble the four anatomical operations into a runnable local workflow.
3. Validate. Run the reconstructed workflow locally on the anatomical (T1w) images of a subset of the OpenNeuro testbench (Assemble the fMRIPrep testbench dataset drawing from OpenNeuro fmriprep-equivalence#3, in the making), and assess reliability with Bland–Altman analyses against the Nipype reference implementation of the same workflow (Nipype baseline workflow niworkflows#1053).

Outcomes/impact

1. A grounded Claude Code /workflow skill.
2. A working, KB-grounded, locally-executed simple anatomical workflow — the MVP proving Claude Code can replace a workflow engine.
3. The reference workflow and reliability baseline the other engine-skill milestones (Engine skill — bash #15–Engine skill — Snakemake #20) reuse.

Impact. The project's first end-to-end grounded reconstruction, and the anchor the whole engine-skill family builds on.

Direct upstream issues

(issues this one depends on / builds upon)

• Wiki KB v0.5: fMRIPrep coverage fmri-preprocessing-llm-wiki#12
• Nipype baseline workflow niworkflows#1053

Direct downstream issues

(issues that depend on this one)

• Workflow Architect Agent 1.0 #1
• Engine skill — bash #15
• Engine skill — CWL #16
• (optional) Engine skill — Pydra #18
• Engine skill — Dask #19
• Engine skill — Snakemake #20