Workspace hygiene

Reuse before recreating. One canonical folder per artifact kind. Parametrize scripts. Never write artifacts into a skill bundle.

Slug: workspace-hygiene Baseline: intrinsic — injected into EVERY agent's base prompt by default (not assignable).

The rest of this page is the exact guidance this skill injects into an agent's system prompt.

---

Workspace hygiene

This is how you keep the SHARED workspace usable. It is a durable, common asset — not a scratch pad — and everything below applies to what you put THERE.

You have one when your ## Workspaces block lists a folder labelled shared — that block is what your tools can actually reach, so it is the only thing to check. When a ## Shared workspace block is also present, it lists what is already in there; its absence means the listing was not built for this turn, never that the folder is missing. If no shared folder is listed at all, work in the folder your ## Workspaces block names and ignore this skill's folder layout.

Reuse before recreating

Before building a workflow, script, or document, check the inventory: if a file already covers the need, load it and adapt it (file_read, then edit or pass different arguments). Recreating an existing artifact under a new name is a failure mode, not a fresh start.

One folder per kind — inside the shared workspace

These are the canonical folders of the shared workspace. When you save there, use them and never invent parallel ones (no outputs_v2/, my_workflows/, files dumped at the root):

• workflows/ — workflow definitions (JSON)
• outputs/ — generated artifacts (images, exports)
• scripts/ — reusable scripts
• documents/ — reports, notes, deliverables

This layout applies to the shared workspace ONLY. If your ## Workspace block names a folder of your own, that folder is where your work goes — do not invent a shared/ path inside it, and do not reach for the folders above.

One workflow = one graph

A saved workflow file is a reusable GRAPH (models, samplers, node wiring) — not a snapshot of one run. FORBIDDEN: saving a workflow file whose only difference from an existing one is the prompt, seed, or other run values — that is clutter, never value. Pass those as execution arguments instead (e.g. run_workflow.py --args {"prompt": …, "seed": …}) against the EXISTING file. The only reason to save a new workflow file is a change to the graph itself (different model, different node chain) — and it is named after the graph, not the scene.

Scripts must be reusable

A script you write takes its variable values (ids, paths, prompts) as ARGUMENTS — never hardcoded in the file. If you find yourself editing a script only to change a value, parametrize it instead. Before writing a helper, check whether an installed skill already ships one (skill_view lists its scripts).

Skill bundles are code, not storage

Never write generated artifacts into an installed skill's folder. Point a skill script's output argument at a real workspace instead — your own if you have one, otherwise the shared workspace, whose absolute path is in the NODAL_SHARED_WORKSPACE environment variable of every script/command you run. If a warning reports bundle writes, move those files out before finishing.

Leave it clean

Temporary diagnostic files (probe scripts, dumps, logs) are deleted before you finish the task — or not written to the workspace at all when stdout suffices.