agentleFS
millhqmillhq/mill/skills/mill-use/SKILL.md

mill-use

Skill1 starsChanged 2 months agoApache-2.0View on GitHub

Drive Mill (guardrailed workflow automation + Atlas knowledge board) over MCP — compose, run, and inspect workflows; read and write Atlas; respect the approval gate. Use when connected to a Mill MCP server or asked to automate through Mill.

The file

We split this skill into the rules it gives an agent

3 rules. About 0 are common, 2 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Never attempt to bypass a gate.

    line 36
  2. Very specific:

    Run with test: true while iterating (excluded from the human's production metrics).

    line 50
  3. Distinctive:

    Ask the registry, don't assume: step IDs and contracts are the source of truth, and they evolve.

    line 96

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Using Mill as an agent

Mill is a local desktop app exposing its full surface over MCP. You get what the human gets — same tools, same guardrails. Full product docs in one file: userdocs/llms-full.txt in the Mill repository.

Connect

The human finds their address under Settings → MCP access (default http://127.0.0.1:8091/mcp). Writes may be deny-all until they enable agent access there.

The mental model

  • A workflow = one trigger + connected steps. Every step has a typed contract: what payload kind it consumes and produces (text/html/markdown/json/any/none, or passthrough). Read the registry resource before composing — the server validates edges against these contracts and rejects definite mismatches with the reason.
  • Payload = the single artifact flowing through a run. Attributes = the workflow's declared typed fields; steps read and write them by name. Structured data goes in attributes.
  • Configure entities (integrations, lists, MCP servers, AI providers, execution environments) are referenced by ID from steps. Create or look them up first; never inline a secret — secrets live in the OS keychain via the entity's credential field.
  • Effect classes: steps marked external PARK the run for human approval unless a guardrail rule allows them. Expect runs to sit in parked state; that is working-as-designed, not an error. Never attempt to bypass a gate; surface the pending approval to the human instead.

The working loop

  1. Tool arguments use the EXPORTED-WORKFLOW JSON envelope: workflow tools take a json string in the same shape export_workflow returns (steps/edges arrays with nodeTypeId), not an inline object — export a seeded workflow once and mirror its shape.
  2. List step types / read the registry resource; check each step's Takes/Produces before wiring.
  3. Author or update the workflow (IDs are stable; labels are yours).
  4. Validate — fix any issue the server reports (they name the step and edge).
  5. Run with test: true while iterating (excluded from the human's production metrics); use the stepped-run tool to debug one step at a time. Production runs omit the flag.
  6. Inspect the run's step records — each carries input, output, and failure text prefixed by the failing step's ID.

Testing an integration

test_request executes ONE real HTTP call -- the Configure Try-it panel's agent twin. Pass requestId to test a configured integration (stored config and secret fill in) or the full draft inline (baseUrl/authType/secret/openApiSpec). Nothing is persisted, but it is still a gated write (a real outbound call): poll check_write_status for the statusCode/body result. Test a draft BEFORE proposing its import.

Checking an AI provider

get_ai_provider_availability is a cache-only read: it never resolves a secret or calls the provider. start_ai_provider_check requests a guarded metadata inspection and returns immediately; poll the read tool, and expect awaiting-approval while Review is pending. cancel_ai_provider_check cancels the exact provider/check ID pair returned by start, including a pending approval or retry wait.

Metadata evidence can show that an endpoint responded and that a model ID or alias resolves. It does not execute a completion. Keep text, structured, and classification support unknown when the evidence is only provider-metadata; operation support needs sample-test or an exact adapter-contract entry. Treat stale and not-checked as evidence states, not as proof of provider failure. The checked destination describes the configured endpoint and does not prove where computation runs.

Atlas

Cards are typed by kinds; links by link kinds, unique per (from, to, kind). Mirrored markdown renders in cards (mermaid fences included). Prefer updating existing cards (match by source URL or ID) over creating duplicates; card writes are gated like any write. Kinds themselves are writable via atlas_propose_kind_write (create/update/delete, gated): on update, omit fields to keep the existing schema -- a provided list must retain every saved key, and deleting a kind is refused while any live card uses it.

Conduct

  • Ask the registry, don't assume: step IDs and contracts are the source of truth, and they evolve.
  • Small validated changes over big blind ones; leave workflows publishable (the human runs the published version).
  • Anything ambiguous about intent, data destinations, or approvals is the human's call — park and ask.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.