CLI and Skill #

Manual is meant to be used by both people and agents.

The CLI gives agents a stable local tool surface. The skill gives agents the instructions and templates needed to turn natural language requests into Manual workflow actions.

Current CLI #

The current cli crate supports a small command surface:

cargo run -p cli -- about
cargo run -p cli -- validate-skill crates/skill/templates/default-skill

The about command reports the workspace name and package list. The validate-skill command checks that a skill directory contains a SKILL.md file.

Current Skill Binary #

The skill crate provides the manual-skill binary:

cargo run -p skill -- template-path
cargo run -p skill -- validate-bundled

It can report the bundled template path and validate that the built-in skill template has the required entrypoint.

Bundled Skill Template #

The default template lives at:

crates/skill/templates/default-skill/SKILL.md

It defines the minimum structure for a Codex or Claude-style agent skill package:

• SKILL.md as the required instruction entrypoint.
• references/ for supplemental human and agent guidance.

Planned Manual CLI #

The product direction calls for a Rust manual CLI that can manage workflows, jobs, runs, costs, artifacts, sandbox policies, and local visualization.

Likely commands:

manual init
manual workflow list
manual workflow show <workflow-id>
manual workflow validate <file>
manual workflow import <file>
manual workflow export <workflow-id>
manual job run <workflow-id>
manual job list
manual job show <job-id>
manual cost report <job-id>
manual artifact list <job-id>
manual sandbox run --policy <policy> -- <command>
manual serve

The current crate names are still simple (core, node, cli, skill, agent, script, sandbox, app). A later productization step can rename binaries and packages when the command surface is ready.

Skill Responsibility #

The Manual Skill should help an agent:

1. Understand the repeatable work the user wants to automate.
2. Ask only the minimum clarifying questions required to produce a reusable workflow.
3. Write a human-readable workflow spec.
4. Validate or import the workflow with the Manual CLI.
5. Run jobs when requested.
6. Summarize artifacts, costs, failures, and next improvements.

The user should not need to learn low-level workflow syntax before receiving value.