Skip to content

Shapes of a Skill

A shape is what a skill looks like on disk. Smithy works out the shape from the conversation about purpose, user, and output, and names it in the execution plan with one line on why. You do not need to pick one. If you name a shape up front, Smithy takes it into account and says if the conversation points elsewhere.

ShapeWhat you getYour need lands here when
Plain skillA SKILL.md, with references that load only when neededYou want an agent to do one job well on request
Agent with capabilitiesA named persona with a greeting and a menuYou want someone to work with across turns, who owns related capabilities
Memory agentAn agent that keeps a memory folder and wakes as the same selfThe agent must build up knowledge across sessions
Rendered skillSteps rendered from switches and prose blocksSeveral teams want different steps, and a fork would drift
Script-backed utilityA thin SKILL.md over tested scriptsMost of the job is mechanical

Registration is a separate choice. Any of these shapes can stand alone as a single-skill module, join a module, or stay outside the registry. See Modules and Registration.

Most needs land here, including long multi-step workflows. A plain skill can hold:

FileHolds
SKILL.mdFrontmatter with name and description, then the instructions
references/<topic>.mdA branch that only some runs need, loaded when reached
scripts/<name>.pyWork with one right answer per input, with a test beside it
assets/Templates and files the skill emits
customize.tomlWhat a team can change without forking
bmod.tomlHow the skill registers with BMad

Only SKILL.md is required. A skill small enough to be one file is one file. A complex workflow stays a plain skill when each branch sits in its own reference, and it is the right size at any total if one run loads only what it needs.

A BMad skill resolves its configuration through the project’s _bmad/scripts when it activates. A skill for use outside BMad has no activation section, only plain paths. A skill is one or the other.

An agent is a skill named <prefix>-agent-<name> with a persona and a menu. Choose an agent when a user will work with it across turns and it owns several related jobs. Choose a plain skill when the job has one output and needs no persona.

The persona lives in an [agent] table in customize.toml: name, title, icon, role, identity, communication_style, and principles. The persona guides the agent when no capability fits, so it is written in the agent’s own voice.

Each [[agent.menu]] item has a code, a description, and exactly one of these:

  • skill, when an installed skill does the job and the agent invokes it.
  • prompt, holding the whole instruction in a sentence or two, or a line that reads a file under references/.

A capability needs its own file only when there is know-how to hold. One big enough to stand alone should become its own skill.

The agent’s description lists its name and the phrases of every capability built into it, because nothing else triggers those capabilities. A menu item that runs an installed skill adds no phrases, because that skill triggers itself. In a module, an agent also gets a roster.toml entry, which is how bmad and party mode know it exists.

A memory agent keeps its identity and memory in a project folder and reloads them each session. Build one only when the agent must build up knowledge over time. The test: used daily for a month, would the thirtieth session differ from the first? See Agents with Memory.

A rendered skill keeps its steps as a template with switches and prose blocks. Before the skill runs, render_skill.py fills the template from the skill’s customize.toml merged with team and personal overrides, and the agent reads the result. A team can switch off a step, add a rule, or change wording without editing the skill. bmad-build is this shape.

It is worth the extra machinery only when teams really differ in which steps run. A skill with one set of steps and a few settable values is a plain skill with a customize.toml.

A thin SKILL.md names the outcome and when to run which script, and tested scripts do the work. It fits jobs with one right answer per input: transforming files, counting, checking structure, and generating from fixed rules. The prose keeps the part that needs judgment, which is what to do with the result. See Scripts.

A skill can change shape through edit mode when its needs grow. Adding a menu to a plain skill or memory to an agent are shape changes, and Smithy says so before making them.