Skip to main content
BMad Builder generates the agents, skills, and workflows your team would otherwise hand-write. Here is the short version.

BMad Builder - Build Agents You Can Trust

BMad Builder generates the agents, skills, and workflows your team would otherwise hand-write. Here is the short version.

Mor Dabastany By Mor Dabastany • • 694 words • 4 min
ai agents agentic-ai agentic-sdlc bmad skills claude

TL;DR

BMad Builder is the generator inside the BMad Method ecosystem. You describe a capability through a guided conversation, it produces the agent, skill, workflow, or module - compliant, linted, ready to install. You stop writing SKILL.md files and bespoke memory layouts by hand. The Builder writes them for you.

The point isn’t just to save typing. It’s domain encoding: capturing your team’s expertise, processes, and voice into reusable agents, instead of letting that knowledge live in five people’s heads and ten Slack threads.

Everything it outputs follows the Agent Skills open standard, so the artifact drops straight into .claude/skills, .codex/skills, .agent/skills, or any compliant host.

A small note up front: as of today, I’m part of a small group inside CyberArk by Palo Alto Networks leading a change in ASDLC (Agentic SDLC), shaping a process close to BMad. So this post is both a primer and a reference I’ll keep coming back to.

What It Builds

Four things, all packaged as Agent-Skills-compliant folders:

  • Skills - the universal unit. A folder with a SKILL.md and optional resources/, scripts/, templates/.
  • Agents - skills with persona, capabilities, and optional persistent memory.
  • Workflows - structured, sequential processes.
  • Modules - shareable bundles of the above.

You don’t author these from scratch. You run a builder, answer questions, get the folder.

Agents That Remember

Agents come in three flavors: stateless, memory (persistent identity across sessions), and autonomous (memory plus scheduled headless wake-ups).

The clever bit: memory lives in your project at _bmad/memory/<agent-name>/, not in the skill folder. The skill stays portable; your context stays yours. Each session, the agent loads its identity, mission, and curated long-term memory, then greets you. If it doesn’t remember something, it says so - no faked continuity.

A discipline worth flagging early: every memory file loads every session, so every token pays rent. Long-term memory is kept ruthlessly small (typically under 200 lines) through active curation. Raw session logs sit in a separate folder, get distilled into the long-term file, and are pruned after 14 days. That is intentional, not an oversight.

The Builder Commands

One generator per artifact. That’s the whole story:

CommandBuilds
BA (Agent Builder)Agents
BW (Workflow Builder)Workflows and utilities
CW (Convert)Converts any existing skill into a BMad-compliant one
QO (Quality Optimize)Validates and improves agents/workflows
IM (Ideate Module)Plans a module
CM (Create Module)Packages skills into a module
VM (Validate Module)Checks a module is clean

Every builder is conversation-driven - you don’t fill out config files, you answer questions. Two modes: Guided for production-quality output, YOLO for quick prototypes. All builders support --headless so you can script them. You’re not stuck in a chat window.

Building a Module, End to End

Four steps:

  1. IM - ideate the module, get a plan doc.
  2. BA and BW - build the individual skills.
  3. CM - package them into an installable module.
  4. VM - validate. Re-run until clean.

Done. Push to Git. There is no proprietary package format - a module is a Git repo with a .claude-plugin/marketplace.json manifest at the root. Consumers install with:

npx bmad-method install --custom-source https://github.com/your-org/my-module \
  --tools claude-code --yes

GitHub, GitLab, Bitbucket, self-hosted, or a local path. All work.

Why It Matters

The whole point: stop building agentic infrastructure by hand. Hand-rolled prompt frameworks don’t scale - someone always forgets the right preamble, the persona drifts, the memory model is one-off, the packaging is bespoke. BMad Builder replaces that with a generator and a standard format. What comes out is versioned, reviewed, and distributed like normal software, not like a forum of clever prompts.

That is exactly the substrate we need for ASDLC at CyberArk - and the reason I’m spending real time on it.

If you want to feel the difference, give it a weekend:

  • Install BMad, run bmad-bmb-setup
  • IM a module, BA/BW a couple of skills
  • CM and VM to package
  • Push to a private repo, install with --custom-source

You’ll learn more in one full lap than from ten posts (including this one) describing it.

Reference

Related Posts