dont-b-mad

A fork of BMAD v6.3.0 with AI tracking baked into every workflow. Measures AI adoption across the full SDLC without adding any overhead to developers.

Works with Cursor and Claude Code.

What This Adds

Every git commit gets three trailers recording what phase of work it represents and whether AI was involved:

feat: implement wave planning task assignment

AI-Phase: code
AI-Tool: cursor/claude-sonnet-4-20250514
Story-Ref: 1-1-wave-planning

Trailers are appended automatically. BMAD workflows fill them with the actual tool/model used. A git hook catches manual commits and tags them with AI-Tool: manual. Nobody types trailers by hand.

A dashboard script (Pulse) reads git history and prints adoption rates grouped by phase:

======================================
  Pulse — AI Adoption Dashboard
======================================

  PLANNING (3 commits)
  --------------------------------
  prd                  100%  (target: 90%)  [2/2]
  story                100%  (target: 90%)  [1/1]

  DEVELOPMENT (8 commits)
  --------------------------------
  code                  75%  (target: 80%)  [6/8]
  test                  50%  (target: 85%)  [2/4]
  review                62%  (target: 95%)  [5/8]

  TOTAL: 11 tracked commits
======================================

Install

Skills publish to the user-level skill directories (~/.claude/skills, ~/.cursor/skills) so a single copy serves every workspace. Workspace-specific files (rules, project registry, team config, dashboard) install at the workspace root, and git hooks install into each repo inside the workspace.

git clone https://github.com/ElasticRun/dont-b-mad.git

# Full install: skills (user-level) + workspace files + hooks in every repo
bash dont-b-mad/scripts/install.sh ~/Workspace

# Skills + workspace files, no hooks
bash dont-b-mad/scripts/install.sh ~/Workspace --skills-only

# Hooks only (into repos discovered inside the workspace)
bash dont-b-mad/scripts/install.sh ~/Workspace --hooks-only

# Skills only (no workspace files)
bash dont-b-mad/scripts/install.sh --global

Run with --dev-link (or simply run from inside this repo with bash scripts/install.sh .) to symlink user-level skills back to the source so edits apply live.

The installer also installs and initializes Git AI (agent hooks via git-ai install-hooks). Restart your terminal and IDE after install so git-ai is on your PATH. Use git ai stats and git ai blame for line-level attribution. Set DONTBMAD_SKIP_GIT_AI=1 to skip the download step.

If a prepare-commit-msg hook already exists in a repo, the installer backs it up to .bak before replacing.

What Gets Installed

Location	What	Scope
~/.cursor/skills/bmad-*, ~/.cursor/skills/dontbmad-*	All BMAD + custom skills (Cursor)	User home
~/.claude/skills/bmad-*, ~/.claude/skills/dontbmad-*	All BMAD + custom skills (Claude Code)	User home
~/.claude/commands/bmad-*.md, ~/.claude/commands/dontbmad-*.md	Slash-command symlinks pointing at each skill's SKILL.md	User home
.cursor/rules/bmad-workspace-resolution.md	Teaches agent how to resolve {project-root}	Workspace root
.cursor/rules/bmad-team-customization.md	Teaches agent to read custom team names	Workspace root
.cursor/rules/dontbmad-graph-first.md	Prefer knowledge graph over reading full source	Workspace root
.cursor/rules/dontbmad-caveman-activate.md	Always-on terse output (caveman mode)	Workspace root
.claude/rules/bmad-*.md, .claude/rules/dontbmad-*.md	Same rules for Claude Code	Workspace root
_bmad/workspace.yaml	Maps project directories in the workspace	Workspace root
_bmad/_config/team.yaml	Custom agent display names	Workspace root
scripts/adoption-dashboard.sh	Reads git trailers, prints adoption rates	Workspace root
<repo>/.git/hooks/prepare-commit-msg	Auto-tags manual commits with AI trailers	Per repo
~/.git-ai/bin/git-ai	Git AI — line-level AI attribution in commits	User home

Workspace layout example

~/.claude/skills/bmad-*/           <- skills published once at user level
~/.cursor/skills/bmad-*/           <- (same)
~/.claude/commands/bmad-*.md       <- slash-command symlinks

~/Workspace/                       <- open Cursor / Claude Code here
├── .cursor/rules/bmad-*.md        <- workspace, team, graph-first, caveman rules
├── .claude/rules/bmad-*.md        <- workspace, team, graph-first, caveman rules
├── _bmad/
│   ├── workspace.yaml             <- project registry
│   └── _config/team.yaml          <- custom agent names
├── scripts/adoption-dashboard.sh  <- dashboard
├── project-a/                     <- git repo + BMAD project
│   ├── _bmad/bmm/config.yaml     <- project's own config + output paths
│   └── .git/hooks/prepare-commit-msg
├── project-b/                     <- git repo + BMAD project
│   ├── _bmad/bmm/config.yaml
│   └── .git/hooks/prepare-commit-msg
└── docs/                          <- not a git repo, skipped

Multi-project output isolation

Each project keeps its own _bmad/ config tree. Output paths like {planning_artifacts} and {implementation_artifacts} are resolved from that project's config.yaml, so artifacts stay inside the project that produced them.

The workspace-level resolution rule (.cursor/rules/ and .claude/rules/) teaches the agent to pick the right {project-root} based on which files are being discussed. If ambiguous, the agent asks. You can set a default_project in _bmad/workspace.yaml to skip the prompt.

After initializing BMAD in a new project, re-run the installer with --force to refresh the registry:

bash dont-b-mad/scripts/install.sh ~/Workspace --force

Customize your team

Every BMAD agent has a default display name. To rename them, edit _bmad/_config/team.yaml (installed automatically by the installer):

agents:
  dev: Arjun
  pm: Priya
  architect: Kiran
  analyst: Meera
  tech-writer: Ravi
  ux-designer: Ananya
  brainstorming: Vikram
  problem-solver: Deepak
  design-thinking: Kavita
  innovation: Nitin
  presentations: Pooja
  storyteller: Rohit

Each key maps to an agent role. Change the name and it takes effect immediately -- no reinstall needed. Agents without an entry keep their default name from the skill files.

Token Compression (Caveman Mode)

Built-in output compression based on caveman by Julius Brussee. Cuts agent output tokens by ~75% and artifact input tokens by ~46% without losing technical substance.

Three pieces ship with the fork:

Skill	What it does
/dontbmad-caveman	Switches agent to terse mode. Levels: lite, full (default), ultra. Say "stop caveman" to revert.
/dontbmad-compress-artifacts	Compresses planning docs (PRDs, architecture, stories) for cheaper agent reads. Originals saved as .original.md.
--caveman flag on party mode	All subagents respond terse. Combine with --model haiku for max savings.

The activation rule (dontbmad-caveman-activate.md) is installed to .cursor/rules/ and .claude/rules/ by the installer, making caveman always-on by default. Delete the rule file to disable.

Code blocks, file paths, commands, and BMAD deliverable artifacts (PRDs, stories) are always written in normal prose.

Dashboard Usage

# Current repo
bash scripts/adoption-dashboard.sh

# Specific repo
bash scripts/adoption-dashboard.sh --repo ./project-a

# All repos in the workspace
bash scripts/adoption-dashboard.sh --workspace

# All repos in a specific workspace path
bash scripts/adoption-dashboard.sh --workspace ~/Workspace

# With Story-Ref filter
bash scripts/adoption-dashboard.sh --workspace "1-*"

Trailers Reference

Three trailers per commit. One commit = one phase of work.

Trailer	Records	Values
AI-Phase	What phase this commit belongs to	prd, architecture, ux-design, epics, sprint-plan, story, code, test, review, deploy
AI-Tool	AI tool/model used, or manual	Tool/model identifier (e.g. cursor/claude-sonnet-4-20250514), or manual
Story-Ref	What story or artifact this belongs to	Story key (e.g. 1-1-wave-planning) or artifact ref (e.g. prd-aieye)

How It Flows

create-story  -->  commits with AI-Phase: story
     |
  dev-story   -->  commits with AI-Phase: code
     |
 code-review  -->  commits with AI-Phase: review
     |
 retrospective --> queries git trailers; surfaces adoption metrics by phase

Manual commits (hotfixes, config changes) get auto-tagged by the git hook with AI-Tool: manual.

Modified Workflows (from upstream BMAD v6.3.0)

Planning workflows (auto-commit artifacts with AI trailers on completion):

• bmad-create-prd -- commits PRD
• bmad-create-epics-and-stories -- commits epics
• bmad-create-architecture -- commits architecture doc
• bmad-create-ux-design -- commits UX design
• bmad-sprint-planning -- commits sprint status

Development workflows (AI Engineering Record + commit trailers):

• bmad-create-story -- AI Engineering Record table in template, commits story on creation
• bmad-dev-story -- fills record rows, creates commits with trailers, checklist updated
• bmad-code-review -- fills review row, creates review commit with trailers
• bmad-quick-dev -- appends trailers to commits (both step-05 and one-shot paths)
• bmad-retrospective -- queries git for AI adoption metrics by phase, includes in retro output
• dontbmad-ai-tracking -- new skill: hook template, dashboard, install instructions
• dontbmad-graphify -- new skill: knowledge graph setup, query reference, workflow integration docs
• dontbmad-caveman -- new skill: terse output mode (~75% token reduction), based on caveman
• dontbmad-compress-artifacts -- new skill: compress planning artifacts for cheaper agent reads (~46% input token savings)
• dontbmad-grill -- new skill: relentless one-question-at-a-time interrogation of a plan, with a recommended answer for every decision; integrated into bmad-create-architecture step 4 as a G option in the A/P/G/C menu, and into bmad-create-story step 5b as an auto-invoked ambiguity-resolution gate (skipped when the story's analysis surfaced no open questions). Default intensity is standard for direct invocation, light for auto-invocation. Adapted from grill-me by Matt Pocock.
• bmad-party-mode -- added --caveman flag for terse multi-agent roundtables

Credits

Built on BMAD v6.3.0 by the BMAD community. This fork adds the AI tracking layer. All upstream skills are included unmodified except where noted above.

License

MIT. See LICENSE.