Momentum

A practice framework for agentic engineering.

Momentum is a philosophy and process for building software with AI agents as primary code producers. It defines how specifications govern code generation, how quality is enforced when AI writes the code, and how the practice itself improves over time. The principles apply whether you're a solo developer or a team — anyone directing AI agents to produce code faces the same quality challenges.

Momentum is currently implemented using BMAD Method and Claude Code, but the principles and process are tool-agnostic. Any agentic coding tool and workflow framework could serve as the implementation layer.

---

Quick Start

Claude Code (Tier 1 — full enforcement):

npx skills add https://github.com/iamsteveholmes/momentum --all

Then in Claude Code:

/momentum

Impetus (the Momentum orchestrator) runs first-time setup: installs global rules, configures hooks, sets up MCP integrations, and orients you to available workflows.

Cursor or other tools (Tier 2 — advisory):

npx skills add https://github.com/iamsteveholmes/momentum -a cursor

Skills install and are invocable immediately. No additional setup required.

No tooling (Tier 3 — philosophy only):

Read the Principles section below and the Philosophy section for the full framework. No installation needed.

---

Enforcement Tiers

Momentum operates at three tiers depending on your tool environment. The same skill files install at every tier — what changes is the enforcement level.

Tier 1: Full Deterministic — Claude Code

Claude Code provides the complete enforcement stack:

• Hooks fire automatically on file changes — linting, formatting, and file protection run without developer action
• Rules auto-load into every session from ~/.claude/rules/ — authority hierarchy, anti-pattern rules, and coding standards are always in context
• Subagent isolation via context: fork — verification agents run in separate contexts with restricted tool access, enforcing producer-verifier separation
• Model routing — skills specify model: and effort: frontmatter; Claude Code routes to the appropriate model tier
• Impetus orchestration — /momentum provides session orientation, sprint awareness, install/upgrade management, and workflow access

What "full deterministic" means: Quality standards at the deterministic enforcement tier (hooks, test gates, file protection) execute automatically. They cannot be skipped, deprioritized, or forgotten. Structured enforcement (workflow steps, review checklists) executes as part of skill workflows. Advisory enforcement (rules in context) is always loaded.

Install: npx skills add https://github.com/iamsteveholmes/momentum --all, then /momentum

Tier 2: Advisory — Cursor and Other Tools

Tools that support the Agent Skills standard can install Momentum skills:

• Skill instructions execute — all SKILL.md files are available and functional
• Guidance is advisory — the tool provides instructions to the agent, but enforcement depends on the agent following them
• Extra frontmatter is silently ignored — context: fork, model:, effort:, and allowed-tools fields are Claude Code-specific; other tools skip them without error

What is NOT available at Tier 2:

• Hooks (no automatic linting, formatting, or file protection)
• Global rules (no ~/.claude/rules/ auto-loading)
• Subagent isolation (no context: fork — all work happens in a single context)
• Model routing (no model: frontmatter enforcement)
• Impetus orchestration (no /momentum session management)

Skills still provide significant value at Tier 2 — workflow structure, acceptance criteria enforcement, review checklists, and quality guidance all function as advisory instructions.

Install: npx skills add https://github.com/iamsteveholmes/momentum -a cursor

Tier 3: Philosophy Only — No Tooling

Momentum's principles are designed to be valuable without any tooling. A developer or team can adopt the practice by reading the documentation: