llm-flow

Deploy the /implement workflow templates to Claude and Gemini.

llm-flow is a small CLI that installs a consistent, opinionated software-development workflow (Analyze → Design → Gate → Implement → Tests → Review → Done) into your AI coding assistants. The workflow tracks state in Jira via the MCP — not via a Python backend.

What it does

• Deploys commands/implement.md + commands/implement/init/*.md into .claude/commands/ and .gemini/commands/ of a project or user directory.
• Converts Markdown to Gemini TOML automatically.
• Works cross-OS (Linux, macOS, Windows).
• Supports updating templates from GitHub without reinstalling the package.

What it does not do

• It does not call Jira directly. The /implement templates instruct your LLM (Claude, Gemini) to use your MCP Jira server for all ticket operations (status, labels, comments, search). Configure your MCP separately in your assistant's settings.
• It does not adapt to a specific programming language. Templates are language-agnostic — build/test/lint commands come from each project's own CLAUDE.md.

Installation

pip install llmflow-cli

Quick start

# First run triggers an interactive setup wizard (target, platforms)
llm-flow setup

# Deploy into the current project (.claude/, .gemini/)
llm-flow deploy

# Or deploy into user-level directories (~/.claude/, ~/.gemini/)
llm-flow deploy --target user

# Preview what would happen
llm-flow deploy --dry-run

# Update templates from GitHub, then deploy
llm-flow deploy --update

# Show or modify configuration
llm-flow config --show
llm-flow config --set-target user

Configuration

Two levels:

• Global (user-level): ~/.config/llm-flow/config.yaml (Linux/macOS) / %APPDATA%\llm-flow\config.yaml (Windows).
• Project (./.llm-flow.yaml): overrides global settings for a specific repo.

Schema is minimal:

default_target: project        # or "user"
platforms:
  - claude
  - gemini

Coverage thresholds, language choices, or Jira credentials are not managed by llm-flow — they belong in each project's own CLAUDE.md and in your MCP Jira server configuration respectively.

Updating templates

Templates are bundled with the installed wheel, but you can pull newer versions from GitHub without reinstalling:

llm-flow update            # fetch latest
llm-flow update --force    # re-download everything
llm-flow deploy --update   # update then deploy

Cache location: ~/.config/llm-flow/templates/ (or %APPDATA%\llm-flow\templates\).

CLI reference

Command	Role
llm-flow version	Show version
llm-flow setup	First-time interactive setup wizard
llm-flow init	Create a .llm-flow.yaml in the current project
llm-flow config [--show] [--set-target X] [--set-platforms X,Y]	Inspect or edit configuration
llm-flow deploy [--target] [--platform] [--dry-run] [--force] [--update] [--no-cache] [--quiet]	Deploy templates
llm-flow update [--force] [--quiet]	Pull latest templates from GitHub into the cache

The /implement workflow

Once deployed, use these commands from Claude or Gemini:

Command	Role
/implement	Main continuous flow (phases 1–6)
/implement:init:1-project	Discovery conversation, initialize Jira epics and tickets
/implement:init:2-architecture	Define technical architecture
/implement:init:3-standards	Define code conventions

Phases set a label on the Jira ticket (via your MCP): phase:analyze, phase:design, phase:impl, phase:tests, phase:review, phase:done. Status transitions (To Do → In Progress → Done) are separate and happen at the start of Phase 1 and the end of Phase 6.

All ticket types are supported (Story, Bug, Task, Subtask, etc.).

MCP Jira

llm-flow assumes your LLM has an MCP Jira server configured. The templates invoke it in natural language ("via the MCP Jira, fetch the next ticket of epic X") — no specific tool names are hardcoded, so any MCP Jira implementation works.

Configure your MCP Jira in your Claude/Gemini settings separately.

Development

git clone https://github.com/mxdumas/llm-workflow
cd llm-workflow
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

pytest -v
ruff check src/
ruff format src/

The templates live at commands/ at the repo root. Editing them in place, then running llm-flow deploy --no-cache in a test project is the fastest feedback loop.

License

MIT