Claude Decide

Autonomous product operator for Claude Code -- researches your codebase, generates PRDs, and builds improvements in continuous cycles.

Installation

In Claude Code, run:

/plugin install maxwhoppa/claude-decide

This installs the plugin with its skills (decide and decide-loop) directly into Claude Code.

Alternative: Manual Installation

Clone the repo and run the install script to copy skills into ~/.claude/skills/:

git clone https://github.com/maxwhoppa/claude-decide.git
cd claude-decide
bash install.sh

Quick Start

Open Claude Code in any project directory and run:

/decide

On first run, the operator analyzes your repo and walks you through a short onboarding interview to understand your product, priorities, and constraints. After that, it begins its first research cycle.

Invocation Modes

/decide -- Interactive (default)

Runs one operator cycle at a time. You review and approve PRDs before execution, and re-invoke /decide between cycles.

/decide

/decide force -- Autonomous

Runs one full cycle without user interaction. PRDs are auto-approved, open questions are resolved by a subagent, and the collaboration phase is skipped.

/decide force

You can also run force mode from the terminal using the launcher script:

bash skills/decide/scripts/launcher.sh --force
bash skills/decide/scripts/launcher.sh --force --max-cycles 10

/decide auto -- Confidence-Gated

A middle ground between interactive and force mode. PRDs with a backlog priority score above a configurable threshold are auto-approved; lower-scored PRDs go through the interactive collaboration phase.

/decide auto              # threshold 0.75 (default)
/decide auto 0.8          # threshold 0.80

Also available from the terminal:

bash skills/decide/scripts/launcher.sh --auto
bash skills/decide/scripts/launcher.sh --auto=0.8

/decide-loop -- Continuous

Runs /decide force in a continuous loop. Each cycle executes in a fresh Claude session via claude -p, ensuring full superpowers skill access (brainstorming, writing-plans, executing-plans) on every cycle.

/decide-loop              # run until stopped (default max: 50 cycles)
/decide-loop 10           # run 10 cycles then stop
/decide-loop edit         # review backlog before starting loop
/decide-loop 5 edit       # review backlog, then run 5 cycles

The edit flag presents all queued backlog items for review before starting. You can approve, reject, or reprioritize items.

To stop gracefully between cycles:

touch .claude-operator/stop

Utility Commands

| Command | Description | |---------|-------------| | /decide status | Show current cycle, phase, mode, and backlog summary | | /decide backlog | Show all queued items sorted by priority | | /decide reset | Delete .claude-operator/ and start fresh (offers to preserve backlog) | | /decide rollback | Revert the most recent cycle's commit via git revert |

Operator Cycle Phases

Each cycle moves through these phases sequentially:

| Phase | What happens | |-------|-------------| | Research | Dispatches 3-5 parallel subagents (adaptive selection based on product stage) to audit code quality, find product gaps, check security, analyze market fit, and more. Synthesizes findings into a prioritized backlog. | | Propose | Generates a full PRD for the highest-priority backlog item, then runs it through a 5-lens self-critique. Sanitizes filenames to prevent path traversal. | | Collaborate | (Interactive/auto mode only) Presents the PRD for your review. You can approve, request changes, reject, redirect, or ask for more research. | | Execute | Implements the PRD via the superpowers pipeline: brainstorming, writing-plans, executing-plans. Creates a stuck.json report if it can't resolve after 10 attempts. | | Update Memory | Records results in memory.json, updates the backlog, writes a cycle log, commits all state changes, and optionally collects a satisfaction rating (1-5). |

In force mode, the Collaborate phase is skipped and PRDs are auto-approved. In auto mode, only PRDs below the confidence threshold go through Collaborate.

State Validation

Run the validation script to check structural integrity of all operator state files:

bash skills/decide/scripts/validate.sh

Checks JSON syntax, enum values, required fields, and cross-references. Exits 0 if all checks pass, 1 if any fail.

Project Structure

claude-decide/
  install.sh                          # copies skills to ~/.claude/skills/
  skills/
    decide/
      SKILL.md                        # full phase router and operator logic
      prompts/                        # prompt templates for subagents
        execution.md                  # execution constraints
        onboarding-repo-analysis.md   # repo analysis prompt
        prd-template.md               # PRD generation template
        prd-critic.md                 # 5-lens PRD critique
        research-*.md                 # research agent prompts
        state-templates.md            # JSON schemas for state files
      scripts/
        launcher.sh                   # terminal launcher for force/auto mode
        validate.sh                   # state file validation script
    decide-loop/
      SKILL.md                        # continuous loop skill definition

When the operator runs in a target project, it creates a .claude-operator/ directory:

your-project/
  .claude-operator/
    state.json       # current phase, cycle number, mode
    memory.json      # product context, features, constraints, decisions
    backlog.json     # prioritized improvement ideas
    prds/            # generated PRD documents
    logs/            # per-cycle result logs
    inputs/          # drop customer feedback, research, etc. here
    agents/          # custom .md research agent prompts
    experiments/     # experiment tracking

Customization

Custom research agents: Add .md files to .claude-operator/agents/ in your target project. Each file is used as a prompt template and dispatched alongside the standard research agents during the Research phase.

User inputs: Drop any files (markdown, text, JSON) into .claude-operator/inputs/ to feed external context -- customer feedback, competitor analysis, user interviews -- into the research phase.

Stopping the Operator

• Interactive mode: Simply stop re-invoking /decide.
• Force mode / Loop: Run touch .claude-operator/stop from another terminal. The operator finishes the current cycle and exits.
• Ctrl+C: Stops the launcher immediately (force mode via terminal only).
• Stuck: If the execution subagent fails after 10 attempts, the operator pauses and waits for you to unblock it.

License

See repository for license details.