Flow
Plan first, work second
Most failures come from weak planning or drifting from the plan. Flow fixes both with structured research upfront, explicit plan reuse, and plan re-read between tasks.
5 commands. 6 agents. 7 skills. One disciplined workflow that actually ships.
NOTE — This isn't strictly an app, but it belongs here given my focus on AI-native software development. Flow represents how I think about structured agent workflows.
THE PROBLEM
Most agent failures aren't about model capability. They're about process:
- ×Starting to code before understanding the codebase
- ×Reinventing patterns that already exist in the repo
- ×Forgetting the original plan mid-implementation
- ×Missing edge cases that were obvious in hindsight
Flow enforces the discipline that makes agents reliable.
COMMANDS
/flow:interviewDeep interview to flesh out spec/bead (optional)
/flow:planResearch + produce plan (auto-review if rp-cli)
/flow:workExecute plan end-to-end (auto-review if rp-cli)
/flow:plan-reviewCarmack-level plan review via rp-cli
/flow:impl-reviewCarmack-level impl review (current branch)
CORE WORKFLOW
/flow:planProduces plans/<slug>.md with references and acceptance checks.
- 01Setup: choose research depth + review pref
- 02Run three research agents in parallel
- 03Run flow gap check
- 04Write plan with references + acceptance checks
- 05Auto-review (if opted in)
- 06Offer next step (work, create issue)
/flow:workTakes a plan file and executes it with discipline.
- 01Confirm plan + clarify blockers
- 02Setup branch or worktree
- 03Turn plan into TodoWrite tasks
- 04Execute task loop with plan re-read
- 05Test + optional audit
- 06Ship with Definition of Done
- 07Auto-review (if opted in at start)
HOW IT WORKS
/flow:interview spec.md/flow:plan Add OAuth logingap-analyst15 uses · 56.6k · 1m 46splans/oauth.mdorBeads epic/flow:work plans/oauth.mdAUTO-INTEGRATED WHEN REPOPROMPT rp-cli DETECTED
RESEARCH DEPTH
When RepoPrompt rp-cli is detected, /flow:plan asks which research approach to use before starting.
repo-scoutFAST- Tools: Grep, Glob, Read
- Speed: ~45 seconds
- Tokens: ~65k
- Output: Line refs, conventions
- Best for: Quick bug fixes, simple changes
context-scoutDEEP- Tools: RepoPrompt rp-cli
- Speed: Slower (builder takes time)
- Tokens: ~45k (30% less)
- Output: Architecture, function sigs
- Best for: Complex features, architecture
Without RepoPrompt: repo-scout is used automatically—no prompt needed.
BUILT-IN REVIEWS
When RepoPrompt rp-cli is detected, both /flow:plan and /flow:work ask upfront which review mode to use:
Review via rp-cli chat with a different model (GPT-5.2 High recommended)
Export context to file for ChatGPT Pro, Claude web, or any LLM
Skip review entirely
CROSS-MODEL — Different models catch different blind spots. Same-model self-review misses things a fresh perspective finds.
EXPORT MODE — Creates a markdown file with full context: file tree, code maps, selected files, and the Carmack-level review prompt. Opens automatically so you can paste into your preferred LLM.
> /flow:plan Add OAuth login
? Review — Run Carmack-level review after?
a) RepoPrompt chat
b) Export for external LLM
c) No
> b
... (research + planning) ...
✓ Exported to ~/Desktop/plan-review-oauth.mdManual invocation still available:
/flow:plan-review plans/oauth.md/flow:impl-reviewBEADS INTEGRATION
Flow optionally integrates with Beads (bd) for dependency-aware issue tracking. When .beads/ exists, Flow adapts automatically.
Standard Mode
- Plans →
plans/<slug>.md - Tasks →TodoWrite
- Context →Plan file
Beads Mode
- Plans →Epic with child tasks
- Tasks →
bd ready / bd close - Context →
bd showoutput
# With Beads configured
/flow:plan Add OAuth login
# Creates epic bd-a3f8e9 with child tasks
/flow:work bd-a3f8e9
# Tracks progress via bd ready/update/close- ›No configuration—detects
.beads/automatically - ›Graceful fallback—works without Beads installed
- ›Dependency tracking—knows which tasks block others
SPECIALIZED AGENTS
repo-scoutSearches codebase with Grep/Glob/Read for patterns and conventions. Fast, ~65k tokens.
context-scoutUses RepoPrompt AI builder for architecture understanding. Slower but 30% fewer tokens.
practice-scoutResearches community best practices and proven approaches for your specific task.
docs-scoutFetches relevant official documentation for frameworks, libraries, and APIs involved.
flow-gap-analystIdentifies edge cases, missing user flows, and potential gaps in your specification.
quality-auditorOptional security and quality review to catch issues before they ship.
SKILLS
Progressive disclosure: ~100 tokens at startupinterviewDeep interview to flesh out specs/beads
flow-planPlanning workflow logic
flow-workExecution workflow logic
flow-plan-reviewPlan review via rp-cli + chat
flow-impl-reviewImpl review via rp-cli + chat
worktree-kitSafe parallel git workspaces
rp-explorerToken-efficient codebase exploration
Skills auto-trigger from natural language. Say "help me plan adding OAuth" and Claude triggers flow-plan automatically.
BEYOND PLAN → WORK
Flow's core workflow is /flow:plan → /flow:work, but you can invoke individual components directly for specific tasks.
- Run in current conversation
- See full chat history
- Trigger via natural language
- Run as isolated subprocess
- Fresh context per task
- Trigger via "Use X to..."
Skill Triggers
interview"interview me about this spec"rp-explorer"use rp to explore how auth works"flow-plan"help me plan adding OAuth"worktree-kit"create a worktree for feature X"Agent Invocation
context-scout"Use context-scout to understand the payment system"repo-scout"Use repo-scout to find all API endpoints"practice-scout"Use practice-scout for React Server Components"Quick Reference
/flow:interview/flow:plan/flow:work/flow:plan-review/flow:impl-reviewRule of thumb: Use skills for inline help, agents for delegated deep dives.
INSTALLATION
/plugin marketplace add https://github.com/gmickel/gmickel-claude-marketplace/plugin install flowclaude --plugin-dir ./plugins/flowClone the marketplace repo and point Claude Code at the plugin directory for development.
OPTIONAL — Install RepoPrompt for auto-review. Flow detects rp-cli and offers integrated Carmack-level reviews.
- ›Plans live in
plans/or Beads epics - ›Prefer reuse of centralized code
- ›Tests and linting are part of the plan
- ›Issues auto-detect GitHub/Linear/Beads
Developers who want Claude Code to ship reliably, not just generate code. If you've ever had an agent "finish" a task only to realize it forgot half the requirements—this is for you.
Flow doesn't make Claude smarter. It makes the workflow disciplined enough that capability translates to results.