Flow.
Plan first, work second.
Retired. Flow was the first plugin to put research and an explicit plan in front of an agent, then keep re-reading that plan between tasks. Its successor, Flow-Next, carries the same discipline further with durable specs, cross-model review, and receipts. This page stays as a record of where the ideas came from.
- Platform
- Claude Code
- Surface
- 5 commands / 6 agents
- Successor
- Flow-Next
- License
- MIT
00 / Status
This one is history.
Flow is retired. The plugin was removed from the marketplace in Flow-Next 1.0.2 and is no longer installable; nothing on this page is a live instruction.
Everything Flow did well is in Flow-Next, which replaced the plan file with a durable spec, the single review pass with cross-model gates at every handover, and “done” with evidence on disk. Start there. The description below is kept for anyone tracing how the workflow got its shape.
01 / Problem
Agents fail for predictable reasons.
Most agent failures trace back to process rather than model capability.
- 01
Coding before context
Starting to code before understanding the codebase.
- 02
Reinventing patterns
Ignoring conventions already in the repo.
- 03
Losing the plan
Forgetting the original plan mid-implementation.
- 04
Missing edges
Obvious-in-hindsight edge cases slipping through.
Flow enforced that discipline. Flow-Next proves the result.
02 / Workflow
/flow:plan → /flow:work
Two steps. Plan produces plans/<slug>.md with references; work executes with discipline.
Plan
- 01
Setup
Choose research depth and review preference.
- 02
Research
Three research agents run in parallel.
- 03
Gap check
Detect missing flows before writing.
- 04
Write plan
References and acceptance checks into plans/<slug>.md.
Work
- 01
Confirm + branch
Clarify blockers; set up branch or worktree.
- 02
Tasks
Turn plan into a TodoWrite list.
- 03
Execute
Loop with plan re-read between tasks.
- 04
Ship
Tests + optional audit, then Definition of Done.
03 / Agents
Specialised research agents.
Each scout has a narrow job; you compose them via /flow:plan.
- 01
repo-scout
Fast pattern finder via Grep, Glob, Read. ~65k tokens.
- 02
context-scout
Deep architecture via RepoPrompt. 30% fewer tokens.
- 03
practice-scout
Community best practices for the task.
- 04
docs-scout
Official documentation for the stack in play.
- 05
flow-gap-analyst
Edge cases and missing user flows.
- 06
quality-auditor
Optional security and risk review.
04 / Review
Cross-model reviews.
When RepoPrompt rp-cli is detected, plan and work ask upfront which review mode. Different models catch different blind spots.
> /flow:plan Add OAuth login ? Review · Run Carmack-level review after? a) RepoPrompt chat b) Export for external LLM c) No > b ... research + planning ... OK Exported to ~/Desktop/plan-review-oauth.md
Export mode creates a markdown file with full context (file tree, code maps, selected files, and the review prompt) that you paste into ChatGPT Pro, Claude web, or any LLM you prefer.
05 / Commands
Surface area.
- 01
/flow:interview
Deep interview to flesh out a spec or bead.
- 02
/flow:plan
Research + produce plan, with optional auto-review.
- 03
/flow:work
Execute plan end-to-end, with optional auto-review.
- 04
/flow:plan-review
Carmack-level plan review via rp-cli.
- 05
/flow:impl-review
Carmack-level implementation review.
06 / Install
Install Flow-Next instead.
/plugin marketplace add \ https://github.com/gmickel/flow-next /plugin install flow-next /flow-next:setup
The Flow plugin was removed from the marketplace in Flow-Next 1.0.2, so its old install commands no longer resolve. Flow-Next is the maintained successor and installs in three lines.
- Flow
- Retired, no longer installable
- Successor
- Flow-Next
- Platforms
- Claude Code, Codex, Droid, Cursor, Grok, OpenCode, Muse Code
- License
- MIT
- Claude Code
- Plugin
- Retired