13 open-source plugins that give Claude Code superpowers — session memory, video understanding, disciplined engineering, cost tracking, rollback, scope control, and more.
Quick Start • Plugins • Contributing • License
- Python 3.10+
- uv — manages all dependencies automatically
curl -LsSf https://astral.sh/uv/install.sh | shFor video-insight, you also need ffmpeg:
# macOS
brew install ffmpeg
# Ubuntu / Debian
sudo apt install ffmpeg
# Windows (chocolatey)
choco install ffmpegInside Claude Code, run:
/plugin marketplace add gowtham012/Claude-plugins
/plugin install carry-forward@Claude-plugins
/plugin install video-insight@Claude-plugins
/plugin install deep-engineer@Claude-plugins
/plugin install cost-guard@Claude-plugins
/plugin install rubber-duck@Claude-plugins
/plugin install rollback@Claude-plugins
/plugin install regression-sentinel@Claude-plugins
/plugin install time-capsule@Claude-plugins
/plugin install scope-lock@Claude-plugins
/plugin install diff-narrator@Claude-plugins
/plugin install dependency-doctor@Claude-plugins
/plugin install pair-mode@Claude-plugins
/plugin install codebase-onboard@Claude-plugins
Install any combination — each plugin works independently.
Every time you start a new Claude Code session, it forgets everything — what you were building, decisions you made, files you touched, what comes next. carry-forward fixes that permanently.
It auto-saves a structured summary of your work after every Claude response and auto-loads it when you start a new session. Zero friction, zero commands after the one-time setup.
Session 1 Disk Session 2
─────────────────── ─────────────────── ───────────────────
/carry-forward:setup --> CLAUDE.md gets Open Claude Code
@carry-forward/ Claude reads context.md
... work normally ... context.md automatically at startup
Stop hook fires --> log.jsonl grows Claude already knows:
after every response (auto, silent) - Current task
- Files in play
/carry-forward:save --> context.md updated - Decisions made
(rich summary) - Next steps
Close Claude Code Just say "continue"
/carry-forward:setup
This creates a carry-forward/ directory in your project, wires up CLAUDE.md with an @import, and you're done. Restart Claude Code once — from then on, context loads automatically at every session start.
---
last_saved: 2026-03-07T10:30:00Z
project: my-app
---
## Current Task
Implementing JWT refresh flow in useAuth.ts
## Files Being Worked On
- src/hooks/useAuth.ts — adding silent token refresh
- src/middleware/auth.py — fixing exp field parsed as string
## Key Decisions
- JWT stored in httpOnly cookie (not localStorage) — security
- React Context over Redux — simpler scope for this feature
## Next Steps
1. Add refresh_token field to auth state
2. Wire up axios interceptor for silent refresh
## Blockers
- Token refresh strategy (silent vs. re-login) undecided| Skill | What it does |
|---|---|
/carry-forward:setup |
One-time setup — creates directory, wires CLAUDE.md @import |
/carry-forward:save |
Save a rich structured summary of the current session |
/carry-forward:load |
Review what was saved from your last session |
/carry-forward:clear |
Reset context and archive the log |
your-project/
├── CLAUDE.md <-- @carry-forward/context.md added here
└── carry-forward/
├── context.md <-- auto-loaded every session start
├── log.jsonl <-- auto-appended after every response
└── log.2026-03-07T10-30.jsonl.bak <-- created by /clear
- Python 3.10+
uvx(installed automatically withuv)
video-insight deeply understands video recordings. Point it at any screen recording, product demo, 3D walkthrough, or marketing video and it extracts 13 structured signals — then use 17 skills to turn that data into frontend code, design specs, tests, PRDs, user flows, changelogs, Storybook stories, and more.
No API keys needed. No cloud uploads. Everything runs locally.
| Signal | Description |
|---|---|
| Keyframes | One representative frame per scene, extracted via scene detection |
| OCR text | Every string visible on screen, timestamped and organized by scene |
| Color palette | Hex codes ranked by dominance + semantic token assignments (background/surface/accent/text) |
| Motion events | Per-scene animation classification (scroll, fade, slide, cut, none) |
| Transcript | Full speech-to-text with timestamps via faster-whisper |
| Cursor path | Click positions and movement tracking across scenes |
| Scroll indicators | Scroll events detected per scene |
| Loading states | Spinners, skeleton loaders, and progress bars detected |
| Burst frames | High-FPS capture of fast animations for precise reproduction |
| Scene diffs | Pixel-level visual change detection between consecutive keyframes |
| Font hints | Font size, weight, and role estimation from text bounding boxes |
| Confidence scores | Reliability rating per extracted signal |
| Video hash | SHA-256 deduplication — skips re-analysis of unchanged files |
| Skill | What it does |
|---|---|
/video-insight:analyze-video |
Full pipeline — extracts all 13 signals, produces structured markdown report |
/video-insight:extract-colors |
K-means color clustering across keyframes with semantic token assignments |
/video-insight:write-copy |
All visible text via OCR + full narration transcript, organized by scene |
/video-insight:user-flow |
Step-by-step user journey from scene transitions, motion, and text changes |
/video-insight:describe-3d |
Tuned for 3D walkthroughs — camera movement classification per scene |
/video-insight:compare-videos |
Structural A/B diff — visual, text, flow, motion, similarity score |
| Skill | What it does |
|---|---|
/video-insight:build-from-video |
Pixel-perfect React or HTML from a screen recording |
/video-insight:generate-tests |
Infers user actions, generates Playwright or Cypress test file |
/video-insight:playwright-tests |
TypeScript Playwright tests with typed selectors and assertions |
/video-insight:generate-animations |
CSS @keyframes or Framer Motion from animated scenes |
/video-insight:storybook |
Storybook .stories.jsx for every detected UI component |
| Skill | What it does |
|---|---|
/video-insight:design-spec |
Figma-style spec — color tokens, component inventory, copy, motion, spacing |
/video-insight:export-tokens |
Colors as Tailwind config, CSS custom properties, or Figma tokens |
/video-insight:figma-tokens |
All three formats (Figma Variables, CSS, Tailwind) from one recording |
/video-insight:generate-prd |
Full PRD — overview, user stories, functional requirements, UI specs |
/video-insight:changelog |
User-facing changelog from before/after recordings |
| Skill | What it does |
|---|---|
/video-insight:watch |
Scan a directory, auto-analyze new videos, skip unchanged files via hash cache |
- Python 3.10+
uv— manages all ML dependencies automatically on first runffmpeg(includesffprobe) — required for video metadata extraction- Optional: GPU with CUDA support for faster transcription
Claude Code's biggest pain points: it hardcodes solutions for your specific example instead of solving generally, skips edge cases you didn't mention, and writes code before thinking. deep-engineer fixes all three with mechanical enforcement — not just instructions, but hooks and tools that physically prevent shortcuts.
User says "fix X"
│
▼
Phase 1: GENERALIZE ──────── Restate as general problem, list 5+ scenarios
│ save_task REJECTS fewer than 5 scenarios
│ PreToolUse hook BLOCKS Write/Edit
▼
Phase 2: HYPOTHETICAL TEST ── Walk through approach mentally against ALL scenarios
│ Document pass/fail table, revise until all pass
│ PreToolUse hook BLOCKS Write/Edit
▼
Phase 3: TDD ─────────────── Write tests covering ALL scenarios
│ PreToolUse hook ALLOWS test files only
│ BLOCKS implementation files
▼
Phase 4: IMPLEMENT ────────── Write code to pass all tests
│ Anti-hardcoding checklist enforced
│ All Write/Edit allowed
▼
Phase 5: VERIFY ──────────── Run tests, fix failures, loop until green
│ complete_task REJECTS unless all phases done
▼
Done ── Task archived to history/
This isn't prompt engineering. The plugin uses three enforcement layers:
| Layer | Mechanism | What it does |
|---|---|---|
| PreToolUse hook | Intercepts Write/Edit before execution | Physically blocks code files during Phases 1-2. During Phase 3, only allows test files. |
| MCP tool validation | Server-side rejection | save_task rejects <5 scenarios. update_phase rejects skipping/backwards. complete_task rejects until verify phase. |
| Stop hook | Injects additionalContext JSON |
Claude sees phase reminders and violation warnings in every turn. |
/deep-engineer:setup
/deep-engineer:solve Fix the date parser to handle ISO 8601 dates
Claude will:
- Restate as "parse date strings in multiple formats" (not just ISO 8601)
- List 5+ scenarios (with timezone, empty string, invalid format, unix timestamp, etc.)
- Walk through approach mentally against all scenarios
- Write tests covering all scenarios — implementation files are blocked by hooks
- Write implementation — no hardcoded values
- Run tests and loop until all pass
| Skill | What it does |
|---|---|
/deep-engineer:setup |
One-time setup — creates directory, wires CLAUDE.md @import |
/deep-engineer:solve |
Main engineering loop — all 5 phases with enforcement |
/deep-engineer:status |
Current phase, scenarios, warnings, test runner info |
Supports 20+ frameworks out of the box:
| Language | Frameworks |
|---|---|
| Python | pytest, unittest |
| JavaScript/TypeScript | vitest, jest, mocha |
| Go | go test |
| Rust | cargo test |
| Ruby | rspec, rake |
| Java | maven, gradle |
| PHP | phpunit |
| Elixir | mix test |
your-project/
├── CLAUDE.md <-- @deep-engineer/current-task.md added here
└── deep-engineer/
├── current-task.md <-- active task (auto-loaded every session)
├── log.jsonl <-- phase transitions + violation log
└── history/
└── task-2026-03-09T....md <-- archived completed tasks
- Python 3.10+
uvx(installed automatically withuv)
Estimates cost per session, warns at thresholds, and can hard-block tool calls when budget is exceeded.
| Skill | What it does |
|---|---|
/cost-guard:setup |
One-time project setup |
/cost-guard:budget |
Set or view spending budget |
/cost-guard:report |
Detailed cost breakdown |
Forces Claude to describe its approach in plain English and get user approval before writing any code. Write/Edit tools are physically blocked until approved.
| Skill | What it does |
|---|---|
/rubber-duck:setup |
One-time project setup |
/rubber-duck:explain |
Start or submit an explanation |
Automatically snapshots files before/after every Write/Edit. Roll back any number of actions instantly.
| Skill | What it does |
|---|---|
/rollback:setup |
One-time project setup |
/rollback:rollback |
Undo the last N actions |
/rollback:rollback-list |
Show recent tracked actions |
Runs tests automatically after every file edit. Detects regressions instantly and can auto-revert breaking changes.
| Skill | What it does |
|---|---|
/regression-sentinel:setup |
One-time project setup |
/regression-sentinel:configure |
Set test command and auto-revert |
/regression-sentinel:status |
View test pass/fail history |
Creates lightweight git stash snapshots before risky edits and on demand. Browse and restore any checkpoint.
| Skill | What it does |
|---|---|
/time-capsule:setup |
One-time project setup |
/time-capsule:checkpoint |
Create a named checkpoint |
Define exact files/directories Claude can touch. PreToolUse hooks block Read/Write/Edit/Bash/Glob/Grep outside the boundary.
| Skill | What it does |
|---|---|
/scope-lock:setup |
One-time project setup |
/scope-lock:lock |
Lock scope to specific paths |
/scope-lock:unlock |
Remove scope restrictions |
/scope-lock:status |
View current scope |
After every edit, generates a plain-English summary. Maintains a running changelog and produces PR-ready descriptions.
| Skill | What it does |
|---|---|
/diff-narrator:setup |
One-time project setup |
/diff-narrator:summary |
View recent change narrations |
/diff-narrator:pr-description |
Generate PR-ready description |
Checks vulnerabilities, license compatibility, and package info before installation. Blocks unapproved installs.
| Skill | What it does |
|---|---|
/dependency-doctor:setup |
One-time project setup |
/dependency-doctor:analyze |
Analyze a package |
/dependency-doctor:audit |
Audit all project dependencies |
/dependency-doctor:approve |
Approve a package for install |
Claude writes a block, pauses for human review, then continues. Tracks approval rate and enforces review cadence.
| Skill | What it does |
|---|---|
/pair-mode:setup |
One-time project setup |
/pair-mode:pair |
Start a pair session |
/pair-mode:stats |
View pair stats |
Analyzes your codebase and generates an onboarding guide: architecture overview, tech stack, directory structure, key files, and getting started.
| Skill | What it does |
|---|---|
/codebase-onboard:onboard |
Generate onboarding guide |
/codebase-onboard:find |
Search the guide |
/codebase-onboard:refresh |
Refresh the guide |
We welcome contributions! See CONTRIBUTING.md for development setup and guidelines.
- Fork the repo
- Create a feature branch (
git checkout -b feature/my-feature) - Make changes and add tests
- Open a PR
Please read our Code of Conduct before contributing.
To report a vulnerability, see SECURITY.md. Do not open a public issue for security bugs.
MIT - see LICENSE for details.
Use freely, attribution appreciated.
Built by Gowtham • Follow on X
