Philosophy: You are the architect; agents are the builders
Advanced Claude Code plugin that turns PRDs (product requirements documents) into autonomous development. Think of it as plan mode on steroids, powered by context and agent orchestration.
Claude Features
Worktree Isolation
Each task executes in its own git worktree. No conflicts, no race conditions.
Mar 2026
Sub-Agents
Spawn focused child agents for parallel task execution across your codebase.
Jul 2025
Agent Teams
Coordinate multiple agents working on related tasks simultaneously.
Sep 2025
Skills
Reusable capability modules that extend agent knowledge and behavior.
Jan 2026
Hooks
Lifecycle hooks for pre/post task automation and validation steps.
Jun 2025
Model Routing
Route tasks to optimal models based on complexity and cost constraints.
Jul 2025
Commands
Custom slash commands to streamline repetitive workflows and operations.
Mar 2025
Branch Assertion
PM Agent verifies branch state before and after each operation.
Sep 2025
Loop Detection
Automatic detection of revision loops prevents infinite cycles.
Sep 2025
Crash Recovery
Execution state reconstructed from git. Resume exactly where you left off.
Sep 2025
Worktree Isolation
Each task executes in its own git worktree. No conflicts, no race conditions.
Mar 2026
Sub-Agents
Spawn focused child agents for parallel task execution across your codebase.
Jul 2025
Agent Teams
Coordinate multiple agents working on related tasks simultaneously.
Sep 2025
Skills
Reusable capability modules that extend agent knowledge and behavior.
Jan 2026
Hooks
Lifecycle hooks for pre/post task automation and validation steps.
Jun 2025
Model Routing
Route tasks to optimal models based on complexity and cost constraints.
Jul 2025
Commands
Custom slash commands to streamline repetitive workflows and operations.
Mar 2025
Branch Assertion
PM Agent verifies branch state before and after each operation.
Sep 2025
Loop Detection
Automatic detection of revision loops prevents infinite cycles.
Sep 2025
Crash Recovery
Execution state reconstructed from git. Resume exactly where you left off.
Sep 2025
Worktree Isolation
Each task executes in its own git worktree. No conflicts, no race conditions.
Mar 2026
Sub-Agents
Spawn focused child agents for parallel task execution across your codebase.
Jul 2025
Agent Teams
Coordinate multiple agents working on related tasks simultaneously.
Sep 2025
Skills
Reusable capability modules that extend agent knowledge and behavior.
Jan 2026
Hooks
Lifecycle hooks for pre/post task automation and validation steps.
Jun 2025
Model Routing
Route tasks to optimal models based on complexity and cost constraints.
Jul 2025
Commands
Custom slash commands to streamline repetitive workflows and operations.
Mar 2025
Branch Assertion
PM Agent verifies branch state before and after each operation.
Sep 2025
Loop Detection
Automatic detection of revision loops prevents infinite cycles.
Sep 2025
Crash Recovery
Execution state reconstructed from git. Resume exactly where you left off.
Sep 2025
THE PROBLEM
Plan Mode Doesn't Scale
Plan mode works great for boilerplate code. The limiting factor is quadratic attention, and the workaround is harness engineering. With higher task complexity and duration, we need shared and optimized context for humans and agents.
SOLUTION
Compound Orchestration
You drive this loop. KARIMO auto-detects your project config, scans your codebase for patterns, then interviews you across 5 rounds to produce a structured PRD. You decide when the plan is ready to execute.
Auto-detect project settings and create .karimo/config.yaml — the single source of truth for all KARIMO operations.
Scan your codebase and the web for patterns, conventions, and context before planning.
A 5-round PRD interview that produces a structured plan with tasks, waves, and dependencies. Rounds are template-driven — edit the interview skill to reshape the process.
LIVE EXAMPLE
A Real Migration, Start to Finish
Recently, we migrated an entire Framer website to a custom Next.js codebase. This can never be done in one plan mode. We migrated all the content, images, compressed them, and built new web page structures.
ENCODING
Git Timeline
Main
Research
/karimo:research
Create PRD
/karimo:plan
Task Briefs
/karimo:run
Dependency Graph
/karimo:run
Research
/karimo:research
Create PRD
/karimo:plan
Task Briefs
/karimo:run
Dependency Graph
tasks.yaml
Planning Phase
You are the architect. Research discovers external dependencies, standards, and internal patterns. You can layer in your own research, upload screenshots, and collect assets in a PRD folder. A structured interview captures your requirements into a PRD. Task briefs and a dependency graph are generated automatically.
How It Works
- →Research scans external deps, API docs, standards + internal patterns, components, conventions — layer in your own research and assets
- →Structured PRD interview captures requirements (~10 min via /karimo:plan)
- →Task briefs generated from research + PRD artifacts
- →Dependency graph maps execution order into parallelizable waves
CONTEXT
Thoughtful Architecture
In plan mode, you get a single 1M-token session that is static, flat, and isolated. KARIMO expands that into progressive, compounding context through progressive disclosure, context multiplication, and compound learning. Agents only load what they need, every session builds on the last, and feedback persists across every future PRD.
Context Compression
Each stage distills context into artifacts optimized for the next session. A 1M token window becomes a building block across multiple sessions with shared information and objectives.
1M × tasks × research depth
Progressive Disclosure
Built on the OpenViking protocol — agents load abstracts first, overviews second, and full definitions only when executing. No wasted tokens.
One-line summaries per agent, command, and skill. Used to verify items exist before loading more. Near-zero cost.
Context Compression
Each stage distills context into artifacts optimized for the next session. A 1M token window becomes a building block across multiple sessions with shared information and objectives.
Scans your codebase and the web. Produces findings.md — patterns, conventions, and external context that agents would otherwise have to rediscover.
Your answers + research findings become a structured PRD with tasks, waves, and dependencies. Context is distilled, not duplicated.
Each task gets a self-contained brief with only what it needs — scope, constraints, and findings relevant to that task. No cross-contamination.
Each brief spawns a fresh agent in its own worktree with a clean 1M window. Tasks can run in parallel or dependent, pulling relevant context when needed.
Scans your codebase and the web. Produces findings.md — patterns, conventions, and external context that agents would otherwise have to rediscover.
Compound Learning
At any point, run /karimo:feedback to capture issues or potential improvements. Observations move through a capture stage, then get stored as summarized learnings in the KARIMO learnings folder — patterns that work, anti-patterns to avoid, execution rules, and product-specific notes. These compound over time: every future PRD and task brief loads in relevant learnings for that specific task. Agents never repeat the same mistake twice.
Open source — contribute on GitHub to help improve KARIMO as models evolve.
ADOPTION
Three Simple Steps
Set up your environment so AI can quickly work as a cheat sheet. Run your first research — 8 documents generated for internal and external resources. Then kick off an interview to create a PRD that gets decomposed into tasks.
Feel the difference between prompting and orchestrating
Frequently Asked Questions
LET'S BEGIN...
Get started today
Clone the repo and run the install script from your terminal.
Paste this prompt into Claude Code and it will handle the rest.
Works in Claude Code CLI, desktop app, or IDE extensions.