Kaddo — Knowledge Driven Development

Prepare any codebase for AI-assisted evolution. Kaddo helps your repo remember why the code exists.

Kaddo is an open-source Knowledge Driven Development (KDD) toolkit for AI-assisted software engineering. It scans your repo, structures product knowledge into four layers (Business, Product, Tech, Delivery), manages the full Work Item lifecycle — from captured intent through implementation handoff, evidence collection and verification — and keeps knowledge alive as the system evolves.

It works in two layers:

• The CLI does the deterministic work — no AI, no API key.
• Your LLM does the interpretation — using Kaddo agents in your chat (Claude, ChatGPT, Cursor, Copilot, Windsurf…).

Knowledge Driven Development ≠ Kaddo. KDD is a prior concept in software engineering and knowledge management. Kaddo is a practical implementation of KDD principles — it applies them; it did not invent them.

Quick Start

npx @kaddo/cli init       # configure project state, structure & language (en/es)
kaddo bootstrap            # new projects: seed Business → Product → Tech → Delivery knowledge
kaddo scan                 # deterministic technical inventory → .kaddo/scan.json
kaddo add agents           # install agent prompt packs
kaddo context              # LLM context pack → .kaddo/context-pack.md
kaddo understand           # guided CLI → LLM handoff plan

Then use your LLM with the context pack and agents to extract capabilities, architecture and a roadmap. Continue the delivery loop:

kaddo create --from roadmap   # roadmap candidate → Work Item draft
kaddo verify                  # evidence collection + AC verification
kaddo owners suggest          # declare code: ownership
kaddo guard                   # detect possible knowledge drift
kaddo explain                 # summarize what Kaddo currently knows

Lost? Run kaddo understand — it always answers "What should I do now?" from the real state of your project.

Proof of Concept mode

Use POC mode when the immediate question is whether an assumption is true, rather than how to plan a complete product increment. It is independent from project state and keeps the work focused on a hypothesis, observable success criteria, minimal technical context, experiment evidence, and a conclusion.

kaddo init --mode poc
# or change an existing project
kaddo project mode poc
kaddo add agents
kaddo add skills
# define the hypothesis and success criteria in knowledge/delivery/poc.md
kaddo context
kaddo understand
kaddo create spike

The canonical record is knowledge/delivery/poc.md. It concludes as validated, rejected, or inconclusive; Work Items retain their normal lifecycle, with a spike often being the right first experiment. Token savings can result from this focused context, but the goal is better evidence and decisions, not prompt optimization.

POC mode does not require the standard Business/Product knowledge baseline. Add technical context or Project Resources only when they are relevant to the experiment.

Install the agent prompts and reusable skills as part of the initial POC setup. They provide the guided refinement and implementation context; they do not add a Business/Product baseline.

After recording validated, rejected, or inconclusive, run kaddo poc report to prepare a deterministic, source-attributed final-report handoff. The report is optional; a reviewed proposal is persisted only with confirmation as immutable knowledge/delivery/poc-report-v001.md (later versions never overwrite earlier ones). It is missing, current, or stale based on a source fingerprint; new evidence never regenerates it automatically.

The Knowledge Model

Kaddo organizes project knowledge into four layers under knowledge/:

Each layer feeds the next. Agents don't guess — they read structured context.

The Kaddo Lifecycle

Work Items follow the Kaddo-native Build Contract — a lifecycle that ensures every change is traceable from intent to verified delivery:

Captured Intent → Refinement → Human Review → Ready
    → Implementation Handoff → Implementation
    → Evidence Collection → Verification → Completed → Learning

Work Items physically move between directories: draft/ → ready/ → in-progress/ → completed/. The lifecycle is enforced by convention — no lock-in, no platform dependency.

Interfaces

Kaddo exposes its capabilities through four interfaces:

The MCP server is mostly read-only. Lifecycle tools (import, ready, handoff, evidence, verify) can write under knowledge/delivery/work-items/; derived tools regenerate artifacts under .kaddo/. Everything else is strictly read-only.

{ "mcpServers": { "kaddo": { "command": "npx", "args": ["@kaddo/mcp"], "cwd": "/path/to/project" } } }

See the MCP Server docs.

Integrations

Kaddo connects to external work systems through the Integration Adapter Foundation:

External system (GitHub Issues, Jira, Azure DevOps…)
    → discovery: browse & filter external items
    → import: bring an item into Kaddo as a Work Item draft
    → refinement: enrich with ACs, scope, affected modules

Import is always human-confirmed — Kaddo never creates Work Items silently. Secrets are managed through environment variables and never exposed. Configure integrations through the Admin UI or @kaddo/integrations programmatically.

Knowledge Capsules

For repos you can't (or don't want to) map as multirepo modules — other teams, restricted access, integration-only context — exchange Knowledge Capsules:

kaddo capsule export   # share a minimal summary of this project
kaddo capsule add      # import an external capsule into the context pack

No multirepo mapping needed, no source access required. See the Knowledge Capsules docs.

Operating Moments

Kaddo matures a project's knowledge through four moments — Base → Definition → Projection → Execution:

• Base — init · bootstrap · scan · add agents · context · understand (set up the workspace and knowledge structure).
• Definition — business-agent · product-agent · capability-agent · codebase-agent · architecture-agent (turn the idea into clear knowledge).
• Projection — roadmap-agent · backlog-agent · create --from roadmap · work-item-agent · ownership-agent (turn knowledge into a delivery plan).
• Execution — implementation-agent · verify · scan · owners suggest · guard · explain (build, verify and keep knowledge in sync).

See Operating Moments for the full breakdown.

Commands

Each command answers one question and has a clear next step.

Supporting commands: kaddo status, kaddo learn, kaddo classify, kaddo history, kaddo module, kaddo modules map|list, kaddo capsule export|add, kaddo graph export, kaddo report impact, kaddo savings, kaddo guard --record, kaddo drift, kaddo questions, kaddo ignore, kaddo add <module>, kaddo work-item import, kaddo adapters install|list|status.

Multirepo Modules & Global Artifacts

Map secondary repositories as living modules of one system:

kaddo modules map    # register a secondary repo as a module
kaddo modules list   # list mapped modules

Per-module knowledge (module-design, stack, security, standards) is generated under knowledge/tech/modules/<id>/. Global artifacts cover cross-cutting concerns:

kaddo add standards    # knowledge/tech/standards.md
kaddo add security     # knowledge/tech/security.md
kaddo add stack        # knowledge/tech/stack.md
kaddo add git-strategy # knowledge/tech/git-strategy.md + .kaddo/git.yml

Skills

Skills standardize how agents do common things well (writing an ADR, refining a Work Item, planning implementation, collecting evidence). Agents orchestrate; skills standardize. Install with:

kaddo add skills                  # recommended (delivery + tech)
kaddo add skills --all            # every skill
kaddo add skills --group tech     # one group: delivery | tech | integration

Adapters (AGENTS.md / CLAUDE.md)

Generate native instruction files so AI coding tools work inside a Kaddo repo:

kaddo adapters install codex        # AGENTS.md for Codex
kaddo adapters install claude       # CLAUDE.md for Claude Code
kaddo adapters install kiro         # AGENTS.md for Kiro
kaddo adapters install opencode     # AGENTS.md for OpenCode
kaddo adapters install antigravity  # AGENTS.md for Antigravity

Both are compact projections of the project's knowledge map — Kaddo stays the source of truth. See Custom Adapters for the shared Adapter Contract.

Kaddo Power

The official kaddo-power/ packages Kaddo Skills, KDD workflow guidance and @kaddo/mcp integration using the Agent Plugins specification. Kiro Powers is the first supported consumer.

Templates

Kaddo ships templates for its main artifacts — organized into six categories: core, business, architecture, module, operations and legacy. Each carries a purpose, when-to-use, output path and quality checklist. See the Templates docs.

Examples

The examples/ folder has reproducible demo repositories:

Each includes a prompt-flow.md with a Mermaid diagram, CLI↔LLM split and copy/paste prompt handoffs. See the Examples docs.

Self-hosting

Kaddo builds Kaddo with Kaddo. The project uses its own lifecycle to manage Work Items, collect evidence, verify ACs and capture learnings — the same workflow it provides to any project. The knowledge/ directory contains Kaddo's own product knowledge; .kaddo/ holds its derived artifacts. This self-hosting validates the lifecycle end-to-end and ensures that friction encountered internally feeds back into improvements.

How Ownership and Guard Work

Ownership is declared in artifact front matter — no central mapping file:

---
type: feature
id: WI-001
title: "Add payment retry logic"
status: in-progress
code:
  - src/payments/**
  - src/shared/payment/**
---

kaddo guard reads git diff, finds artifacts whose code: globs match changed files, and shows a non-blocking FYI when the artifact was not updated in the same diff. Guard is silent when no artifacts declare ownership — no noise on day one.

What Kaddo Does Not Do

• Not a code generator
• Not an agent execution framework (it ships agent prompts, it does not run them)
• Not a replacement for Jira, Linear or documentation tools
• Not a platform
• Does not call an LLM or require an API key

About the Author

Kaddo is created and maintained by Julian Dario Luna Patiño — Cloud Solutions Architect Lead, AWS Community Builder and content creator at TryCatch.tv. It is the result of years designing software architectures, leading development teams and documenting systems. Kaddo applies Knowledge Driven Development principles to AI-assisted software development; it does not claim to have invented KDD. GitHub · LinkedIn · About.

Contributing

See CONTRIBUTING.md.

License

MIT