CRAFT — Context-Reusable Agentic Folder Tree

An open standard for organizing projects where humans and AI ship together.

AI output is bounded by input quality, so context deserves the same discipline as code. CRAFT splits slow-moving human signal (vision, domain, references) from fast-moving AI output (plans, builds) and keeps each at a stable, predictable path.

your-project/
├── AGENTS.md       # entry point for agents — directory map + conventions
├── specs/          # what & why — vision, domain, scope          (human)
├── references/     # supporting inputs — docs, data, feeds       (human)
├── prompts/        # reusable patterns — skills, jobs, QA        (human)
├── plans/          # plans, reports, drafts (TLDR-first)         (AI)
└── build/          # shippable code repos                        (AI)

Use this repo as a scaffold — top-level directories ship empty, each with a short README explaining what goes there:

git clone <this-repo> my-project
cd my-project && rm -rf .git && git init

For a fully-populated worked example, see references/example-projects/lighthouse/ — a fictional invoice-management SaaS that fills in every directory.

Still iterating — feedback welcome via GH issues.

These are markdown files that provide a high degree of human signal in the what and why of the project.

Example files for software product projects:

• specs/vision.md: The business and product vision of the project.
• specs/domain.md: Important domain knowledge, terminlogy and industry context.
• specs/product/mvp-scope.md: A definition of the Minimum Viable Product.
• specs/marketing/value-proposition.md: An outline of the marketing message

Note: We can imagine certain tooling that creates a write-only file (e.g. a Google Doc vision doc that outputs to specs/vision.md)

These can be a combination of static or dynamic artifacts. They can be documents, data sets or feeds from external systems.

Example files for Rekall organization:

• references/datasets/example-customer-export/: Data sets that can be relevant (e.g. for data model planning)
• references/slack/<channel>/: Append-only feed/dump of Slack channel.
• references/linear/<board>/: Append-only feed/dump of Linear tickets.
• references/google-docs/<file>.md: Markdown representation of Google Docs.
• references/google-meets/<meeting-date>.md: Transcripts from relevant Google Meets.
• references/figma/<project>/: Figma export.
• references/claude-design/branding/: Example export of a design project.
• references/claude-design/marketing-website/: Example export of a design project.

This folder could hold generic prompting templates, but would have specific folders for skills and jobs.

• prompts/jobs/: Schedulable AI prompts with triggers.
• prompts/skills/: Reusable agent capabilities.
• prompts/checklists/: Checklists for agents (e.g. project-kickoff-requirements.md)
• prompts/qa/: Quality assurance routines (e.g. mvp-golden-path-acceptance-test.md).

This is the place for agentic planning (with skills).

Examples:

• plans/engineering/mvp-data-model.md: Data model used downstream.
• plans/marketing-website/implementation-plan.md: Full plan in md, TLDR at the top.
• plans/marketing-website/implementation-plan.html: Visualized plan in rich HTML.
• plans/style.css: Common styles for HTML plans

Agenticly produced artifacts. Examples:

• Apps
• Services
• Infrastructure
• Libraries
• Plugins