A minimal, tool-agnostic architecture for doing serious work with AI. Instead of losing context every session, you give the AI a project to join: documentation it reads before it speaks, rules that get tighter over time, and a shared memory that survives tool switches.
- Download this folder. Click the green Code button above → Download ZIP. Unzip it and rename it to your project.
- Open it in any AI tool that reads files: Claude Cowork, Claude Code, Cursor, Codex, Antigravity, or similar.
- Say: "Read SETUP.md and help me get started." The AI needs you to ask — it will not start on its own.
The AI will run a short interview (4 rounds of questions, ~5 minutes) and populate your project files automatically. When it is done, you are ready to work. You can start a task immediately or close the session and come back later — everything is saved in the files, and any AI tool you open this folder with will pick up where you left off.
If you use GitHub: click Use this template (top right) to create your own copy. Then clone and follow step 2.
If you already know what you want, skip the interview and edit the files directly:
CLAUDE.md— Operating instructions. Who you are, what the project is, how work gets done.STYLE.md— Voice and preference rules. Starts nearly empty, grows after every session.INDEX.md— Reading order. Controls what the AI loads and in what sequence (progressive disclosure).TASKS.md— Active tasks, backlog, and done log.SESSION-REVIEW.md— Instructions the AI follows at the end of every working block to update the memory.
See the examples/ folder for what populated files look like in a mature project. These examples are from a specific author (a climate science writer). Your files will look different depending on your project and voice. Use the examples as a pattern, not a checklist.
├── SETUP.md ← Entry point. AI reads this and interviews you.
├── CLAUDE.md ← Operating instructions for the AI
├── STYLE.md ← Voice rules (grows over time)
├── INDEX.md ← Reading order (progressive disclosure)
├── TASKS.md ← Task tracking
├── SESSION-REVIEW.md ← End-of-session review protocol
│
├── knowledge-base/
│ ├── memory/ ← Decisions, strategies, accumulated context
│ ├── docs/ ← Structured notes and reference material
│ └── sources/ ← Raw inputs: reading notes, exports, transcripts
│
├── projects/
│ └── my-project/ ← Drafts, assets, working files
│
├── workflows/ ← Guided processes for recurring tasks
│ ├── organize-material/ ← Sort existing notes into the knowledge base
│ ├── build-research-db/ ← Structure reading notes into a searchable database
│ ├── draft-from-seed/ ← Develop an idea into a first draft
│ ├── revision-session/ ← Review a draft against your style guide
│ └── voice-capture/ ← Process a dictated idea into the knowledge base
│
├── examples/ ← What a mature setup looks like
│ ├── CLAUDE-example.md
│ ├── STYLE-example.md
│ └── SESSION-REVIEW-example.md
│
└── skills/ ← Optional: reusable AI commands (advanced)
└── session-review/
You do not need to understand these to start. They will become obvious as you use the system.
Progressive disclosure. The AI reads only what it needs for the current task. INDEX.md controls the reading order. Less context produces better output.
Session reviews. At the end of every working block, you and the AI update the memory files together. Mistakes become permanent rules. The system gets smarter after every session.
The writer as filter. AI generates volume. Your job is to decide what stays, what gets cut, and what needs to be rewritten. The work is what survives your judgment.
After setup, the template includes guided workflows for common tasks. Say "Read workflows/draft-from-seed/WORKFLOW.md and help me" to start one. Or say "What workflows are available?" to see the full list.
Workflows cover five common tasks: organizing existing material, building a research database, drafting from a seed idea, running a revision session, and processing voice captures. Each one is a structured instruction file the AI follows step by step. See workflows/README.md for details.
The skills/ folder contains reusable AI instructions that some tools (Claude Code, Cowork) can execute as commands. Currently includes a session-review skill that automates the end-of-session protocol. You can create additional skills for workflows you run frequently (drafting, revision, voice capture). See skills/README.md for setup. If you are just getting started, skip this entirely.
The system needs periodic pruning. Stale context is worse than no context. Every month or so, spend a session reviewing the knowledge base: archive completed projects, delete outdated reference material, consolidate overlapping notes, and remove completed tasks from the Done log. This keeps the memory sharp and prevents the AI from reading information that no longer applies.
The intelligence lives in the files, not in the client. Switch tools and the new one picks up where the last one left off. No special setup is needed when you change AI tools — the new tool reads the same files and follows the same instructions.
This architecture comes from a writing practice described in "The Agentic Studio". You do not need to read the essay to use the template, but it explains the philosophy behind the design. More writing at jonadas.com/writing.