project-brain

Different session. Same brain.

A folder structure + collaboration protocol that lets your AI assistant pick up a project after a context wipe — across new sessions, new windows, new collaborators.

中文版 →

Part of Sprout Labs — Local-first AI memory & agent safety, built independently by Ethan (Independent Product Designer & AI Builder, since 2016). ethanflow.com · LinkedIn · @ethanflow_lab · GitHub

The problem

Long-running projects with AI coding assistants (Claude Code, Cursor, Copilot, etc.) hit a counter-intuitive wall:

Larger context windows don't solve the problem. Better information structure does.

A wider window doesn't mean it gets read. Read doesn't mean located. Located doesn't mean prioritized. Without structure, every new session starts with "wait, what's going on here?" — and the AI either reads too much (wasting tokens) or misses the critical pieces.

Symptoms you'll recognize:

• One README mixing project pitch + current status + decision history + how-to-run
• Sprawling docs where boundaries blur (one architecture file containing design + ops + history + bugs)
• Decisions buried in commit messages, chat logs, and footnotes — never traceable when you need them
• Cross-window handoff loses the "fresh stuff in your head" — the next session starts blind
• Projects running multiple parallel workstreams (e.g., dev + ops + outreach as three parallel streams) lose track of which session is on which line — switching between them re-orients from scratch every time

project-brain is a structural answer: a brain/ folder layout + a small set of protocols, designed so a fresh AI session can read 2-3 files and be productive.

Quick start

Option A — install as a Claude Code plugin (recommended for Claude Code users)

In Claude Code, run these two slash commands:

/plugin marketplace add Ethan-YS/project-brain
/plugin install project-brain@sprout-labs

If /plugin isn't exposed in your Claude Code environment (some embedded / SDK contexts strip it), the CLI form does the exact same thing — run from any terminal:

claude plugin marketplace add Ethan-YS/project-brain
claude plugin install project-brain@sprout-labs

That's it. From any project after that, just say:

• "set up project brain" — kick off a new brain
• "resume this project" — load MAP.md + STATUS.md + (if exists) HANDOFF.md
• "I'm switching windows" / "context's getting full" — write a HANDOFF before context dies
• "update the project brain" — propose updates with reasons, you approve per item

The skill handles four workflows: new-project kick-off, startup resume, window-switch handoff, and updates. Auto-trigger requires an explicit user request — it intentionally does not activate just because a brain/ folder exists.

Already installed the old way? If you previously cloned to ~/.claude/skills/project-brain, remove it first: rm -rf ~/.claude/skills/project-brain. The plugin install is the supported path going forward.

Update later: /plugin marketplace update sprout-labs to fetch the latest version.

Option B — manual scaffold (works with any AI assistant)

If you don't use Claude Code, or you just want the scaffold script:

git clone https://github.com/Ethan-YS/project-brain.git

# Scaffold into your project (defaults to all four AI adapters, English)
./project-brain/scripts/scaffold.sh /path/to/your/project

# Or pick specific adapters:
./project-brain/scripts/scaffold.sh /path/to/your/project --tools claude,cursor

# For projects documented in Chinese — gives Chinese brain/ + CLAUDE.md
# (other adapters stay English — they're consumed by AI tools, not humans):
./project-brain/scripts/scaffold.sh /path/to/your/project --lang zh

# Fill in brain/PROJECT.md on day one
# Walk through ⚠️ TODO ⚠️ placeholders

What the scaffold gives you:

When a new AI session opens your project, the auto-loaded instruction file directs it to read brain/MAP.md + brain/STATUS.md. If brain/HANDOFF.md exists, it picks up the previous session's "still-warm-not-yet-written-down" thoughts.

Health check (any time after scaffolding)

./project-brain/scripts/doctor.sh /path/to/your/project

Read-only structural checks: missing core files, STATUS over 80 lines, decisions without "rejected alternatives," topics/ files not registered in MAP, stale ⚠️ TODO ⚠️ placeholders, etc. Reports issues but never fixes them — that's still your call. See scripts/doctor.sh.

See a fully-filled example

If the empty templates feel abstract, see examples/small-saas/ — a fictional SaaS project ("Quill," a local-first notes app at v0.3) with every file filled in. Reading it is the fastest way to understand what each brain/ file looks like in real use.

Structure

brain/
├── PROJECT.md       What is this project? What do we explicitly NOT do?
├── MAP.md           Module structure + document index ("where do I find X?")
├── STATUS.md        Current state — overwritable, soft cap 80 lines
├── DECISIONS.md     Decision log — append-only, requires "rejected alternatives"
├── HANDOFF.md       Cross-window bridge — what's still in head but not yet written
├── handoffs/        Archive of past HANDOFFs (timestamped)
└── topics/          Per-problem-dimension docs
    ├── systems/      → "How is this designed?"
    ├── operations/   → "How do I operate it / what to do each release?"
    ├── planning/     → "What are we going to build / how to plan it?"
    └── feedback/     → "What is reality / users telling us?"

Core principles

1. Time-scale separation

The five core files in brain/ are split by how often they change, not by topic:

Why volatility-based? Because mixing low-frequency content with high-frequency content forces the low-frequency content to be re-edited every session — the most common mode of doc rot.

2. Problem-dimension classification (not by module)

brain/topics/ is divided into 4 dimensions: systems / operations / planning / feedback. A single business module (say, "payment") has its design in systems/, deploy logs in operations/, pricing strategy in planning/, user feedback in feedback/.

Why? Modules grow, shrink, get renamed, get merged. The four problem dimensions are far more stable.

3. Decision logging requires "rejected alternatives"

Every entry in DECISIONS.md must include what was considered and rejected, and why. Without this, the file degrades into a worse version of CHANGELOG.md (which already records "what was done"). The unique value of decision logs is the paths not taken — that's where a project's judgment lives.

4. Judgment Division

When the user says "update the project brain":

• The user decides "should we record now" (high-level pacing instinct)
• The AI decides "specifically what to record / how to write each entry" (specialized understanding of how each file works)
• The user approves or rejects the AI's judgment

The AI should not push specialized judgments back ("which files do you want me to update?") — that hands the wrong layer of judgment to the wrong person.

Since v2.6, routine bookkeeping writes are tiered (METHODOLOGY §4.1): STATUS / HANDOFF / mechanical MAP registrations are write-then-announce — the AI writes after the work lands and reports it, with git as the review surface. DECISIONS / PROJECT remain ask-before-write. The explicit "update the project brain" checkpoint above keeps per-item approval either way.

5. Multi-workstream mode (v2.1, optional)

Some projects naturally have parallel independent workstreams (e.g., development + operations + outreach in one product). For these, STATUS and HANDOFF split per workstream:

brain/
├── PROJECT.md           ← shared
├── MAP.md               ← shared
├── DECISIONS.md         ← shared
├── STATUS_dev.md        ← split per workstream
├── STATUS_ops.md
├── HANDOFF_dev.md
├── HANDOFF_ops.md
├── handoffs/
│   ├── dev/
│   └── ops/
└── topics/              ← shared

Single-workstream projects ignore this — they keep the default STATUS.md / HANDOFF.md.

Why this exists — the evolution story

This methodology didn't appear from a design session. It evolved through real use across multiple projects — each iteration triggered by friction we hit and couldn't ignore.

v1 (April 2026): After accumulating 15+ scattered docs and one 32KB monolithic file in a single project, every new AI session needed to re-read everything to figure out "what's going on here?" We refactored across 4 commits and established the basic structure: meta/ (continuity layer) + docs/ (problem-dimension classified). The core insight was separating files by volatility — stable / slowly-changing / per-session / append-only — to prevent the most common doc-rot pattern: low-frequency content getting re-edited because it lives next to high-frequency content.

v2 (April 30, 2026): We surveyed community approaches (Prompt Shelf 2026's three-file architecture, softaworks/agent-toolkit's session-handoff skill, Anthropic's official skills repo). We found:

• Community solutions were stronger at runtime mechanics (auto-handoff triggers, staleness detection, handoff chains)
• Our v1 was structurally deeper in three places: the "what we explicitly DON'T do" forcing function in PROJECT.md, mandatory "rejected alternatives" in decision log, and classification by problem-dimension instead of by module

The plan was "borrow the runtime layer, keep the structural layer." But across 8 rounds of refinement, the user systematically rejected mechanism creep. Every time the AI proposed "let me design auto-detection / triggering / staleness checking," the user said "no — that's a judgment I'll make myself."

This crystallized into the Judgment Division Principle (see Core principle 4): the user decides "should we record now"; the AI decides "what specifically to record"; the user reviews. Don't hide specialized judgment behind automation.

Companion design changes: merged meta/ + docs/ into a single brain/, added HANDOFF.md for cross-window continuity, replaced "hard keyword detection" with gentle inquiry ("does this count as decided?" instead of asserting "we decided X"), made placeholders visually loud (⚠️ TODO ⚠️), and made the git prerequisite explicit.

v2.1 (same day, evening): A second project — non-development, running parallel workstreams (official ops + outreach) — broke v2's hidden assumption that "one project = one workstream." The user had naturally evolved a workaround using STATUS_<workstream>.md naming. We folded it into the methodology as Multi-workstream Mode (Core principle 5): project-level files stay shared, status and handoff split per workstream.

The meta-takeaway

The healthy evolution pattern wasn't "design everything upfront." It was "build what works now, let real-world friction drive the next version." Each version came from a specific, named problem. Each refinement was triggered by a real moment of pain.

If you take one thing from this repo: resist the urge to design comprehensive automation up front. Real friction tells you which mechanisms are worth building, and which would just hand specialized judgment to the wrong layer.

Documentation

• METHODOLOGY.md — full methodology including all 16 traps, judgment division mechanics, multi-workstream details, and migration paths
• CHANGELOG.md — version history
• skills/project-brain/SKILL.md — Claude Code skill manifest (loaded automatically once the plugin is installed)
• .claude-plugin/ — plugin.json (manifest) + marketplace.json (Sprout Labs marketplace entry)
• templates/ — drop-in templates for brain/ + 4 AI tool adapters
• examples/small-saas/ — a fully-filled example brain/ folder
• scripts/scaffold.sh — one-command scaffold into any project
• scripts/doctor.sh — read-only health check (8 structural checks covering the most common traps)

Compatibility

This methodology is AI-agnostic. The repo includes adapter templates for:

• Claude Code — install as a plugin (Option A above) for the full skill experience, or just drop in CLAUDE.md for project-level instructions
• Cursor — .cursorrules
• GitHub Copilot Chat — .github/copilot-instructions.md
• Codex CLI / Aider / Continue — AGENTS.md (the agents.md convention)
• Any other AI assistant that respects a project-level instruction file (write your own pointer to brain/MAP.md + brain/STATUS.md)

./scripts/scaffold.sh copies all four by default; use --tools claude,cursor to pick specific ones.

Requirements:

• Git — several mechanisms (HANDOFF archival, decision traceability, file blame) assume git history
• That's it.

Status

🌱 v2.4 — methodology stable, 140+ stars, multiple projects in active use. v2.4 ships as a one-command Claude Code plugin (the Sprout Labs marketplace). v2.3 added scripts/doctor.sh (structural health check) and a fully-filled example project (examples/small-saas/). v2.2 introduced the Claude Code skill manifest and adapter templates for Cursor / Copilot / AGENTS.md. External effectiveness data still pending — battle-tested by the maintainers daily; field reports from other users still being collected.

Authors

Built by Ethan, the person behind Sprout Labs. The methodology came out of a Jarvis-style AI collaboration workflow: eight rounds of iteration in a single working day, each triggered by a specific friction point — not a design session.

The methodology itself models the working approach: clear judgment division, refusal to automate away decisions that should be made by humans, and willingness to admit when v1 needs to become v2.

Released under Sprout Labs.

License

MIT — use it, modify it, share it. If it helps your workflow, a star on the repo is appreciated.

Contributing

Issues and PRs welcome. The methodology is intentionally minimalist — proposals to add mechanisms should clearly identify the specific friction they solve. New traps (mistakes you've hit using this) are especially welcome.