AutoFlow Standard

Documentation-as-code for AI-driven projects. An 11-file convention + autonomous daemon + cross-agent CLI that keeps your Obsidian vault, AI agents (Claude/Codex/Hermes), and Discord in sync — automatically.

What it is

A standardized 11-file structure for every project under 02-Projects/<Name>/:

02-Projects/<Name>/
├── 📝 MANUAL (you write once, rarely change)
│   ├── CLAUDE.md           # routing card for AI agents
│   ├── VISION.md           # why + Out-of-Scope
│   ├── CONSTITUTION.md     # rules + Why per rule
│   └── ONBOARDING.md       # setup + credentials refs
│
├── 🤖 AUTO-GENERATED (daemon refreshes, do not edit)
│   ├── SYSTEM.md           # architecture (Arch-Hub linked)
│   ├── RUNBOOK.md          # incident response (buglog-aware)
│   └── PROCESS-MAP.md      # n8n / automation map
│
├── 📊 SESSION-DRIVEN (hook updates)
│   ├── STATUS.md           # current state
│   └── handoff.md          # next-session context
│
└── 📚 APPEND-ONLY (never deleted)
    ├── decisions.md        # why X was chosen
    └── mistakes.md         # what failed and why

Why bother

Problem	This solves
AI agents load 50KB of project docs every query → expensive	11 files = load only what relevant. 5× cheaper context.
Multiple agents editing one mega-doc → merge hell	11 files = 4 writers concurrently, 0 conflicts.
Manual + auto docs in same file → daemon overwrites human edits	Separated. AUTO-START/END markers protect manual sections.
Decisions get rewritten or lost	Append-only with INSERT markers. Never deleted.
New employee onboarding takes weeks of "where is...?"	ONBOARDING.md + CLAUDE.md routing → 30 min to productive.
Documentation goes stale	Daemon refreshes auto-sections every 5 min. PostToolUse hook triggers on edit.
Different agents (Claude/Codex/Hermes) need same context	Unified vault CLI + per-project CLAUDE.md routing card.

Features

• ✅ 11-file convention with manual/auto/session/append-only separation
• ✅ vault CLI (Bash, stdlib-only Python) — agent-agnostic
• ✅ 3 auto-refresh skills (Python): runbook, process-map, anatomy
• ✅ macOS LaunchAgent daemon — auto-refreshes every 5 min
• ✅ PostToolUse hook — refreshes immediately on file edit
• ✅ Claude Code integration — 7 slash commands (/decision, /mistake, /note, /where-am-i, /refresh, /verify, /handoff)
• ✅ Hermes Discord integration — bot skill for remote !decision, !mistake, !note
• ✅ Obsidian Templater scaffolder — 7 questions → 11 files
• ✅ Obsidian Dataview dashboard — compliance check across all projects
• ✅ Token leak detection — heuristic secrets scan in vault verify
• ✅ Idempotent — runs twice = no double writes
• ✅ Hebrew + English — UI Hebrew, code English

Quick install

# 1. Clone
git clone https://github.com/YOUR_USERNAME/autoflow-standard.git
cd autoflow-standard

# 2. Set vault location
export AUTOFLOW_VAULT="$HOME/vault"

# 3. Run installer
./bin/install.sh

# 4. Verify
vault help
vault verify

5-minute walkthrough

Scaffold a new project

# Option A: bash
mkdir -p "$AUTOFLOW_VAULT/02-Projects/MyProject"
cd "$AUTOFLOW_VAULT/02-Projects/MyProject"
# (manually copy templates/_ProjectStandard/*.md and fill placeholders)

# Option B: Obsidian Templater (interactive — recommended)
# Cmd+P → "Templater: Create new note from template" → "Project-Init"
# Answer 7 questions → 11 files created

Log a decision

cd "$AUTOFLOW_VAULT/02-Projects/MyProject"
vault decision "Switched from REST to GraphQL — needed batched queries"
# Edits 02-Projects/MyProject/decisions.md (append-only, datestamped)

Check compliance

vault verify              # all projects
vault verify MyProject    # just one + secrets scan

From Claude Code

/decision Switched from REST to GraphQL
/mistake Tried polling instead of webhooks — too slow
/note Idea: weekly digest workflow
/where-am-i

From Discord (via Hermes bot)

!decision Switched stack to Kubernetes
!note Bug idea — race condition in cache invalidation
!verify

Auto-refresh

• LaunchAgent runs vault-daemon every 5 min
• PostToolUse hook triggers refresh on every Edit/Write inside 02-Projects/
• Idempotent — runs only if source files changed

Architecture

                 [files on disk: 02-Projects/<Name>/...]
                          ↑ ↓
                  [vault CLI: bash + python]
              ↑         ↑         ↑         ↑
        Claude Code   Codex    Hermes    macOS LaunchAgent
        (slash cmds  (AGENTS  (Discord    (daemon every
        + hooks)     .md)     skill)      5 min)

Comparison to alternatives

Approach	Pros	Cons
Single big README	Easy to read	Context expensive, drift, no autonomy
OpenWolf (cerebrum/anatomy/memory)	Daemon model	In practice 95% empty, single-codebase mindset
Notion/Linear/Confluence	Polished UI	Vendor lock-in, not Markdown, no AI context
AutoFlow Standard	Markdown, autonomous, multi-agent, cheap context	Setup effort: 30 min

See docs/comparison.md for detailed comparison.

Documentation

• Architecture: why 11 files?
• Getting started
• Comparison to OpenWolf and single-file approaches
• vault CLI reference
• Daemon & autonomy
• Claude Code integration
• Hermes Discord integration
• Obsidian setup (Templater + Dataview)

Status

Production. In use on 7+ active projects (WhatsApp automation, e-commerce pipelines, CRM, voice bots, lead gen).

License

MIT — see LICENSE.

Credits

Created by Daniel Moshe (AutoFlow). Inspired by the gaps in OpenWolf, Notion templates, and single-file READMEs.

For AI agents working in this repo

See AGENTS.md — canonical instructions for Claude Code, Codex, Cursor, Aider, and any AI agent loading this repository.

Reference example

docs/examples/example-project/ — a fully filled-out fictional project ("TaskTracker") showing what all 11 files look like when populated. Use this to learn the patterns before scaffolding your own.