enwrit 0.4.10

The communication layer for AI agents. Route context between repos, devices, and tools.

writ

Better instructions. Connected agents.

The quality and communication layer for AI coding agents. Lint your instructions, compose context across tools, and connect agents across repos and devices.

pip install enwrit
writ lint CLAUDE.md               # Instant quality score (0-100) for any instruction
writ init                         # Initialize writ in repo
writ search "code reviewer"       # Find instructions from Hub database
writ add code-review-agent        # Install instruction; from library, project or Hub

Try the live lint demo -- paste any instruction, get an instant quality score. Or browse 6000+ instructions on the Hub.

---

Lint Your Instructions

97% of AI instructions have quality defects. writ lint catches them. Scores 0-100 across 6 dimensions: Clarity, Verification, Coverage, Brevity, Structure, Examples. Works on any .md, .mdc, .txt, or YAML instruction file -- no writ init required.

writ lint .cursor/rules/my-rule.mdc
# Score: 34 / 100
# - 6 instances of vague language ("try to", "consider", "if possible")
# - No verification commands (agents can't check their own work)
# - No code examples
# Dimension       Score  Summary
# Clarity            43  Moderate              
# Structure          44  Moderate            
# Coverage           31  Needs improvement              
# Brevity            51  Moderate              
# Examples           15  Critical
# Verification       10  Critical
# Suggestions:
#   Replace vague phrases with imperative commands
#   Add verification: test/build/lint commands (2-3x quality impact)
#   Add 1-2 code examples showing desired patterns

Several options depending on user preferences - fast, local & cloud. The ML-powered solution is TF-IDF + LightGBM based off teacher-student distillation of 6,000+ ratings & suggestions of instructions. writ-lint-08B is a fine-tuned Qwen3.5-0.8B on the same dataset.

writ lint CLAUDE.md                     # Score any file (ML-powered, local, free)
writ lint AGENTS.md --deep              # AI-powered analysis (Gemini, via enwrit.com)
writ lint AGENTS.md --deep-local        # Local AI analysis (writ-lint-0.8B, GPU-accelerated)
writ lint rules.mdc --json              # Machine-readable JSON output for CI
writ lint --ci --min-score 60           # Exit 1 if score too low (CI gate)

Try it in your browser -- paste any instruction, get an instant score.

---

30-Second Quickstart

# 1. Install
pip install enwrit

# 2. Initialize in any repo (auto-detects your stack, installs writ-context rule)
writ init

# 3. Search and add from 6,000+ instructions on the Hub
writ search "code review"
writ add code-review-agent         # Fetches from Hub, saves to .writ/, writes to IDE dirs

That's it. Your IDE now has battle-tested instructions -- no copy-paste, no manual conversion.

What Makes writ Different

Capability	What it means
Instruction linting	6-dimension quality scoring (0-100) for any AI instruction file. Code-based + AI-powered.
Context composition	Layer project + team + agent + handoff context into one coherent instruction set
Multi-format export	Cursor .mdc, Claude Code, Kiro steering (auto-detected); AGENTS.md, Copilot, Windsurf (opt-in)
Agent communication	Structured conversations between agents across repos and devices
Personal library + cloud sync	writ save → writ add --lib on any device. Your instructions follow you.
Hub with 6,000+ instructions	Semantic search across rules, agents, programs (PRPM + enwrit). writ search <query> / writ add <name>
MCP server	One line in your config and any agent can search/install from the Hub

How It Works

writ writes to native IDE files -- it does NOT call LLM APIs.

Tool	writ writes to	Mode
Cursor	.cursor/rules/writ-*.mdc	Auto-detected
Claude Code	.claude/rules/writ-*.md	Auto-detected
Kiro	.kiro/steering/writ-*.md	Auto-detected
AGENTS.md	AGENTS.md	Opt-in (--format)
GitHub Copilot	.github/copilot-instructions.md	Opt-in (--format)
Windsurf	.windsurfrules	Opt-in (--format)

When you run writ add reviewer, the tool composes all relevant context and writes it directly into the files your IDE already reads.

Context Composition

The core innovation. Each agent's context is composed from 4 layers:

Layer 4: Handoff context      ← Output from another agent
Layer 3: Agent's instructions  ← The agent's own role
Layer 2: Inherited context     ← From parent agents
Layer 1: Project context       ← Auto-detected (languages, frameworks, structure)

writ add implementer                     # Compose with project context

MCP Server

Optional supplement for MCP-only users or features the CLI can't provide (e.g. writ_compose, writ_chat_send_wait).

writ mcp install      # Auto-detects Cursor, VS Code, Claude Code, Kiro, Windsurf
                      # Writes MCP config (slim mode), preserves existing servers
writ mcp uninstall    # Remove writ MCP config from all IDEs

Slim mode (default via writ mcp install): exposes only 2 MCP-exclusive tools -- the CLI already provides everything else with less token overhead.

Full mode (for MCP-only users without the CLI):

{"mcpServers": {"writ": {"command": "uvx", "args": ["enwrit", "mcp", "serve"]}}}

Hub: Browse & Install

The enwrit Hub has 6000+ curated instructions across three tiers:

• Rules -- passive context that shapes agent behavior (verification loops, commit hygiene, no-secrets)
• Agents -- on-invocation workers (code review, git commit, documentation, security audit)
• Programs -- autonomous metric-driven loops (test coverage optimizer, dependency freshness, dead code eliminator)

writ search "code review"          # Search from CLI
writ add code-review-agent         # Install into your project + activate in your IDE

Templates

Bootstrap an agent team in seconds:

writ init --template fullstack     # Architect + implementer + reviewer + tester
writ init --template python        # Python developer + reviewer
writ init --template typescript    # TypeScript developer + reviewer
writ init --template rules         # Project rule + coding standards

Personal Library & Cloud Sync

writ save my-reviewer              # Save to library (local + cloud if logged in)
writ save my-reviewer --local      # Save locally only
writ login                         # Authenticate for cross-device sync
writ add my-reviewer --lib         # Load from library on any machine
writ list --library                # See everything (local + remote)

Cross-Project Memory

writ memory export research-insights     # Export from current project
writ memory import research-insights     # Import in another project

Agent-to-Agent Communication

Agents can have structured conversations across repos:

writ peers add partner-repo --path ../partner-repo
writ chat start --with partner-repo --goal "Review API design"
writ chat send <conv-id> "Here's my proposed schema..."
writ inbox                          # Check for responses

All Commands

Command	Description
writ init	Initialize in current repo (auto-installs writ-context rule)
writ add <name>	Add instruction: project -> library -> Hub -> create new; auto-writes to IDE dirs
writ add <name> --lib	Force fetch from personal library
writ add <name> --from prpm	Install directly from PRPM registry
writ add --from <url>	Import instruction from any URL
writ add <name> --format cursor	Export to a specific format
writ add --file <path>	Import markdown file(s) or directory
writ list	List all instructions in project
writ list --library	List personal library (local + remote)
writ remove <name>	Remove instruction
writ save <name>	Save to personal library (syncs to cloud if logged in)
writ save <name> --local	Save locally only (skip cloud sync)
writ search <query>	Semantic search across Hub (6,000+ instructions)
writ search <query> --min-score N	Filter search results by quality score
writ search <query> --type agent	Filter by type (agent, rule, program, skill)
writ publish / unpublish	Make publicly discoverable
writ login / logout	Authenticate with enwrit.com
writ register	Create account
writ lint [file|name] [--deep] [--deep-local]	Quality score (0-100, 6 dimensions)
writ lint <file> --json	Machine-readable JSON output
writ lint <file> --ci --min-score N	CI gate: exit 1 if score below threshold
writ diff <file>	Compare lint score vs previous git commit
writ upgrade [name]	Pull latest version of installed instructions from source
writ sync	Bulk bidirectional library sync (confirmation for large ops, --undo to revert)
writ mcp install	Auto-configure MCP server in detected IDEs (slim mode)
writ mcp uninstall	Remove writ MCP server from IDE configs
writ mcp serve	Start MCP server (21 tools full / 2 slim, called by IDE)
writ chat start/send/inbox	Agent-to-agent conversations (supports --file attachments)
writ memory export/import	Cross-project memory
writ handoff create	Create agent handoff
writ review <name>	Browse/submit reviews
writ threads	Knowledge threads
writ approvals	Human-in-the-loop approval management
writ peers add/list/remove	Manage peer repo connections
writ connect	Interactive peer setup wizard

GitHub Action

Gate PR quality with writ lint in CI. You choose which files to lint:

- uses: enwrit/writ@main
  with:
    files: ".cursor/rules/*.mdc"    # Glob pattern for files to lint
    min-score: 50                   # Fail if any file scores below this

Not all instruction types need the same lint emphasis. E.g. skills and planning instructions may score lower on verification/examples by design -- choose your files and thresholds accordingly.

Instruction Format

Instructions are YAML files in .writ/, routed by task_type:

name: reviewer
description: "Code reviewer for TypeScript"
version: 1.0.0
task_type: agent    # agent | rule | context | program | template
tags: [typescript, review]
instructions: |
  You are a code reviewer specializing in TypeScript.
  Focus on: type safety, component composition, performance.
composition:
  inherits_from: [architect]
  project_context: true

Users interact primarily with markdown. Use writ add --file to import .md, .mdc, or .txt files directly.

Development

git clone https://github.com/enwrit/writ.git
cd writ
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate
pip install -e ".[dev]"
pytest                    # 540+ tests
ruff check src/ tests/

License

MIT

Links

• Hub: enwrit.com/hub -- Browse and install instructions
• Website: enwrit.com
• GitHub: github.com/enwrit/writ
• PyPI: pypi.org/project/enwrit