Conjure

Bind the daemon. Ship the code.

The production-grade harness kit for Claude Code — the sigil that keeps your AI inside the circle.

Quickstart • Why Conjure? • Features • Stack profiles • Compliance • Migration • Docs • Compare

---

What is Conjure?

Conjure is the missing init kit for Claude Code. It sets up the four-layer harness Anthropic recommends — CLAUDE.md + lazy-loaded Skills + isolated Subagents + deterministic Hooks — and ships with safe migrations from every other AI assistant, 9 stack profiles, 4 compliance overlays, persistent knowledge-graph integration, and a CLI that makes it all one command.

Built for high-stakes work where adherence matters. People-lives-depend-on-it defaults: backup-before-mutate everywhere, size caps enforced, no curl | sh foot-guns, cross-platform Node.js hooks.

The problem: most CLAUDE.md setups fail the same way — a 500-line monolith that Claude ignores after the first hundred lines. Conjure enforces the ≤100-line cap, splits content into lazy-loaded skills, and promotes non-negotiables to deterministic hooks.

✨ Why Conjure?

• 🕯️ Four-layer binding circle — CLAUDE.md (advisory) + Skills (lazy) + Subagents (isolated) + Hooks (deterministic). Four sigils, one bound daemon.
• 🧠 Knowledge-graph aware — first-class graphify integration. Persistent codebase knowledge that survives sessions.
• 🛡 Safe migrations — backup-before-mutate from Cursor, Aider, Continue, GitHub Copilot, Windsurf, and existing .claude/ configs.
• 🎯 9 stack profiles — Java-Spring, Python-FastAPI, TS-Next, Rust-Axum, Go-Gin, Node-Nest, Monorepo, Polyglot, Data Science.
• 🔒 4 compliance overlays — HIPAA (with PHI scan hook), SOC 2, GDPR, PCI.
• 🧪 112 self-tests, all green — change confidently; CI runs on every PR.
• 🌍 Cross-platform — bash hooks for POSIX, Node.js .mjs hooks for native Windows.
• 📦 Plugin-ready — installable via Claude Code Marketplace.
• 🔁 Compound engineering — Stop hook proposes new rules from session corrections.
• 📐 Eval-backed sizing — caps from 2,455-evaluation study; less context = better adherence.

🚀 Quickstart

# 1. Install
curl -sSL https://raw.githubusercontent.com/mohandoz/conjure/main/install.sh | bash

# 2. Initialize a project (auto-detects new or existing)
cd /path/to/your/repo
conjure init existing --profile=python-fastapi .

# 3. Open Claude Code, paste PROMPT.md, watch the daemon obey

That's it. Run conjure audit anytime to verify health.

🧰 Features

Core scaffold ≤100-line root CLAUDE.md (hard cap enforced) 17 skill templates with progressive disclosure 6 subagent definitions for isolated context 5 hook scripts with correct exit codes (bash + .mjs) .claudeignore, .editorconfig, .gitattributes Per-project .claude/README.md + EVENT-LOG.md Standard docs: ARCHITECTURE, RUNBOOK, GLOSSARY, ADR template	Production guard-rails Backup-before-mutate on every change Version pinning per project (.claude/.conjure-version) Audit with size caps, schema validation, anti-pattern detection Pre-flight checks for tool availability Failure-mode docs for every common breakage JSON schemas for skill/agent frontmatter (IDE-validated) Compound-engineering loop for continuous improvement
Tool integrations 🕸 graphify — persistent knowledge graph 📚 context7 — live framework docs 🌐 firecrawl — web research 🔍 ast-grep — structural code search 📦 repomix — full-codebase context dumps 🗄 Postgres MCP — schema introspection 🐙 GitHub MCP — PRs / issues / Actions 🧮 Sequential Thinking MCP	Stack & compliance 9 stack profiles (Java/Python/TS/Rust/Go/Node/...) 4 compliance overlays (HIPAA/SOC2/GDPR/PCI) 6 migration paths (Cursor/Aider/Continue/Copilot/Windsurf/Claude) Monorepo support (nested CLAUDE.md per package) GSD workflow integration

🎯 Stack profiles

conjure init existing --profile=<stack> .

Profile	Stack	Build	Test
java-spring	Java 17+ / Spring Boot / Gradle	./gradlew build	./gradlew test
python-fastapi	Python 3.11+ / FastAPI / uv	uv sync	uv run pytest
ts-next	TypeScript / Next.js 15 / pnpm	pnpm build	pnpm test
rust-axum	Rust / Axum / cargo	cargo build	cargo nextest
go-gin	Go / Gin	go build ./...	go test ./...
node-nest	Node / NestJS / pnpm	pnpm build	pnpm test
monorepo	Turborepo / Nx / pnpm workspaces	per-package	per-package
polyglot	Mixed stacks	Make/Just	per-language
data-science	Python / Jupyter / dbt	uv sync	pytest + nbqa

🔒 Compliance overlays

Layer one or more on top of a profile:

bash $CONJURE_HOME/compliance/hipaa/apply.sh .
bash $CONJURE_HOME/compliance/soc2/apply.sh  .
bash $CONJURE_HOME/compliance/gdpr/apply.sh  .
bash $CONJURE_HOME/compliance/pci/apply.sh   .

Each overlay adds: CLAUDE.md non-negotiables, pre-commit guard hooks (e.g. PHI pattern scan for HIPAA), and a controls checklist under docs/compliance/.

⚠️ Overlays make the AI less likely to produce non-compliant code. They do NOT make you compliant — that requires people + process + audit.

🔄 Migration from other tools

Backup-before-mutate is automatic. Rollback is mv .claude.backup-<ts> .claude.

Source	Command
Existing hand-rolled .claude/	conjure migrate from-claude .
Cursor (.cursorrules, .cursor/rules/)	conjure migrate from-cursor .
Aider (.aider.conf.yml, CONVENTIONS.md)	conjure migrate from-aider .
Continue (.continue/config.json)	conjure migrate from-continue .
GitHub Copilot (.github/copilot-instructions.md)	conjure migrate from-copilot .
Windsurf (.windsurfrules)	conjure migrate from-windsurf .

See MIGRATION-GUIDE.md for details.

🧪 Quality

$ bash tests/run.sh
═══════════════════════════════════════════════════════════════════
PASS: 112    FAIL: 0
═══════════════════════════════════════════════════════════════════

Every PR runs:

• shellcheck on all .sh files
• JSON Schema validation on all .json files
• Frontmatter validation on every SKILL.md / agent .md
• Size cap enforcement (CLAUDE.md ≤100, SKILL.md ≤200, agent ≤80)
• @import detection (forbidden in CLAUDE.md)
• Exit code check on hooks (must be exit 2, never exit 1)
• Migration coverage (every documented source has a working script)
• Profile + compliance coverage (every documented overlay applies cleanly)

📖 Documentation

Doc	What
PROMPT.md	The master prompt — paste into Claude Code
MIGRATION-GUIDE.md	Safe migration playbook
FAILURE-MODES.md	Symptom → cause → fix for every common breakage
COMPARISON.md	How Conjure compares to alternatives
checklists/NEW-PROJECT.md	Greenfield step-by-step
checklists/EXISTING-PROJECT.md	Brownfield step-by-step
checklists/AUDIT.md	Periodic health check
checklists/ONBOARDING.md	Onboard new dev to a Conjure repo
reference/BEST-PRACTICES.md	2026 consolidated, eval-backed
reference/TOOLS-CATALOG.md	Every CLI tool worth knowing
reference/MCP-SERVERS.md	Which MCPs to install + configs
reference/ANTI-PATTERNS.md	What NOT to do (with evidence)
reference/SIZING.md	Line counts + token budgets
reference/COMPACTION.md	Surviving context compression
reference/PROMPTING-PATTERNS.md	Trigger-action and other patterns
planning/ROADMAP.md	What's planned next
planning/GSD-INTEGRATION.md	GSD workflow integration
CHANGELOG.md	What changed per version
CONTRIBUTING.md	How to contribute
SECURITY.md	Security policy
CODE_OF_CONDUCT.md	Contributor Covenant
SUPPORT.md	Where to get help

🎓 Core principles (10)

1. Less context = better output. ETH Zurich + Anthropic eval data.
2. @imports load eagerly. Use prose refs or skills instead.
3. Skills are the real lazy loader. Progressive disclosure.
4. Hooks > advisory rules. Promote non-negotiables to hooks.
5. Subagents isolate context. Verbose work in fresh windows.
6. Persistent graph > re-reading files. Build once, query forever.
7. Trigger-action format. "WHEN X, DO Y" beats general guidance.
8. Order matters. Top of CLAUDE.md survives compaction; non-negotiables first.
9. Compound engineering. Every correction → candidate rule promotion.
10. Cite file:line. So future Claude can verify.

🤝 Contributing

PRs welcome. See CONTRIBUTING.md.

• Bug? Open an issue.
• Feature? Use the feature request template.
• Question? Use Discussions.
• Security? See SECURITY.md — do not open public issues.

🛣 Roadmap

• v0.3.0 — test fixtures per profile, skill firing telemetry, cost estimator.
• v0.4.0 — marketplace publication, Homebrew formula, Docker image, conjure publish-skill.
• v0.5.0 — interactive 3-way merge for updates, drift detector, auto-PR bot.
• v0.6.0 — workspace mode, cross-repo graphify orchestration.
• v1.0.0 — frozen schemas, signed releases, ≥10 production teams.

See planning/ROADMAP.md.

⚖ License

MIT — use freely. Attribution appreciated, not required.

🌟 Star history

View live star history chart →

_{Inline chart will appear here after the repo crosses ~10 stars (star-history.com rate-limits anonymous renders for fresh repos).}

🙏 Built on

• Claude Code by Anthropic
• graphify for persistent knowledge graphs
• context7 for live docs
• repomix, ast-grep, firecrawl
• Eval data from Anthropic's published benchmarks + the 2,455-eval community study
• ETH Zurich research on context size and task success
• GSD workflow (Get Shit Done) for our own development

---

Conjure — Code grows on Conjure the way a vine grows on a garden frame: structured but free.

Made with ⚙ by mohandoz