ZenKit

Disciplined workflows for Claude Code.

Add structured specs, plans, audits, checkpoints, and handoffs to any project. Native slash commands and skills. No daemon, no framework, no lock-in.

Landing Page | Playground | npm

npx zenkit init claude

That's it. This creates:

.claude/commands/zenkit-spec.md
.claude/commands/zenkit-plan.md
.claude/commands/zenkit-build.md
.claude/commands/zenkit-audit.md
.claude/commands/zenkit-checkpoint.md
.claude/commands/zenkit-handoff.md
.claude/skills/zenkit-audit/SKILL.md
.claude/skills/zenkit-handoff/SKILL.md
.claude/skills/zenkit-checkpoint/SKILL.md
CLAUDE.md

Then in Claude Code, use the slash commands:

/zenkit-spec "add user profile page"
/zenkit-plan
/zenkit-build
/zenkit-audit
/zenkit-checkpoint

Every command produces structured output with: context, assumptions, constraints, decision, deliverable, risks, open questions, next handoff.

Most AI-assisted development drifts. Agents wander from the plan, burn tokens on narration, hide uncertainty behind confident prose, and lose context between sessions. ZenKit adds discipline through file-based conventions that Claude Code already supports — slash commands, skills, and project memory.

Start with layer 1. Add layer 2 when you want machine-verifiable workflows.

npm install zenkit

import { validate, createHandoff, loadFeatureSpec } from 'zenkit'

const result = validate('handoff', myData)
if (!result.valid) console.error(result.errors)

const handoff = createHandoff({
  context: 'Completed auth module',
  assumptions: ['Redis available'],
  decision: 'JWT with refresh tokens',
  deliverable: { type: 'code', description: 'Auth module' },
  next_agent: 'frontend-architect',
})

npx zenkit init claude              # Install Claude Code commands + skills
npx zenkit init                     # Scaffold full protocol structure
npx zenkit validate handoff data.json   # Validate JSON against schema
npx zenkit benchmark:all            # Run all benchmark specs
npx zenkit audit                    # Full audit with structured report
npx zenkit status                   # Project health check

ZenKit includes a benchmark harness that verifies acceptance criteria against the actual implementation.

• 5 feature specs, 44 acceptance criteria, 131+ checks
• 7 verification types: file_exists, file_contains, schema_count, examples_valid, schemas_consistent, test_passes, json_path_equals
• Telemetry is always labeled as estimated or actual — never fabricated
• Every result includes uncertainty and limitations arrays

npx zenkit benchmark:all     # Run all specs
npx zenkit audit             # Run benchmarks + produce audit report

1. Thin over grand — Files, not frameworks.
2. Protocol over persona — Contracts, not theatrical agent identities.
3. Bounded autonomy — Assumptions explicit, uncertainty recorded.
4. Validation over narration — Artifacts over prose.
5. Claude-native — Slash commands and skills, not a parallel universe.

See CONTRIBUTING.md.

• Philosophy | Architecture | Commands
• Agents | Benchmarking | Self-Audit
• Roadmap

MIT — see LICENSE.