by elb-pr
SRE thinking applied to Claude Code, based on Boris Cherny's Q&A. It enforces a strict 4-stage pipeline with gates between each step. You literally cannot skip verification. You cannot ship without approval.
# Add to your Claude Code skills
git clone https://github.com/elb-pr/claudikins-kernelGuides for using ai agents skills like claudikins-kernel.
Last scanned: 5/30/2026
{
"issues": [],
"status": "PASSED",
"scannedAt": "2026-05-30T16:30:32.423Z",
"npmAuditRan": true,
"pipAuditRan": true
}claudikins-kernel is an open-source ai agents skill for AI coding assistants such as Claude Code, Codex CLI, and ChatGPT, built by elb-pr. SRE thinking applied to Claude Code, based on Boris Cherny's Q&A. It enforces a strict 4-stage pipeline with gates between each step. You literally cannot skip verification. You cannot ship without approval. It has 117 GitHub stars.
Yes. claudikins-kernel passed SkillsLLM's automated security scan — a dependency vulnerability audit plus prompt-injection heuristics — with no high-severity issues. You can read the full report in the Security Report section on this page.
Clone the repository with "git clone https://github.com/elb-pr/claudikins-kernel" and add it to your Claude Code skills directory (see the Installation section above).
claudikins-kernel is primarily written in Shell. It is open-source under elb-pr on GitHub, so you can review or fork the full source.
Yes. SkillsLLM lists many other AI Agents skills you can browse and compare side by side. Open the AI Agents category from the badge at the top of this page, or use the Related Skills and comparison links further down to weigh claudikins-kernel against similar tools.
No comments yet. Be the first to share your thoughts!
You asked Claude for a bug fix. He refactored half your codebase.
You asked Claude for a feature. He placed a bunch of stubs that look a little bit real.
You asked Claude if you should drink that coffee you forgot about, now you're feeling sick and threw up on your keyboard, which cost too much just to hear nice click clack noises whilst you code, which actually isn't that great because it wakes your dog up at night, then you have to take him out in the cold to poop whilst you're still sick and well...okay, maybe that one was just me, but the point still stands!
claudikins-kernel applies SRE discipline to AI workflows. It enforces a strict 4-stage pipeline with gates between each step. You literally cannot skip verification. You cannot ship without the Cynic's approval.
Constraint is freedom. By preventing shortcuts, you get code that actually works.
┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐
│ /outline │────▶│ /execute │────▶│ /verify │────▶│ /ship │
└──────────┘ └──────────┘ └──────────┘ └──────────┘
│ │ │ │
▼ ▼ ▼ ▼
taxonomy- babyclaude catastrophiser git-perfectionist
extremist spec-reviewer cynic
code-reviewer
Each arrow is a gate. Try to /ship without /verify passing? Blocked. Try to /execute without a plan? Blocked. The system enforces this - not guidelines, guardrails.
# Prerequisites: jq (JSON processor)
# Windows: winget install jqlang.jq
# Ubuntu/Debian: sudo apt install jq
# macOS: brew install jq
# Add the Claudikins marketplace
/marketplace add elb-pr/claudikins-marketplace
# Install the plugin
/plugin install claudikins-kernel
Restart Claude Code. Then:
# Start your first disciplined session
/claudikins-kernel:outline "Add user authentication to the app"
These aren't generic "agents". They're your synthetic staff - each with a job and a personality.
| Agent | Role | Personality |
|---|---|---|
| taxonomy-extremist | Researcher | The librarian. Categorises everything. Reads your codebase, external docs, the web - returns structured findings. |
| babyclaude | Implementer | The eager junior. Does exactly what you specify. One task, one branch, fresh context. No scope creep. |
| spec-reviewer | Compliance | The auditor. Did you do what you said you'd do? Mechanical check against acceptance criteria. |
| code-reviewer | Quality | The critic. Is it actually any good? Error handling? Edge cases? Naming? |
| catastrophiser | Verification | The QA lead who assumes everything will break. Runs your code, takes screenshots, curls your endpoints. Sees it working, doesn't trust tests alone. |
| cynic | Simplification | The senior engineer who hates complexity. If it can be done in 5 lines, won't let you use 10. |
| conflict-resolver | Merge Handler | The diplomat. When branches collide, proposes resolutions. |
| git-perfectionist | Documentation | The pedant. README not updated? Changelog wrong? Blocked until it's right. |
/outline - "Let's figure out what we're building"Iterative brainstorming until you have a solid plan.
Output: plan.md with a task table that /execute can parse.
/execute - "Let's build it"Execute the plan task by task with fresh agents and code review.
Key feature: Each babyclaude gets fresh context. No pollution between tasks.
/verify - "Does it actually work?"Claude must see the code working. Not trust. Verify.
Output: verify-state.json with unlock_ship: true if approved. Plus file hashes so /ship can detect tampering.
/ship - "Send it"Merge to main with proper docs and PR.
gh CLI.Output: Code on main. PR merged. Done properly.
| Protection | What it does |
|---|---|
| Cross-command gates | Can't skip steps. Execute needs plan. Ship needs verify. |
| State files | Each command writes to .claude/. Resume if context dies. |
| File locking | flock prevents race conditions on state writes. |
| Code integrity | SHA256 hashes ensure shipped code = verified code. |
| Session management | Stale session (4+ hours)? Warns you research might be outdated. |
| Human checkpoints | Nothing auto-proceeds. You approve every phase. |
Industrial-grade patterns adapted for AI workflows.
| Distributed Systems Pattern | Claude Code Adaptation |
|---|---|
| Circuit breakers | Stuck agent detection |
| Distributed tracing | Execution spans |
| Load shedding | Batch size limits |
| Coordinated checkpoints | Batch-boundary saves |
| Deadline propagation | Task time budgets |
| Exponential backoff | Retry with jitter |
Same principles, different scale. Reliability through structure - not speed through parallelism.
# Ubuntu/Debian
sudo apt install jq
# macOS
brew install jq
# Windows
winget install jqlang.jq
| Plugin | Purpose |
|---|---|
claudikins-tool-executor |
MCP access for research and verification |
claudikins-automatic-context-manager |
Context monitoring at 60% |
| Plugin | Purpose |
|---|---|
claudikins-klaus |
Escalation when truly stuck |
v1.2.0 - Full Release - Regularly Maintained.
View the marketplace | Changelog
MIT