kstack

Personal AI software factory. Inspired by gstack, built for GitHub Copilot CLI first, agent-agnostic by design.

kstack is my opinionated set of slash-command skills for AI coding agents. It turns a blank prompt into a structured sprint: Plan → Build → Review → Ship. Each skill is a Markdown file that encodes a specialist persona — planner, reviewer, debugger, release engineer — and a repeatable workflow.

I'm @chokevin. I ship infrastructure and platform code for a living, and I got tired of re-typing the same "be rigorous, check for X, verify Y" prompts every time. kstack is the muscle memory, written down.

This is a work-in-progress. The point is to dogfood it daily and let it evolve.

Why another one?

gstack is phenomenal and I borrow its sprint model shamelessly. But gstack targets Claude Code as the primary host, and it's opinionated around a founder/product workflow. kstack is:

• Copilot CLI first — skills are drop-in under ~/.copilot/skills/ or invoked via custom instructions
• Infra-aware — skills assume you're often working on Terraform, Kubernetes, Helm, Azure, and multi-repo CLIs, not just web apps
• Small surface area — a few skills I actually use, not a library I might use someday
• Portable format — each skill is plain Markdown with YAML frontmatter, works in Copilot CLI, Claude Code, Cursor, etc. with thin host adapters

Skills (v0)

Skill	Role	What it does
/plan	Tech Lead	Turns a vague ask into a scoped plan with a SQL todo list and explicit non-goals.
/research	Researcher	Deep multi-source research (arxiv, GitHub, Reddit/X/HN, postmortems, talks). Steelmans the opposite. Memo ends with adopt / adapt / skip / further-research.
/reckon	Service Re-evaluator	Honest re-evaluation of a long-lived service slice. Produces a one-page memo with a binding decision (Harden / Extend / Refactor / Deprecate / Freeze).
/review	Staff Engineer	Rigorous review of uncommitted / branch changes. High signal-to-noise, no style nitpicks.
/investigate	Debugger	Root-cause methodology. Iron Law: no fixes without a verified hypothesis.
/explore	Subagent Foreman	Fan out parallel read-only subagents to map a codebase / spec an API surface / compare alternatives without burning the main context.
/unstuck	Loop-breaker	You're spinning. Same fix, same error, three iterations in. Stops the loop, names it, forces a layer change.
/ship	Release Engineer	Lint, test, commit, push, open PR. Verifies green CI before declaring done.
/retro	Coach	End-of-session retrospective: what worked, what to add to kstack next.

Domain skills

These only apply if you work on the same systems I do. They live in this repo because it's my personal stack; ignore them if they don't match your world.

Skill	Role	What it does
/airun-triage	Distributed Cluster Triage Engineer	Layered triage (Ray → Kueue → k8s scheduler → MIG → node pool/region) for stuck/failed/slow jobs on a multi-region AKS cluster. Logs every diagnosis to build a corpus over time.
/kernel-recall	Kernel Optimization Archivist	Searches the local kevin-obsidian vault for prior GPU/ML-kernel optimization notes, benchmark results, profiling lessons, and experiment history.
/cpu-kernel-recall	OS/Linux Kernel Module Archivist	Searches local memory and source references for CPU/host kernel modules, Linux driver configs, Secure Boot/AppArmor, OFED/MOFED, and driver-runtime gotchas that can confound GPU benchmarks.
/profiler-handoff	Profiler Artifact Release Engineer	Packages PyTorch/Perfetto/Nsight/HTA profiler results and torchview/Netron/Graphviz model graphs into Hermes-ready artifact_handoff events.
/obsidian-improve	Obsidian Context Gardener	Edits the local kevin-obsidian vault to improve context loading: indexes, context cards, promoted learnings, and validation.
/obsidian-crystallize	Obsidian Knowledge Crystallizer	Mines raw vault material and routes durable knowledge to the right home: learnings, context cards, context maps, research, reckonings, or retros.
/upstream-friction	Integration Boundary Architect	Finds downstream hacks that should be upstream contract fixes, forces an owner decision, and prevents workaround sprawl.

More will land as I find friction in my own workflow.

Install

Requirements: GitHub Copilot CLI, git, bash.

git clone --depth 1 https://github.com/chokevin/kstack.git ~/.kstack
~/.kstack/bin/setup

The setup script:

1. Symlinks ~/.kstack/skills/* into ~/.copilot/skills/ (or your agent's skills dir)
2. Appends a # kstack section to ~/.copilot/instructions.md listing the available skills

Run ~/.kstack/bin/setup --help for other agents (Claude Code, Cursor, Codex).

Use

In any Copilot CLI session:

/plan add retry + backoff to the baker provisioner
/review
/ship

Or reference a skill inline: "follow the kstack /investigate methodology on this flaky test."

Philosophy

1. One skill = one persona = one workflow. No swiss-army knives.
2. Skills compose. /plan writes artifacts /review and /ship can read.
3. Agent-agnostic Markdown. If a skill needs host-specific tooling, it degrades gracefully.
4. Ship the narrowest wedge. Every skill starts as the 20% that covers 80%. Iterate from real use.

License

MIT. Fork it, remix it, make it your own stack.