Agent Skills are a simple idea: a folder of instructions an AI coding agent can load to do a specific task the same way every time. Most of the ones I saw were frontend or generic workflows, so I packaged four checklists I actually use on large-scale backend systems and open-sourced them.
Install them on any agent (Claude Code, Codex, Cursor):
npx skills add shubhamkumbhalkar/skills
distributed-systems-design-review
Point it at a design doc or PR and it reviews for the failure modes that cause real incidents: retries without idempotency, retry storms with no circuit breaker, missing backpressure, cross-store write drift, hot partitions, and missing golden-signal observability. For each gap it gives the concrete failure scenario, the blast radius, and a specific fix, then a Ship / Ship-with-changes / Needs-rework verdict.
zero-downtime-migration
Migrating a live or Tier-1 system is where teams get hurt. This skill produces a phased plan built on one rule: every phase is independently reversible. Expand (dual-write), backfill (idempotent, reconciled), shadow-verify against real traffic, incremental cutover (1/5/25/50/100%), then contract. It also reviews someone else's migration plan for one-way steps and missing rollbacks.
genai-production-readiness
LLM features fail differently: non-deterministic, manipulable by inputs, and they degrade silently. This skill checks the things a demo skips: an eval set with a regression gate, grounding and hallucination handling, prompt-injection defense on untrusted input, a human-in-the-loop boundary for consequential actions, cost/latency bounds and fallbacks, PII handling, and observability. It ends with a go / no-go.
time-series-forecasting-review
Forecasting is deceptively easy to get wrong: leakage and bad backtests inflate offline metrics, then quiet drift erodes them in production. This skill catches data leakage, random-fold validation (instead of rolling-origin backtests), missing naive baselines, mishandled seasonality and known-future covariates, the wrong error metric (MAPE near zeros), and no online accuracy monitoring. Includes notes for foundation models like TimesFM (zero-shot vs fine-tuned, context length, license).
Why
These encode the review discipline I use on systems that process a lot of revenue at high uptime, plus applied GenAI and forecasting in production. They are opinionated on purpose.
Repo (MIT): https://github.com/shubhamkumbhalkar/skills
PRs welcome. If you run them on your codebase, tell me what they caught and what they missed.
Top comments (0)