4 Agent Skills for backend engineers: design review, zero-downtime migration, production GenAI, forecasting

shubhamkumbhalkar — Thu, 17 Sep 2026 00:25:31 +0000

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.

I automated my job search across 7 job boards

shubhamkumbhalkar — Wed, 16 Sep 2026 23:31:01 +0000

Job hunting is a part-time job on top of your job. The worst part isn't the interviews, it's the daily grind of refreshing a dozen career pages, copy-pasting the same details, and rewriting the same cover letter with a new company name.

So I built a small tool to do the repetitive 80% and leave the 20% that actually needs me.

What it does

Every morning it:

Scans 7 job platforms (Lever, Greenhouse, Ashby, SmartRecruiters, Workday, Indeed/Google via JobSpy, Hiring.cafe).

Scores each opening against my resume from 0 to 100 with an LLM.

For anything above my threshold, drafts a tailored cover letter and a "why this company" answer.

Sends me one Slack card per role, with the apply link.

Then in Slack I tap the link and apply, react ✅ to mark it applied (it tracks the pipeline), or react 📞 to get interview prep, STAR stories mapped to likely questions, back in the thread. A weekly report summarizes how many roles I saw, scored, and applied to.

The one rule: keep a human in the loop

The tool does not auto-submit applications. That was deliberate. Blasting auto-filled applications is bad for employers and bad for you. The tool removes the toil (finding, scoring, drafting, tracking); I still make every apply decision and do every interview.

How it's built

Python, with a thin client per platform, so adding a board is one small file.
A pluggable AI backend: a local model for free, or the Anthropic API for speed.
Deduplication by job id and by title+company, a per-company cap per run, and platform balancing so a single board doesn't crowd out the rest.
A cron job runs the daily scan; a Slack bot watches reactions.

What I learned

Scoring against your actual resume, not just keyword matching, cuts the noise dramatically.
The Slack "one card per role, react to act" loop is what made me actually keep up with it.
Keeping a human in the loop made it something I trust to run every day.

Try it

It's open source (MIT): https://github.com/shubhamkumbhalkar/lever-auto-apply

PRs welcome, especially new platform clients. If you try it, tell me what breaks.

DEV Community: shubhamkumbhalkar