Claude Code Skills Pack: 15 Skills for Senior-Engineer Work

Stop re-explaining your engineering standards to your AI agent every session.

Why not just ask your agent?

You can, and a good model will often do fine. But without standing instructions, every session re-learns how you want things done, and the result varies: a review full of style comments, a bug "fixed" by guessing, a PR that says "tested locally" with nothing to back it, a column renamed in one step on a live table, an API key echoed into a log. Each miss costs a re-prompt loop (by our estimate 5-20 minutes including re-review), and the ones nobody catches become production bugs.

You could have your agent write these rules for you. Getting 15 of them to trigger on the right requests, and only those, takes trial and error: in our own testing, two requests landed on competing skills until we added one-line redirects. Expect a day or two to write and tune a set like this yourself.

What this saves you (conservative assumptions, per developer):

• Time: 20-40 re-prompts or fix-ups avoided a month at about 10 minutes each is about 3-7 hours a month.
• Money: at ₹1,000-2,500 an hour, that is ₹3,300-₹16,700 a month. One unsafe migration or leaked key that costs a day of cleanup is another ₹8,000-₹20,000.
• Payback: ₹999 is less than one saved hour at ₹1,000 an hour.

These are estimates from the stated assumptions, not measurements of your team.

Before and after, from our test runs (same model, same repo with a planted bug, same request):

• Without the skill: "Fix: the one-line change is return 0 → return total. Tests: I added two ... All 4 tests pass." The tests were written after the fix.
• With debug-loop: "Before the fix it failed with AssertionError: 0 != 5000, which is the cause I expected. ... The original repro now prints 5000 9999 9000."

What's inside

Review and ship

• pr-description: reviewer-first PR bodies; every testing claim is a real command and result; risk and rollback included.
• diff-review: findings ranked Blocker to Low, each with file:line, a concrete failure scenario and a fix. No style nits.
• merge-coordinator: lands several branches in dependency order with in-memory trial merges; regenerates lockfiles instead of hand-merging.
• release-handover: a handover a teammate and their AI agent can follow without asking you, with a paste-ready agent brief.

Fix and test

• debug-loop: reproduce, isolate, prove with a failing test, fix, verify. No guess-patches.
• test-writer: tests in your repo's style that fail for the right reason first.
• dependency-upgrade: one upgrade per commit, changelogs read, versions pinned, one-revert rollback.

Data and safety

• safe-db-migration: expand/contract, idempotent, batched backfills, single head, nothing destructive by default.
• secrets-safe-ops: secrets never printed or committed; set/not-set checks; staged-change scanning.
• incident-triage: severity, the first 15 minutes, comms templates, blameless postmortems.

Deploy and verify

• live-verification: checks a deploy as a user would, with evidence, and lists what was not checked.
• deploy-without-ci: safe manual deploys with ordered uploads, retries, health gates and an audit note.

Work style

• low-resource-builds: lock and RAM guard so parallel agents and builds do not freeze an 8-16 GB laptop.
• parallel-sessions: briefs, file ownership and worktrees for splitting work across agents.
• ux-copy: plain, specific microcopy in the user's language, with honest paywalls and no dark patterns.

How it works

1. Install: run one script (macOS, Linux or Windows). Skills go into ~/.claude/skills/ or a single project's .claude/skills/.
2. Talk normally: "review my changes", "this test started failing", "rename this column", "is my key set?"
3. The skill kicks in: Claude Code matches your request to the skill's description and follows its rules, steps and definition of done.

Works with Claude Code natively. Also includes rules files generated from the same skills for Cursor, Codex, Antigravity and other agents (AGENTS.md, Cursor .mdc rules, and one RULES.md to paste anywhere). Those files were generated and checked, but not tested inside those tools.

How it was tested

• Trigger selection on 32 natural requests (including 5 that should match nothing), using only the skill descriptions as Claude Code does: every request picked the intended skill.
• Five skills run on real tasks in a throwaway repo, three against a no-skill baseline. Trimmed transcripts are included.
• All 9 scripts syntax-checked and run on Windows (Git Bash, PowerShell 5.1).
• Honest limit: the Claude Code CLI was not available on the build machine, so runs used Claude sub-agents with tool access rather than claude -p sessions. macOS, Linux and PowerShell 7 were not tested.

Who it's for

• Developers and small teams who use Claude Code daily and are tired of repeating their standards.
• Teams that want the same review, PR, migration and release bar whether a person or an agent does the work.
• Anyone running several agents or sessions at once, especially on a laptop.

Who it's not for

• If you want an agent that moves faster with less checking: these skills make it show its evidence.
• Teams that already have detailed agent rules for all of these areas.
• It does not replace CI, code review by people, or an incident process.

FAQ

Does it work with the Claude Code CLI and the desktop app? Yes. Both read skills from ~/.claude/skills/ (all projects) and .claude/skills/ (one project). Start a new session after installing.

Which models? Skills are plain instructions, so they work with whichever Claude model your Claude Code uses. They were tested with a current Claude model; smaller or older models may follow long rules less consistently.

Does it send data anywhere? No. The skills are text files on your machine. There is no telemetry. The scripts only contact URLs you give them (smoke checks and health gates).

Can my whole team use it? Yes. The licence covers you and your team inside your organisation, including committing the skills to your private repos. Resale and public redistribution are not allowed.

Can I edit the skills? Yes. docs/CUSTOMISE.md shows where to add your stack's commands and conventions, and docs/WRITE-YOUR-OWN.md explains how to write new skills that trigger reliably.

Will it overwrite my existing skills? No. The installer skips any skill folder that already exists unless you pass --force, and the uninstaller only removes this pack's skills.