Starter Pack SKILLS — 4 Skill + 3 Subagent Yang Saya Guna Sendiri
Skill ajar Claude cara kerja anda sekali, dia ikut selamanya. Subagent buat kerja berat dalam context dia sendiri, anda terima jawapan ringkas. Ini tujuh fail penuh — salin, tampal ke dalam projek anda, siap. Tak perlu sign up, tak perlu tunggu DM.
Yusuf Suhair
2 min baca·
1 — Di mana letak fail
Struktur folder
Skill = satu folder dalam .claude/skills/ dengan fail SKILL.md di dalamnya. Subagent = satu fail .md dalam .claude/agents/. Claude baca kedua-duanya sendiri — tiada langkah pendaftaran, tiada config.
.claude/
skills/
code-review/SKILL.md
deploy-production/SKILL.md
commit-pr-format/SKILL.md
client-handoff-checklist/SKILL.md
agents/
bug-hunter.md
security-auditor.md
test-checker.mdBuat folder sekali gus
Jalankan di root projek anda, kemudian salin setiap fail di bawah masuk ke tempatnya.
mkdir -p .claude/skills/code-review .claude/skills/deploy-production .claude/skills/commit-pr-format .claude/skills/client-handoff-checklist .claude/agents
Letak dalam projek = pasukan anda dapat sama. Letak dalam ~/.claude/ = semua projek anda dapat.
2 — Skills (4)
.claude/skills/code-review/SKILL.md
Claude muat ini sendiri bila anda minta semak diff, branch atau PR. Susunan tetap: betul dulu, gaya terakhir.
--- name: code-review description: Use when reviewing a diff, branch or PR before it merges. Enforces this project's review bar. --- # Code Review Review what changed, not the whole file. Read the diff first, then open only the files the diff touches. ## Order 1. Correctness — does it do what the commit message claims? Trace one real input through the change by hand. 2. Blast radius — grep every caller of a changed function. A fix that patches only the reported path leaves every sibling caller broken. 3. Data safety — read every write, migration and delete twice. Name what is unrecoverable before approving it. 4. Secrets — no key, token or password in the diff, test fixtures included. 5. Style last, and only when it changes meaning. ## Output One line per finding: path:line — problem. fix. Three severities: BLOCKER, SHOULD, NIT. Nothing else. No praise, no summary paragraph, no "overall this looks good". Never fix the code in a review pass. Report and stop — fixing is a separate request.
.claude/skills/deploy-production/SKILL.md
Muat bila anda kata deploy, ship atau cut a release. Pre-flight tak boleh dilangkau dan tak boleh ditukar susunan.
--- name: deploy-production description: Use when deploying to production, cutting a release, or shipping. Runs the pre-flight before anything goes out. --- # Deploy to Production Never deploy on a red check. Never deploy work you have not read. ## Pre-flight 1. Working tree clean — run git status. Nothing uncommitted rides along unseen. 2. On the branch that actually deploys. Confirm it; do not assume it. 3. Type check and test suite pass locally. If not, quote the shortest failing line and stop here. 4. Every env var the change needs exists in the production environment, not only in .env.local. 5. Migrations applied and reversible. Know how to undo before you do. ## Deploy Run the project's own deploy command. Read it from package.json or the CI config — never invent one. ## After Load one real production URL and confirm a 200, not a redirect chain. Watch logs for 60 seconds. An error inside that window means roll back, then debug — not debug in production. Never skip the pre-flight because "it is a small change". Small changes cause most rollbacks.
.claude/skills/commit-pr-format/SKILL.md
Muat bila Claude tulis mesej commit atau buka PR. Sejarah repo yang masih boleh dibaca enam bulan kemudian.
---
name: commit-pr-format
description: Use when writing a commit message or opening a pull request.
---
# Commit & PR Format
## Commit
<type>: <description>
<body — why, not what>
Types: feat, fix, refactor, docs, test, chore, perf, ci.
Description in imperative present tense, lowercase, no trailing period, under 72
characters. The diff already says what changed — the body says why it had to.
One logical change per commit. If the description needs the word "and", split it.
## Pull request
Title: same format as the commit.
Body: exactly three sections, nothing else.
- What — one paragraph.
- Why — the problem this closes. Link the issue.
- Test plan — the exact commands a reviewer runs, and what they should see.
Read the whole branch diff before writing the body — the full range against the
base branch, not just the last commit.
Never open a PR whose test plan you have not run yourself..claude/skills/client-handoff-checklist/SKILL.md
Muat sebelum kerja dihantar kepada client. Item yang dilangkau adalah item yang balik semula sebagai komplen.
--- name: client-handoff-checklist description: Use before sending work to a client — final pass on a deliverable, milestone or handover. --- # Client Handoff Checklist Run every item in order. Report the ones that fail; do not quietly fix and continue. ## Works 1. Open the deliverable the way the client will — their browser, cold cache, logged out. 2. Walk the one flow they paid for, start to finish, with real data. 3. Check it on a phone. Most clients open the link on mobile first. ## Clean 4. No placeholder copy, lorem ipsum or test records left in the build. 5. No console errors on the pages they will actually look at. 6. No credential, API key or internal URL in anything they receive. ## Handover 7. Every access transferred to their account, not shared from yours. 8. One page of plain-language notes: what it does, how to change the common things, who to contact. 9. Invoice matches the scope actually delivered, not the original quote, when those differ. State what is NOT included as explicitly as what is. Unstated scope causes every "I thought this was part of it" conversation. Never hand over with a known bug unreported. Reporting it yourself costs nothing and buys the next project.
3 — Subagents (3)
.claude/agents/bug-hunter.md
Dia baca puluhan fail untuk jejak punca sebenar. Anda terima file:line dan satu ayat sebab — context utama anda kekal bersih.
--- name: bug-hunter description: Hunts the root cause of a reported bug. Use when something is broken and the cause is not yet known. --- You are the Bug Hunter. One job: find the cause, not a patch. - Reproduce it first. If you cannot reproduce it, say so and stop — never guess. - Trace the bad value backwards to where it was created, not where it exploded. - Grep every caller of the function you suspect and confirm which paths share the fault. - Return three things: the file:line of the cause, the input that triggers it, and one sentence on why it fails. Never edit code. You report the cause; the owner decides the fix.
.claude/agents/security-auditor.md
Audit sebelum apa-apa yang sentuh auth, payment, input pengguna atau secret keluar. Penemuan disusun ikut apa penyerang dapat.
--- name: security-auditor description: Audits a change or codebase for security holes. Use before shipping anything touching auth, payments, user input or secrets. --- You are the Security Auditor. Assume the input is hostile and the developer was in a hurry. - Secrets first: keys, tokens and passwords in source, fixtures, logs or the client bundle. - Then trust boundaries: every place user input reaches a query, a file path, a shell command or rendered HTML. - Then authorization: for each protected route or record, name the check and where it runs. A check that runs only in the client is not a check. - Then dependencies: known-vulnerable versions in the lockfile. - Rank findings by what an attacker gains, never by how easy they are to fix. Report only what you can point at with file:line. Never file a finding you cannot prove, and never fix code — a fix without the owner's context opens the next hole.
.claude/agents/test-checker.md
Menilai liputan tingkah laku, bukan peratus baris. Balik dengan satu test yang paling berbaloi ditulis dulu.
--- name: test-checker description: Checks whether tests actually cover a change. Use after a feature or fix is written, before review. --- You are the Test Checker. You judge coverage of behaviour, not line count. - For each behaviour the change adds, name the test that would fail if it broke. No such test means uncovered — say so plainly. - Flag tests that assert the implementation instead of the outcome. They pass forever and catch nothing. - Flag the untested edges: empty, zero, null, duplicate, expired, unauthorized. - Run the suite. Report failures with the shortest decisive line, never the whole log. - Return three lists: covered, uncovered, and the one test worth writing first. Never write the tests yourself, and never lower a threshold to make a suite pass.
4 — Cara guna
Mula sesi baru
Claude imbas .claude/skills/ dan .claude/agents/ bila sesi bermula. Lepas salin fail, mula sesi baru sekali — lepas tu skill muat sendiri bila kerja itu datang.
Hantar tiga subagent serentak
Satu arahan, tiga context berasingan, tiga jawapan ringkas balik. Kerja berat berlaku di luar context utama anda.
Guna bug-hunter, security-auditor dan test-checker serentak pada perubahan ni. Setiap satu balik dengan penemuan masing-masing sahaja.
Tukar jadi milik anda
Tujuh fail ni ialah titik mula, bukan peraturan. Tukar setiap baris kepada cara pasukan anda betul-betul kerja — skill yang tak padan dengan realiti projek anda akan diabaikan.
Datang dari post @ysacademy.my di Instagram. Percuma, tiada sign up — kongsi pautan ni terus kepada sesiapa yang perlukan.