Skills Installation — Copy-Paste Prompts

What are "skills"?

You already have: VS Code ✅ · Claude Code ✅ · DeepSeek working ✅ (from the Setup Guide). This guide adds the skills — extra toolkits that give the AI specialised abilities — and your first agent.

Some examples of what you'll be able to do:

• Turn a report into a clean PDF in one sentence
• Save your work-in-progress on Friday and restore it Monday morning
• Systematically investigate a number that doesn't match
• Research a question with cited sources

You call a skill by typing /skill-name or just asking in plain words.

Part A — Install the gstack skill suite (one-time, ~10 minutes)

Prompt 1 — Install everything

Paste this whole block into Claude Code, press Enter, and answer the yes/no permission questions as they appear:

I need to install the gstack skill suite on this computer. Please:

1. Check if Git is installed. If not, install it:
   - Windows: winget install Git.Git
   - macOS: brew install git
2. Check if Bun is installed. If not, install it (npm install -g bun).
3. Clone the gstack repository into my user skills folder:
   git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack
4. Run its setup script with the default options:
   cd ~/.claude/skills/gstack && ./setup
   (host: claude — pick defaults if it asks anything)
5. Add a short "gstack" section to my CLAUDE.md so the skills are documented.

Explain each step briefly before you run it, and stop to ask me if anything
fails or needs a decision.

What will happen: Claude checks/installs Git and Bun (one-time), clones the gstack suite, runs its setup script, and wires it up. It will ask you to approve commands — that's normal, read the one-line explanation and approve.

Prompt 2 — Install the karpathy-guidelines skill

What it is: a set of behaviour guidelines distilled from Andrej Karpathy's (co-founder of OpenAI) observations of where AI assistants go wrong. Four principles:

1. Think before coding — surface assumptions and ask questions instead of guessing
2. Simplicity first — do the minimum that solves the problem, nothing speculative
3. Surgical changes — touch only what you must; match the existing style
4. Goal-driven execution — define what "done" looks like, then verify it's actually done

Why it matters for finance: it makes the AI ask before it guesses — exactly what you want when it's handling numbers. Most useful when you ask Claude to write or change scripts (Excel/Python).

Install the karpathy-guidelines skill from the plugin marketplace:
1. Add the marketplace:  /plugin marketplace add forrestchang/andrej-karpathy-skills
2. Install the plugin:   /plugin install andrej-karpathy-skills@karpathy-skills
3. Then summarise the four principles in plain English and give me one example
   of how they change the AI's behaviour.

(You can also type those two /plugin lines yourself, one at a time.)

Already built in — no install needed

Some useful skills ship with Claude Code itself. The one you'll use most:

• /deep-research — a deep research harness: runs multiple web searches, fetches and cross-checks sources, and produces a cited report. Use it for market research, competitor checks, regulatory questions: /deep-research — "What are typical food & beverage gross margins in Singapore in 2026?"
• /verify — confirms a change actually does what it was supposed to (handy after the AI edits a script for you)

Prompt 3 — Discover more skills anytime

Search the web for the most useful Claude Code skills and plugins for finance
and accounting work. Suggest the top 5, explain each in one line, and install
the ones you recommend — ask me first before adding any new marketplace.

Prompt 4 — Check what you've got

List all the skills I currently have installed, grouped by category, and tell me
which 5 are most useful for finance work. Do not install anything.

Part B — Set up your finance workspace (CLAUDE.md)

CLAUDE.md is a briefing file Claude reads at the start of every session. It's how Claude always remembers who you are, how you work, and your team's rules — no matter when you come back to the folder.

Prompt 5 — Create your workspace brief

Create a file called CLAUDE.md in this folder with exactly this content:

---

# Finance Team Workspace — Otter Group

You are working with the finance team of Otter Group, a Singapore F&B operator
(brands: Ji De Chi, Otter Pizza).

- Our work: bank reconciliations, GST (9%) return preparation, payroll review,
  AR aging, management accounts, and the consolidated financial statements
  (parent + 3 subsidiaries, prepared from Excel financials).
- Rule: The AI drafts — the finance team reviews, verifies, and approves before
  anything is shared or published. Never skip the review step.
- Never paste financial data into free/public AI tools. Keep everything in this
  private workspace and private repositories.
- Save all work in dated task folders: YYYY-MM-DD task-name/, each with a README.md.
- Currency: SGD. Timezone: SGT (UTC+8).
- GST rate is 9% (since Jan 2024). When in doubt, ask a team member before proceeding.

---

Then tell me in 3 bullet points what this file does for me.

Part C — Create your first agent (the CFO Reviewer)

An agent is a specialised assistant inside Claude Code that you can call on at any time. This one is your safety net: it double-checks the AI's financial work before you finalise it.

Prompt 6 — Create the cfo-reviewer agent

Create a subagent called "cfo-reviewer" and save it as .claude/agents/cfo-reviewer.md.

Definition:
- Role: A meticulous finance reviewer that double-checks AI-generated financial
  work before it is finalised.
- What it does: verifies arithmetic and totals, cross-checks figures against the
  source data files, flags implausible numbers (e.g. expenses jumping 200%
  month-on-month), confirms GST is applied at 9%, and checks that S$ amounts
  match the supporting files.
- Style: short numbered findings with severity tags (High / Medium / Low).
  It never changes any numbers itself — it only reports.
- Output: ends with a recommendation: Approve / Revise / Investigate.

Then show me one example sentence of how to invoke it.

How to use it: at the end of any task, paste:

Run cfo-reviewer on my draft report before I finalise it.

Type /agents at any time to see it listed.

Part D — Skills you'll actually use, day to day

Example prompts (copy-paste)

• PDF: "Turn my report into a clean PDF." (then send the file wherever you need)
• Investigate: /investigate — "The July bank statement balance (S$ 182,430) doesn't match the GL balance (S$ 185,120). Find the root cause."
• Research: /deep-research — "What are typical food & beverage gross margins in Singapore in 2026?"
• Hand-off: /context-save on Friday → /context-restore on Monday

If something goes wrong

Next step: GitHub Collaboration → — create your account and join the team's private repository.