Installation

Installation — ইনস্টলেশন

TakaBooks is one source tree built into one bundle per platform. This page covers getting a bundle, then installing it in each LLM, then running the Python engine on your own machine. Where a screenshot would help and cannot be shown, the screen is described in words.

Prerequisites

For	You need
Any chat platform (Claude, ChatGPT, Gemini, Kimi, DeepSeek, …)	An account on that platform. Nothing to install.
The numbers	Python 3.11 or newer on your own computer. The engine is standard library only — there is nothing to pip install, and nothing is downloaded at run time. Check with python3 --version (Windows: py -3 --version). Get Python from https://www.python.org/downloads/ if needed.
Cursor, Copilot, Codex, Windsurf	The repository checked out; they read AGENTS.md at its root.
The npm installer (optional)	Node 18 or newer, and a GitHub token (explained below).

A chat platform without the engine can still classify entries, explain rules and prepare the exact commands for you to run; it cannot compute a figure. The core instruction tells it to say so and to ask for the engine's output back rather than do the arithmetic itself.

Step 1 — get a bundle

Option A: download from a release (recommended)

Go to https://github.com/bemoshiur/TakaBooks/releases/latest. Each release carries:

Asset	For
bd-bookkeeping-tax-skill-<version>.zip	Claude (claude.ai upload, Claude Code, Claude API) and every other Agent Skills host (Codex CLI, ChatGPT Skills, Gemini CLI, Cursor, VS Code)
takabooks-chatgpt-<version>.zip	ChatGPT Custom GPT or Project: instructions.md, knowledge/, README-install.md
takabooks-gemini-<version>.zip	Gemini Gem: gem-instructions.md, knowledge/, README-install.md
takabooks-complete-<version>.md	Any other LLM — Kimi, DeepSeek, Llama, Mistral, local models: one paste-anywhere file
AGENTS-<version>.md	Coding agents: Cursor, Copilot, Codex, Windsurf, Gemini CLI
SHA256SUMS.txt	Checksums of every asset

Verify a download before you trust it:

sha256sum -c SHA256SUMS.txt        # macOS: shasum -a 256 -c SHA256SUMS.txt

Option B: build from source

git clone https://github.com/bemoshiur/TakaBooks.git
cd TakaBooks
python3 build/build.py             # every bundle into dist/, plus AGENTS.md at the root

python3 build/build.py --check verifies every platform limit without writing; --target chatgpt (or claude-skill, gemini, universal, agents-md) builds one. The build is deterministic — the same src/ bytes produce byte-identical dist/ bytes.

Option C: the npm installer (GitHub Packages)

The package @bemoshiur/takabooks copies a built bundle to where a platform expects it:

npx @bemoshiur/takabooks install claude        # → ~/.claude/skills/bd-bookkeeping-tax/
npx @bemoshiur/takabooks install chatgpt       # → ./takabooks-chatgpt/
npx @bemoshiur/takabooks install gemini        # → ./takabooks-gemini/
npx @bemoshiur/takabooks install universal     # → ./takabooks-complete.md
npx @bemoshiur/takabooks install agents        # → ./AGENTS.md
npx @bemoshiur/takabooks list

Options: --dest <path> (install elsewhere), --dist <path> (read bundles from a checkout's dist/), --force / -f (replace an existing install — it deletes the destination first), --dry-run / -n, --json, --quiet, --no-bangla, --help, --version. It never clobbers an existing install without --force, and exits non-zero on any failure.

The catch: the package is published to GitHub Packages, not npmjs.com, and GitHub Packages requires authentication even to install a public package. On a machine that has never authenticated, npx @bemoshiur/takabooks returns 401. To use it, put two lines in your ~/.npmrc:

@bemoshiur:registry=https://npm.pkg.github.com
//npm.pkg.github.com/:_authToken=${NODE_AUTH_TOKEN}

and export NODE_AUTH_TOKEN as a GitHub classic personal access token with the read:packages scope (fine-grained tokens do not work with the npm registry). If that is more friction than you want, use Option A — the release assets are the same files.

From a checkout, without any registry: node bin/takabooks.mjs install claude --dist ./dist.

Step 2 — install it where your LLM lives

Claude Code

The skill folder goes under skills/, personal or per project:

# personal — available in every project
unzip bd-bookkeeping-tax-skill-<version>.zip -d ~/.claude/skills/

# project — checked in with the repository
unzip bd-bookkeeping-tax-skill-<version>.zip -d .claude/skills/

Either way you end up with …/skills/bd-bookkeeping-tax/SKILL.md plus references/, scripts/, data/, templates/. The zip has the bd-bookkeeping-tax/ folder at its root, which is why -d …/skills/ is the right target. A personal skill overrides a project skill of the same name; an enterprise-managed one overrides both.

Then, in a project, start Claude Code and ask a Bangladeshi bookkeeping or tax question in Bangla, Banglish or English — the skill is loaded by its description when the question matches. The scripts run through Claude Code's shell; only their output enters the conversation, which is exactly the design.

Claude Code reads CLAUDE.md, not AGENTS.md. To give it the repository's coding-agent instructions too, create a two-line CLAUDE.md at the repository root:

@AGENTS.md

claude.ai (the web and desktop apps)

Screen by screen:

1. Open Settings (your initials or avatar, bottom-left on the web).
2. Go to Features (the section that lists optional capabilities). Make sure code execution is enabled — skills need it.
3. Find Skills and choose Upload skill (a button that opens a file picker).
4. Pick bd-bookkeeping-tax-skill-<version>.zip. The folder must be at the root of the zip; the release asset already is.
5. The skill appears in the list by name (bd-bookkeeping-tax) with its description.

Skills uploaded on claude.ai are per-user: there is no organisation-wide management, and they do not sync to Claude Code or to the API. Install in each place separately.

Claude API

Upload the same zip to the Skills endpoint (/v1/skills). bd-bookkeeping-tax/SKILL.md sits at the top of a single enclosing folder, which is what the endpoint expects. Versions are complete snapshots, not deltas — re-upload the whole file set every time. The API sandbox has no network access and cannot install packages; TakaBooks needs neither.

ChatGPT — Custom GPT (Free, Plus, Pro)

Unzip takabooks-chatgpt-<version>.zip. Then, screen by screen:

1. In ChatGPT, open Explore GPTs (left sidebar) and press Create (top right).
2. Switch from the Create tab to the Configure tab — a form with Name, Description, Instructions, Conversation starters, Knowledge, Capabilities.
3. Instructions: paste the entire contents of instructions.md. The field accepts 8,000 characters; the file is built to fit and the build fails rather than produce an over-long one. If the field shows a character counter, it should be under the limit with a little headroom.
4. Knowledge: press Upload files and select every file in knowledge/. A Custom GPT holds up to 20 knowledge files; the bundle ships fewer. Keep the file names — the instruction refers to them by name.
5. Capabilities: Code Interpreter is useful (the GPT can run the engine on files you upload in a chat), not required.
6. Save (top right) → choose who can use it → Create.

The rates file is shipped as rates-AY<year>.md — Markdown with the TOML inside a fenced block — because .toml is not on any published list of accepted upload types. Its values are byte-identical to src/data/.

The engine scripts are not in the knowledge set: the GPT explains and classifies; the numbers come from running the scripts locally, as knowledge/engine-usage.md describes.

ChatGPT — Projects

Open a Project → Project instructions → paste instructions.md. Then Add files → the knowledge/ files. A project holds fewer files than a Custom GPT (as few as five on the Free plan), so on a small plan add, in this order: the rates file, 30-tax-overview.md, then the reference you need for the task at hand.

ChatGPT Skills and Codex CLI (Business, Enterprise, Edu)

Both read the open Agent Skills format, so use the Claude Skill bundle, not the ChatGPT one:

• Codex CLI: unzip so that ~/.codex/skills/bd-bookkeeping-tax/SKILL.md exists.
• ChatGPT Skills: upload the same zip under Skills → Upload from your computer. Skills in ChatGPT are limited to Business, Enterprise, Healthcare and Edu plans, and an administrator may have to enable them first.

Gemini — Gems

Unzip takabooks-gemini-<version>.zip. Then:

1. Open Gemini → Gems (left sidebar) → New Gem (or Gem manager → New Gem).
2. Give it a name. In Instructions, paste the whole of gem-instructions.md. It is short on purpose — Gems drift away from their knowledge files under long instructions — and written in Google's Persona / Task / Context / Format shape.
3. Under Knowledge, press Add files (or Upload from computer / Add from Drive) and add every file in knowledge/. A Gem supports up to 10 source documents, which is why the bundle's files are merged; it ships exactly ten. The merge is at whole-file level, never a truncation, so every source URL and verified flag survives.
4. Save.

Tip for keeping rates current: Gemini reads the latest version of a Google Doc from Drive automatically, but re-uploads other file types by hand. If you publish the rates knowledge file as a Google Doc and add it from Drive, a rates update reaches every Gem that uses it without a re-upload.

Kimi, DeepSeek, Llama, Mistral, any other LLM — the universal file

takabooks-complete-<version>.md is one self-contained Markdown file: core instruction, every reference, the rates table with the TOML in a fenced block, the engine commands with their output shapes, and the glossary — with every cross-reference an in-document anchor. Nothing outside the file exists.

• Chat web apps (Kimi, DeepSeek, …): paste the whole file as the first message, or upload it as a document if the app supports uploads. It opens with a STOP — READ THIS FIRST block that the model sees before anything else.
• Ollama: put the file inside SYSTEM """…""" in a Modelfile, or attach it as context in whichever front end you use.
• LM Studio / Open WebUI: paste it into the system-prompt field and save it as a preset.

The file tells the model what to do when it cannot run Python — give the classification, the account codes, the tax_tag and the rule in words, print the exact command, and ask for the output back. Small local models (7–13B) follow short instructions better than long ones; if yours drifts, use the universal file as an uploaded document and a shorter system prompt.

Cursor, GitHub Copilot, Codex, Windsurf, Gemini CLI — AGENTS.md

AGENTS.md at the repository root is the cross-agent standard those tools read; the nearest AGENTS.md up the directory tree wins. It is generated by the build and carries the non-negotiables (never invent a rate; integer paisa; stdlib only; src/ is the truth; dist/ is generated), the layout, the commands, and the content rules. Nothing to install: check out the repository and the tool finds it. For Claude Code, add the two-line CLAUDE.md shown above.

Step 3 — the engine on your own machine

Whichever platform you use, the numbers come from Python running locally.

git clone https://github.com/bemoshiur/TakaBooks.git      # or unzip the skill bundle
cd TakaBooks
python3 src/engine/validate.py --help                     # prints usage → Python is fine

In the Claude Skill bundle the same scripts are at bd-bookkeeping-tax/scripts/ and the rates file at bd-bookkeeping-tax/data/; scripts/USAGE.md lists every command.

Windows notes: use py -3 src\engine\validate.py --help if python3 is not on your PATH; run in Windows Terminal or PowerShell so Bangla renders; the engine itself is unaffected by what the console can display.

Then follow Getting Started.

Verifying an install

Platform	How to tell it worked
Claude Code	ls ~/.claude/skills/bd-bookkeeping-tax/SKILL.md exists; a Bangla bookkeeping question makes Claude read the skill and run scripts/…
claude.ai	The skill is listed under Settings → Features → Skills with the name bd-bookkeeping-tax
ChatGPT	The Configure tab shows the instructions with no truncation warning and every knowledge file listed
Gemini	The Gem shows ten knowledge documents
Universal	The model's first reply to "what are your rules?" mentions that it never does arithmetic and names the rates file
Engine	python3 src/engine/validate.py --help prints usage and exits 0

Updating

Bundles are versioned with the release tag. To update, download the new assets and repeat the platform steps — every platform above replaces rather than merges, so re-upload the whole set. Custom skills do not sync across surfaces; update each one. Your books/ directory is yours and is never touched by an update.

See also

• Getting Started · Engine Reference
• Troubleshooting — Python version errors, the npm 401, Bangla rendering
• Disclaimer