AI LLM for Coding

My Current Setup

• OpenCode with free default model
• Code formating via Ruff (python) and biome (node)
• 6 lines of Cavenman in global user home AGENTS.md
• 1 section of Ponytail in global user home AGENTS.md
• Context7 as MCP server for OpenCode
• RTK for token reduction

Recently, I moved all of it into a docker container to prevent LLMs from installing packages on my computer and having full access to all my files. see https://github.com/entorb/devbox

Code checks

See

• Code Checks
• Korrekturleser Repo

Tips

• E2E tests (e.g. Cypress/Playwrigth) are very important
• Use tools for code formatting, linting, type-hints, and dead code
• Configure these tools to produce minimum output to reduce token consumption. e.g. for
  • Prettier: --log-level silent
  • Cypress: cypress run --e2e --quiet
• For Refactoring/renaming: better do it manually, via VSCode, that is a lot faster

Caveman speech

less words = less tokens

Accoring to [1] Caveman speech can be can be reduced to 6 hard coded lines in (global) AGENTS.md:

Respond like smart caveman. Cut all filler, keep technical substance.
- Drop articles (a, an, the), filler (just, really, basically, actually).
- Drop pleasantries (sure, certainly, happy to).
- No hedging. Fragments fine. Short synonyms.
- Technical terms stay exact. Code blocks unchanged.
- Pattern: [thing] [action] [reason]. [next step].

Original Repo: [2]

npx skills add JuliusBrussee/caveman -a github-copilot

It has an option to install to local home dir instead of to project repo.

Ponytail, lazy senior dev mode

add contents of [3] to (global) AGENTS.md

llmfit: Right-sizes LLM models to your system's RAM, CPU, and GPU

https://www.llmfit.org offers a tool that suggests local LLMs that run on your hardware.

Context7

https://context7.com offers an MCP server of latest spec to many languages, enabling the LLM to support latest versions and best practice.

npx ctx7 setup

In prompt write:

use context7. do xxx

RTK - Rust Token Killer

https://www.rtk-ai.app reduces shell command output

brew install rtk
brew upgrade rtk
rtk init --global
rtk init --global --opencode
rtk init --global --gemini
rtk init --global --agent vibe

Codeburn (not tested by me)

https://github.com/getagentseal/codeburn
npm install -g codeburn
# or
brew install codeburn

Coding Agents

OpenCode

https://opencode.ai

brew install anomalyco/tap/opencode

# add context7 MCP
npx ctx7 setup --opencode
 
# initialize Opencode and create/update/improve/compress AGENTS.md file
 /init
 
# update
brew update
brew upgrade anomalyco/tap/opencode

# update models
opencode models --refresh

Config ~/.config/opencode/opencode.json

{
  "$schema": "https://opencode.ai/config.json",
  "share": "disabled",
  "formatter": true
}

Addons

# add Caveman and Ponytail instructions to global
# ~/.config/opencode/AGENTS.md
# see below

# add Ruff Token Killer
rtk init --global --opencode

Model ranking

Most prominent models: https://opencode.ai/data/

SQLite DB

~/.local/share/opencode/opencode.db

SELECT title, 
DATETIME(time_created/1000, 'unixepoch') AS date, -- UTC 
round(cost,2) AS cost, 
tokens_input, tokens_output, tokens_reasoning, tokens_cache_read, tokens_cache_write,
json_extract(model, '$.id') AS 'model-id', 
json_extract(model, '$.providerID') AS 'model-providerID'
FROM SESSION s
WHERE 1=1
--AND project_id = 'fe41f10cbae7d51e25088b6ed35c168da9d3a4f6' -- flashcards
AND directory LIKE '%/flashcards' 
AND model LIKE '%deepseek%'
ORDER BY time_created DESC

Mistral Devstral Vibe CLI

# install
uv tool install mistral-vibe
# setup --> choose automatic
vibe --setup
# run via
vibe

to set Cavemen style global, create ~/.vibe/AGENTS.md (or link to existing file)

connect to context7 MCP server ~/.vibe/config.toml (copy from edit mode)

mcp_servers
name = "context7"
transport = "http"
url = "https://mcp.context7.com/mcp"

[mcp_servers.headers]
CONTEXT7_API_KEY = "XXX"

Claude

Web

Website: [4]

• Pros: Daily reset free quota
• Cons: Need to manually upload single files

Claude Console

Claude Console via API and VISA Pre-Paid: [5]

• Pros: Fully integrated in codebase
• Cons: Costs money (e.g., 5€ are easily spent)

npm install -g @anthropic-ai/claude-code
cd myProject
claude

Uninstall

npm uninstall -g @anthropic-ai/claude-code

GitHub Copilot

global config file:

.copilot\copilot-instructions.md

Google Gemini

• Pros: Free to use with Google account

Create your API key at [6]

Install

npm install -g @google/gemini-cli@latest
gemini
# or just run
npx https://github.com/google-gemini/gemini-cli

Uninstall

npm uninstall -g @google/gemini-cli

Usage

cd /tmp
export GEMINI_API_KEY="xxx"
git clone https://github.com/entorb/meeting-meter
cd myProject
gemini

To switch model from pro to flash:

export GEMINI_MODEL="gemini-2.5-flash"

Secret alternative: .env file

GEMINI_API_KEY=xxx

Google Antigravity

https://antigravity.google/download