DEV Community: borhen saidi

AgentForge – 28 production grade skills that make AI agents ship reliable code

borhen saidi — Thu, 11 Jun 2026 23:39:47 +0000

AgentForge is a control system for AI coding agents. Instead of hoping your agent "writes good code," you give it structured engineering workflows — the same ones senior engineers actually follow.

What it is:

28 skills covering the full lifecycle: define, plan, build, verify, review, ship
Each skill is a structured process with steps, anti-rationalization tables, red flags, and verification gates
Works with Claude Code, OpenCode, Gemini CLI, Copilot, Cursor, Windsurf
Why it exists: I've watched agents confidently ship broken auth, skip error handling, and deploy on Friday afternoon because nobody told them not to. These aren't prompt engineering tricks — they're encoded workflows. The test-driven-development skill makes the agent write a failing test before touching implementation. The shipping-and-launch skill forces a rollback plan. The doubt-driven-development skill makes the agent challenge its own assumptions before continuing.

What's different:

Anti-rationalization tables: Every skill lists the excuses engineers use to skip best practices ("CI is too slow" → "Optimize the pipeline, don't skip it")
Verification gates: Checklists the agent must complete before proceeding
Cross-skill consistency: A quality gate ensures all 28 skills follow the same anatomy and reference each other correctly
The catch: This isn't magic. The agent still needs to follow the skill. But when it does, the output is consistently better — fewer "it works on my machine" patches, more actually-shippable code.

Repo: [https://github.com/borhen68/SkillEngine]

I'd love feedback from anyone running AI agents in production. What's the most expensive mistake your agent has made?

TokenTamer

borhen saidi — Wed, 10 Jun 2026 14:27:32 +0000

🚀 I just open-sourced something I built out of frustration — and it's saving me real money every day.https://github.com/borhen68/TokenTamer
If you use AI coding agents (Cursor, Claude Code, Aider…), you've probably noticed how fast your API bills stack up. The problem? Your agent keeps re-sending the same files — full source code — every single turn. You're paying for tokens you already paid for.

So I built TokenTamer — a drop-in proxy that sits between your agent and the LLM API and quietly cuts token usage by 50–80%. No config changes. No code changes. Just point your API base URL at it and go. ✅

Here's what it does under the hood:

🧠 AST-based compression — strips function bodies from background files, keeps only signatures. The LLM knows what exists, without reading every line.
🔧 Tool-aware compression — skeletonizes stale file reads, keeps the latest one intact.
💾 Prompt cache hijacking — injects Anthropic cache breakpoints so long Claude Code sessions hit the cache instead of paying full price (~73% off on long runs).
💰 Real-time dashboard — watch tokens saved and dollars saved live in your terminal.

It's MIT licensed, works with OpenAI + Anthropic APIs, and takes 5 minutes to set up.

👇 Star it, try it, break it — and tell me what you think.

TokenTamer A proxy that reduces LLM token usage through context compression

borhen saidi — Tue, 09 Jun 2026 09:19:49 +0000

I built TokenTamer, an open-source proxy that sits between AI coding assistants and LLM APIs.

The goal is to reduce token consumption before requests reach the model by applying techniques such as:

Context deduplication
Conversation compression
Intelligent summarization
Smart context filtering

I originally built it after noticing that coding agents often resend large amounts of repeated context, leading to unnecessary token usage and higher costs.

TokenTamer is designed to be lightweight and easy to place in front of existing workflows.

I'd love feedback on the architecture, compression strategies, and potential use cases.(https://github.com/borhen68/TokenTamer)