_SH Log's
Back to Root
EST: 5 min read

The Tools I Built to Make AI Coding Agents 10x Better

I built a toolchain — skills, memory, free-LLM routing — that makes AI coding agents dramatically more capable. Here's every tool and what problem it solves.

#ai-agents#claude-code#open-source#productivity

AI coding agents are powerful out of the box. They become transformative with the right toolchain around them. Over the past year building six products with AI agents as my primary collaborators, I've built and refined a set of tools that make agents dramatically more capable. Here's what I built and why.

The problem: agents are only as good as their context

A coding agent without context makes generic decisions. Add project context — architecture, conventions, gotchas — and it makes the right decisions for your specific codebase.

The tools I've built all solve one version of this problem: giving agents the context they need, when they need it, without requiring me to re-explain it every session.

Tool 1: softco — Full-product lifecycle agent

Problem: Coordinating across PM, architecture, implementation, and deployment requires constant context switching.

Solution: softco packages all of those roles into a single agent skill. One command goes from feature description to deployed code:

/softco build "add webhook delivery retry with exponential backoff to the letx compile queue"

softco produces: user story → architecture note → implementation plan → Go code → tests → CI/CD delta. It knows the whole product lifecycle and won't give you code without tests or deploy scripts.

Impact: Feature development time down ~60%.

Tool 2: common-knowledge — Git-backed agent memory

Problem: Every session starts cold. Agents re-discover gotchas, re-learn conventions, re-read architecture docs.

Solution: A Git-backed knowledge store with a CLI that agents can save to and load from:

/ck save "pgBouncer session mode breaks prepared statements — use transaction mode"
/ck load letx  # loads all letx knowledge into agent context at session start

Saves are Git commits. History is permanent. Searchable. Human-readable Markdown.

Impact: ~30-minute context rebuild time at session start → ~2 minutes.

Tool 3: skill-builder — Build skills from descriptions

Problem: Creating a new agent skill requires knowing the skill format, writing reference docs, and structuring the SKILL.md correctly.

Solution: Describe the skill; skill-builder writes it:

/skill-builder "A skill for Go gRPC services using buf toolchain, 
proto conventions, error handling, and connection pooling patterns"

Output: complete SKILL.md + reference files + pre-commit hook for proto linting.

Impact: New skill creation: 3 hours → 20 minutes.

Tool 4: openrouter-free — Zero-cost LLM routing

Problem: Using Claude API for every task is expensive. Many tasks (summarization, classification, simple code) work fine with free open-source models.

Solution: Routes to the best free model available on OpenRouter for the task type:

/openrouter-free summarize this PR diff
/openrouter-free classify these user feedback entries
/openrouter-free generate test data for this schema

Uses: Llama 3 70B, Mistral 7B, Gemma 2 9B — all free tier on OpenRouter.

Impact: ~40% reduction in Claude API spend.

Tool 5: seo-master — SEO + GEO + AEO

Problem: SEO implementation requires knowing the full 2026 playbook — structured data, AI engine indexes, answer-first writing, Search Console setup.

Solution: A comprehensive skill that knows the full playbook and is framework-aware (Next.js App Router, React SPA, Astro, Remix):

/seo-master blog.shihub.online

Produces: audit findings + implementation priority list + JSON-LD snippets + title rewrites + meta description rewrites.

Impact: Full SEO audit + implementation: 1 day → 2 hours.

Tool 6: frontend-design — Anti-AI-slop UI

Problem: Agent-generated UI tends toward a generic "AI aesthetic" — rounded cards, gradient buttons, floating action buttons. It's forgettable.

Solution: Strict guidelines for production-grade UI with specific references to real design systems, typography rules, and component patterns:

/frontend-design build a data table component for letx document history

Impact: First-pass UI quality significantly higher; fewer design revision rounds.

Tool 7: hermes-agent-aws — Always-on agent on AWS

Problem: Agents are reactive. Many tasks (monitoring, scheduled jobs, webhook handlers) need a persistent process.

Solution: Terraform module + Go agent binary that deploys an always-on AI agent to AWS for ~$17/month. Listens to webhooks, runs cron tasks, monitors GitHub/Slack/CloudWatch.

terraform apply  # deploys in 3 minutes

The compound effect

These tools compound. A session looks like:

/ck load letx                    # load project context (2 min)
/softco build [feature]          # implement with full context (30 min)
/openrouter-free review-tests    # validate tests with free LLM (5 min)
/code-review                     # final review before merge (10 min)
/ck save [anything discovered]   # persist new learnings (1 min)

Context → plan → implement → review → persist. Each tool feeds the next.

FAQ

What's the most impactful tool for solo AI-assisted development? common-knowledge — eliminating the session context rebuild is the highest-leverage change. Second: softco for closing the gap between "describe a feature" and "deployed code."

Are all these tools open source? Yes — all are open source on GitHub at shihabshahrier under MIT license.

Do these tools only work with Claude Code? The skills (SKILL.md files) are Claude Code-specific. common-knowledge and hermes-agent-aws are agent-agnostic. openrouter-free can be used with any agent that supports tool use.

How much do these tools cost to run? The skills themselves cost nothing. common-knowledge uses local Git (free). hermes-agent-aws costs ~$17/month on AWS. openrouter-free reduces costs by routing to free LLMs.


Written by Shihab Shahriar Antor — AI Engineer & Founder of Shahriar Labs. See also: My AI Agent Skills Stack · How I Use Claude Code to Ship Faster.