prompt-eng-toolkit

prompt-eng-toolkit

Defense-first workflow for designing and compressing production single-turn LLM system prompts, packaged as a Claude Code plugin.

Two skills, one shared corpus of references / scripts / fixtures:

• prompt-create — design a new system prompt with injection defenses baked in
• prompt-optimize — compress an existing prompt while preserving (often improving) adherence

What this is

An opinionated, vendor-agnostic, defense-first workflow for production single-turn task prompts (refinement, classification, extraction, formatting). Built from the consensus of Google Gemini, Anthropic Claude, and OpenAI's official 2026 prompting guides, plus a real-world failure-mode catalog from production deployment.

Every workflow step is gated:

• API key present → iterative loop against real provider APIs (validate behavior before writing to source)
• API key absent → theory-only design pass with explicit "untested" disclaimer

What this isn't

Avoid stack-of-shovels syndrome — use the right tool:

Use case	Better tool
Production-grade eval pipeline / regression suite	promptfoo (20k+ stars)
Auto-optimize via ML / DSPy / genetic algos	AutoPrompt / Promptomatix
Catalog of named frameworks (CO-STAR / RISEN / TIDD-EC …)	prompt-architect
Web/Desktop UI tool, multi-tab playground	linshenkx/prompt-optimizer (28k+ stars)
Improve user-prompts to Claude Code itself	claude-code-prompt-improver
User-facing prompt rewriter for 30+ AI tools	prompt-master

Where this fits the stack: use this to author the prompt → optionally pipe the resulting prompt into promptfoo for production-grade regression eval.

What's inside

plugins/prompt-eng-toolkit/
├── skills/
│   ├── prompt-create/SKILL.md     # design workflow
│   └── prompt-optimize/SKILL.md   # compression + hardening workflow
└── shared/
    ├── references/
    │   ├── provider-guidance.md           # 3-provider official-doc consensus
    │   ├── universal-principles.md        # 10 cross-vendor rules + checklist
    │   ├── optimization-playbook.md       # compression patterns + anti-patterns
    │   ├── failure-modes-and-defenses.md  # 7 attack categories + 5-layer defense
    │   └── v4-template.md                 # block-by-block annotated template
    ├── scripts/
    │   └── count_tokens.py                # multi-provider token counter (Gemini / Anthropic / OpenAI)
    └── fixtures/
        └── attack-tests-template.yaml     # 7-category injection-test starter

Install

The plugin works in three agentic CLIs. Pick the section for your tool.

Claude Code (CLI, no TUI required)

Run from any shell:

claude plugin marketplace add RxChi1d/prompt-eng-toolkit
claude plugin install prompt-eng-toolkit@prompt-eng-toolkit

The same commands also work as slash commands inside a Claude Code session (/plugin marketplace add ... / /plugin install ...) if you prefer.

Codex CLI (CLI add, then TUI install)

Codex's CLI only handles marketplace registration; the actual install is a TUI action.

# 1. Register the marketplace from your shell
codex plugin marketplace add RxChi1d/prompt-eng-toolkit

# 2. Open the TUI and install the plugin
codex
# inside Codex: type /plugins → switch to the prompt-eng-toolkit tab
#               (or search "prompt" in All Plugins) → press Enter to install

If the marketplace was added before this commit and the plugin doesn't appear, run codex plugin marketplace remove prompt-eng-toolkit then add again to force a fresh parse.

OpenCode (agent auto-install)

OpenCode has no marketplace mechanism. Easiest path: ask the OpenCode agent to install it for you. In your OpenCode session, paste:

Please follow the install instructions at https://raw.githubusercontent.com/RxChi1d/prompt-eng-toolkit/main/install/opencode.md

The agent fetches the install guide, asks whether you want per-project or global scope, clones the repo to ~/.local/share/prompt-eng-toolkit, creates the required symlinks, verifies the install, and reports back. Re-running the same prompt later updates the install (idempotent).

Manual install (if you prefer not to delegate):

# Clone once
git clone https://github.com/RxChi1d/prompt-eng-toolkit ~/.local/share/prompt-eng-toolkit

# Pick scope
TARGET="$HOME/.config/opencode"        # global
# TARGET="$(pwd)/.opencode"            # per-project