Prompt Engineering Toolkit

What's in this pack

This pack assembles six high-signal prompt-engineering assets and pairs them with a Claude Code subagent that actually applies them. The mix is deliberate: two encyclopedic references, two opinionated framework decks, and two operational tools you can drop into your editor.

The collection is opinionated about the order of operations: read the awesome list to map the territory, copy a framework that matches your task, then run the subagent on your draft and iterate.

Why a "toolkit" rather than another listicle

Search results for "prompt engineering" have collapsed into the same five tips on every page. This pack solves a different problem: once you know the basics, what do you reach for to get better?

The answer turns out to be three things in tension:

• Breadth — see how prompts vary across domains, models, and provider quirks. The Awesome list and Cookbook cover this.
• Structure — pick a frame so your prompts are auditable, comparable, and reusable. The 27 frameworks and Prompt Master deck cover this.
• Iteration — get from draft N to draft N+1 quickly, with a rationale. The Claude Code subagent covers this.

Owning all three at once is what compounds. Picking one or two leaves you re-discovering the same mistakes for months.

Install in one command

# Install the entire pack
tokrepo install pack/prompt-engineering-toolkit

# Or pick the subagent alone
tokrepo install prompt-engineer

The subagent activates inside Claude Code on a @prompt-engineer rewrite this prompt for clarity and falsifiability request. It uses one of the 27 frameworks (default: CRISPE) and outputs a diff plus a rationale section. The framework decks are stored under .claude/skills/prompt-engineering/ so you can reference them from any session.

Common pitfalls

• Treating frameworks as gospel. CRISPE / RACE / RTF are scaffolds, not laws. The subagent picks one based on the task; if its choice feels wrong, override with --framework=<name> rather than fighting the output.
• Skipping eval. A prompt rewrite that "feels better" might score worse on your real test set. Pair this pack with LLM Eval & Guardrails (Promptfoo / DeepEval) so every change has a quantified delta.
• Provider drift. OpenAI Cookbook examples assume OpenAI APIs. The Claude / Gemini equivalents differ in subtle ways (system message handling, tool use schemas). When porting, check the provider's own prompting doc first.
• Over-prompting. Long prompts hide bugs. If a prompt exceeds ~400 tokens, factor parts into a tool definition or retrieval call instead of cramming everything into the system message.
• No version control. Prompts are code. Commit them, diff them, code-review them. The subagent emits diffs precisely so this lifecycle works.

Common misconceptions

• "Prompt engineering is dead because models got smarter." The opposite — better models reward more structured prompts because they can follow more constraints reliably. The skill that died is trick-prompting (jailbreak phrases, magic words). Structured prompting is more valuable than ever.
• "You don't need this if you're using a framework like LangChain." Frameworks compose prompts; they don't write them. The system messages and tool descriptions inside a LangChain chain are still prompts you have to author.
• "The OpenAI Cookbook is OpenAI-specific." The patterns (function calling, structured output, evaluators) port cleanly to Claude and Gemini. The bindings differ; the approach doesn't.

Where this pack stops

The toolkit is end-to-end for single-prompt engineering. It does not cover multi-step agent design (where the prompt is just one node in a graph), retrieval-augmented prompting (where context is injected at runtime), or fine-tuning (where you change the model itself). For each of those, pair this pack with the appropriate adjacency: agent frameworks for graphs, RAG pipelines for retrieval, and fine-tuning recipes for the model layer. The prompt-engineer subagent is honest about its scope — when you ask it to "fix" something that needs RAG instead, it will tell you so explicitly rather than producing a longer prompt that papers over the gap.