llm

LLM Prompt Engineering Toolkit. Build structured prompts from role/context/task components, compare prompt variations side by side, estimate token counts, manage reusable prompt templates, chain multi-step prompts, and evaluate prompt quality with a scored breakdown. All commands run locally in bash with no API keys or network access required.

Commands

prompt — Build a Structured Prompt

Assembles a prompt from modular components: role, context, task, constraints, and output format. The --task flag is required; all others are optional.

Flags:

• --role <text> — Define the AI's persona (e.g., "senior developer")
• --context <text> — Provide background information
• --task <text> — (required) The main instruction
• --constraints <text> — Rules or limitations
• --format <text> — Desired output format

bash scripts/script.sh prompt --role "senior developer" --context "Python Flask app" --task "write unit tests"
bash scripts/script.sh prompt --task "summarize this article" --constraints "max 3 sentences" --json

compare — Compare Prompt Variations

Compare two or more prompt files side by side. Shows each variant with word/line/char/token stats, then a diff --side-by-side of the first two variants, plus a summary table.

Flags:

• --prompts <file1> <file2> [file3...] — Two or more prompt text files to compare

bash scripts/script.sh compare --prompts prompt_a.txt prompt_b.txt
bash scripts/script.sh compare --prompts v1.txt v2.txt v3.txt

tokenize — Estimate Token Count

Estimate the token count for a given text using a cl100k_base-compatible heuristic. Reports characters, words, lines, and estimated tokens.

Input methods:

• --input <text> — Inline text string
• --file <path> — Read from a file
• Pipe via stdin

bash scripts/script.sh tokenize --input "Your prompt text here"
bash scripts/script.sh tokenize --file prompt.txt
echo "some text" | bash scripts/script.sh tokenize
bash scripts/script.sh tokenize --file prompt.txt --json

template — Manage Prompt Templates

Save, list, load, and delete reusable prompt templates. Templates are stored as .txt files in ~/.llm-skill/templates/.

Actions:

• --save <name> --file <path> — Save a template from a file (or pipe via stdin)
• --list — List all saved templates with sizes
• --load <name> — Output the contents of a saved template
• --delete <name> — Remove a saved template

bash scripts/script.sh template --save my_template --file prompt.txt
bash scripts/script.sh template --list
bash scripts/script.sh template --list --json
bash scripts/script.sh template --load my_template
bash scripts/script.sh template --delete my_template
echo "Write a haiku about {{topic}}" | bash scripts/script.sh template --save haiku

chain — Multi-Step Prompt Chains

Run a sequence of prompt steps where each step's output feeds into the next via the {{previous_output}} placeholder. Steps can be specified as individual files or loaded from a JSON config.

Flags:

• --steps <file1> <file2> [...] — Ordered list of step files
• --from <config.json> — Load steps from a JSON configuration file

bash scripts/script.sh chain --steps step1.txt step2.txt step3.txt
bash scripts/script.sh chain --from chain_config.json
bash scripts/script.sh chain --steps brainstorm.txt refine.txt format.txt --json

evaluate — Score Prompt Quality

Score a prompt on four dimensions (0–100 each): Clarity, Specificity, Structure, and Completeness. Returns an overall score (0–100) and letter grade (A–F) with actionable suggestions.

Scoring heuristics:

• Clarity — Penalizes vague words ("something", "stuff"), rewards action verbs ("write", "create", "analyze") and structural markers
• Specificity — Rewards concrete numbers, quoted examples, and sufficient length
• Structure — Rewards headers, bullet lists, numbered steps, and paragraph breaks
• Completeness — Checks for role definition, output format spec, constraints, and examples

bash scripts/script.sh evaluate --input "Explain quantum computing"
bash scripts/script.sh evaluate --file my_prompt.txt
bash scripts/script.sh evaluate --file my_prompt.txt --json

help — Show Help

bash scripts/script.sh help

Global Flags

• --json — Output in JSON format (supported by prompt, tokenize, template --list, chain, and evaluate)

Data Storage

• Templates: ~/.llm-skill/templates/*.txt
• No other persistent state. All commands are stateless except template which manages saved files.

Requirements

• Bash 4+ (uses arrays, [[ ]], process substitution)
• Standard Unix utilities: wc, grep, diff, cat, basename, tr, sed, rm, mkdir
• No external dependencies, API keys, or network access required

When to Use

1. Crafting system prompts — Use prompt to build well-structured prompts from role/context/task components instead of writing them freehand.
2. A/B testing prompt variants — Use compare to see side-by-side diffs and token counts for two or more prompt versions before committing to one.
3. Estimating API costs — Use tokenize to get token estimates before sending prompts to paid LLM APIs, helping you stay within budget.
4. Building reusable prompt libraries — Use template to save, organize, and reuse your best prompts across projects.
5. Quality-checking prompts before use — Use evaluate to score your prompts on clarity, specificity, structure, and completeness, with actionable improvement suggestions.

Examples

# Build a structured prompt for code review
bash scripts/script.sh prompt \
  --role "senior code reviewer" \
  --context "React TypeScript project" \
  --task "review this pull request for bugs and performance issues" \
  --constraints "focus on security vulnerabilities" \
  --format "numbered list of findings"

# Estimate tokens for a long prompt
bash scripts/script.sh tokenize --file system_prompt.txt

# Save a template and reuse it
echo "You are a {{role}}. Your task: {{task}}" | bash scripts/script.sh template --save generic
bash scripts/script.sh template --load generic

# Evaluate prompt quality
bash scripts/script.sh evaluate --input "You are an expert Python developer. Write a function that sorts a list of dictionaries by a given key. Include type hints, docstring, and 3 unit tests."

Powered by BytesAgain | bytesagain.com | hello@bytesagain.com