Heimdall — curated system prompt library

Norse mythology: Heimdall watches from the gateway of the gods. Here: a small Python library + CLI for managing, searching, and analyzing AI system prompts. Ships with 10 opinionated starter prompts.

Install

pip install heimdall-prompts

CLI

heimdall list                          # show all prompts (built-in + local)
heimdall list --category coding
heimdall list --tag tdd

heimdall show code-review-thorough     # metadata + content
heimdall show code-review-thorough --content-only   # just the prompt body

heimdall search TDD                    # fuzzy search

heimdall analyze code-review-thorough  # quality stats (density, structure, readability)

heimdall categories                    # list categories with counts

Library

from heimdall import SystemPrompt, PromptLibrary, PromptAnalyzer

lib = PromptLibrary()
lib.add(SystemPrompt(
    name="my-reviewer",
    content="You are a senior engineer...",
    category="coding",
    tags=["review"],
))

# Search, filter, analyze
results = lib.search("review")
print(PromptAnalyzer().analyze(lib.get("my-reviewer")))

Starter prompts

Ships with 10 opinionated system prompts:

Category	Prompt
coding	code-review-thorough, bug-repro-minimizer, tdd-strict
writing	api-doc-generator, pr-description-writer, terse-explainer
analysis	research-synthesizer
thinking	rubber-duck, devil-advocate
career	interview-star-coach

Each is deliberate about what not to include (no hedging, no history-dumping, no empty praise). Read them in src/heimdall/data/starter_prompts.json.

Adding your own prompts

Drop JSON files in ~/.heimdall/ (or set HEIMDALL_DIR=/path/to/prompts). Each file can be a single prompt object or an array. Schema:

[
  {
    "name": "my-prompt",
    "content": "You are...",
    "category": "coding",
    "tags": ["example"],
    "model_target": "claude"
  }
]

heimdall list picks them up automatically.

What Heimdall isn't

• Not a prompt-engineering framework. No chains, no templates-of-templates.
• Not an LLM wrapper. The prompts are text — you send them wherever.
• Not opinionated about which LLM. model_target is metadata, not enforcement.

Development

git clone https://github.com/MukundaKatta/heimdall.git
cd heimdall
python -m venv .venv && source .venv/bin/activate
pip install -e '.[dev]'
pytest
heimdall list

License

MIT