PromptGate is a prompt refinement toolkit for AI agent workflows.

It turns rough user prompts into structured, actionable requests and can hand the refined prompt to registered external skills in Claude or Codex environments.

PromptGate does not bundle workflow skills. It does not own coding, research, planning, deployment, or review workflows. It only refines prompts and routes them to skills that already exist in the user's environment.

PromptGate = Prompt Refinement Engine + External Skill Handoff Layer

• Refine messy prompts into a clear refined_prompt.
• Separate goal, background, constraints, exclusions, output preferences, assumptions, and solution candidates.
• Treat solution ideas as candidates, not requirements.
• Prefer explicitly mentioned registered skills over inferred matches.
• Automatically hand off to safe registered skills by default.
• Block high-risk or destructive handoff by policy.
• Provide thin Claude and Codex adapters.
• Validate core behavior with deterministic eval fixtures.

• It does not include named workflow skills.
• It does not hardcode downstream skill names.
• It does not run destructive actions automatically.
• It does not guarantee deterministic slash-command invocation across every platform in v0.

Build and install the private wheel:

python3 -m pip install -r requirements-dev.txt
python3 -m build --wheel
python3 -m pip install dist/promptgate-0.1.0-py3-none-any.whl

Verify the installed CLI:

promptgate doctor
promptgate eval

Run the provider quality benchmark and compare it with a saved baseline:

promptgate eval --provider --yes --limit 5 --report-json .promptgate/reports/provider-eval-current.json
promptgate eval compare --baseline .promptgate/reports/provider-eval-baseline.json --current .promptgate/reports/provider-eval-current.json

Preview hook installation:

promptgate hooks install --adapter codex
promptgate hooks install --adapter claude

Run PromptGate over a raw prompt:

promptgate --json "Redis 쓰면 되나 세션이랑 캐시랑 같이 쓰고 싶은데"

Real provider calls require:

export OPENAI_API_KEY=sk-your-openai-api-key
export PROMPTGATE_OPENAI_MODEL=gpt-5

The executable runtime is LLM-first. The provider creates a draft PromptGateResult, and Python validates schema, registry, risk, and mode policy before returning the final result.

Before distributing a private wheel, run:

python3 scripts/verify-wheel-install.py

Validate eval fixtures:

python3 scripts/validate-evals.py

Read the design spec:

docs/superpowers/specs/2026-05-14-promptgate-design.md

core/ Platform-neutral PromptGate policy and contracts
adapters/ Thin Claude and Codex adapter guidance
evals/ Deterministic behavior fixtures
scripts/ Validation tooling
docs/ Public documentation and ADRs

This repository and package are private. Public package publishing remains blocked until maintainers approve a public license and publishing target.