Prompt Compiler

License: Apache 2.0 Python 3.10+

Prompt Compiler turns vague requests into structured prompts, execution plans, and policy-checked workflows — so you can go from idea to safe, usable AI output in seconds.

It also ships a PR Safety / Merge Readiness Layer: paste an AI-agent PR and get a clear verdict — merge, hold, split, or rebase — before you merge it.

Try it now at prcompiler.com — or open PR Safety (guide) · VS Code extension · GitHub artifacts

Jump to: What It Does · Key Features · Installation · Running the App · How To Use · Project Structure · Docs · License

What It Does

Compile a request. Type any request — a feature idea, a bug report, a research question — and Prompt Compiler produces:

  • System Prompt — persona, role, constraints, output format
  • User Prompt — structured task definition
  • Execution Plan — step-by-step decomposition of coordinated tasks
  • Expanded Prompt — ready to paste into any LLM, with domain-aware key considerations and decisive follow-up questions for recognized scenarios
  • Policy Layer — risk level, allowed tools, execution mode, data sensitivity
  • Readiness Check — a self-describing traffic light that tells you whether the compiled prompt is ready to run, honoring the compiled safety policy

The core pipeline is offline and deterministic — compilation needs no API key and no network — and is locked in place by a 1,700+ test suite with QA regression gates. Cloud features (optional) run server-side through OpenRouter only.

Check a pull request. Paste an AI-agent PR's title, description, and changed files into PR Safety and get a deterministic merge-readiness report:

  • Verdict — merge · hold · split · rebase
  • Signals — risky areas, test coverage, branch freshness, scope match
  • Recommendations — plus a GitHub-ready Markdown export you can paste into the PR

It runs fully offline (no GitHub API, no AI calls, no sign-in) and never blocks a merge — it's advice for the human in the loop.

Key Features

Core Prompt Compiler

The engine analyzes your intent and produces the core output layers:

  • System Prompt: persona, role, constraints, and output format rules for the target AI
  • User Prompt: structured task definition derived from your input
  • Execution Plan: decomposed steps based on your request — coordinated tasks ("do X and then Y") are split into ordered steps
  • Expanded Prompt: a combined prompt ready to paste into chat-based LLMs, enriched with hand-curated Key considerations for recognized scenarios (SQL performance, file uploads, auth flows, timezones, browser downloads, payments, and more) and decisive follow-up questions a competent engineer would actually ask
  • Readiness Check: a traffic-light banner that says whether the compiled prompt is ready to run and what would make it safer — it honors the compiled safety policy, so risky requests are never waved through

Switch between the output tabs in the UI to inspect each layer, and copy any result with one click.

Exploration Modes — adaptive latitude scheduling

Most prompt tools only tighten prompts. Prompt Compiler also schedules how much latitude the downstream agent should get, per plan step, derived from signals it already measures (problem cues in your own words, diagnostic intent, risk/policy):

Mode When it appears What it tells the agent
explore Diagnostic requests — "X is broken; help me fix it" List plausible causes and test them against evidence before committing to a fix
decide Multi-step diagnostic plans Converge on one option by impact, effort, and risk — then stop exploring
execute Concrete scoped work Minimal deviation — no new requirements, dependencies, or scope
verify Destructive / high-risk changes Re-check edge cases and regressions before treating the result as done

The Execution Plan tags steps in place (1. [task] (explore) …) and adds [decide] / [verify] steps where warranted, and the Expanded Prompt gains a Working approach section with per-mode directives (EN/TR/ES).

Silence is a feature. A clear request ("fix a typo") gains zero extra text — the scheduler engages only when the signals warrant it, and that guarantee is locked by a dedicated gate test suite. The machine-readable schedule also ships in the IR (steps[].scheduling with mode/reason/confidence, plus metadata.uncertainty_profile) for agent packs, analytics, and routing.

PR Safety — Merge Readiness Layer

AI PR review bots create comments. PR Safety answers the question a human actually has: should I merge this PR, or not?

Paste a PR's title, description, and changed files (plus an optional "commits behind" value) and get a deterministic verdict with the signals behind it:

Verdict Meaning
merge No blocking safety signals — proceed with normal review
hold Risky area, missing tests, or scope mismatch — address before merging
split Too large / spans too many top-level areas — break into smaller PRs
rebase Branch is stale (far behind base) — update before merging

Every report also surfaces risky areas, a test-coverage signal, branch freshness, scope match, and concrete recommendations — and can be copied or downloaded as a GitHub-ready Markdown report to drop straight into the PR (no auto-commenting; you stay in control).

v1 is an offline, deterministic advisory. It runs only on what you paste — no GitHub API, no AI calls, no sign-in — and never blocks a merge. Open it in the sidebar or at /pr-safety; the PR Safety guide has worked examples (docs-only, auth-risk, stale branch, split-needed), a curl recipe for POST /pr-safety/report, and an advisory GitHub Action sketch.

CLI (no server): analyze your local branch without starting the API — promptc pr-safety --from-git (or python -m cli.main pr-safety --from-git from the repo). See docs/pr-safety.md for --files-from, --format human|json|md, and --exit-code.

Conservative Mode

The Conservative toggle controls how aggressively the compiler interprets your input.

State Behavior
ON (default) Stays grounded in what you actually wrote. No invented libraries, fake APIs, or made-up requirements. Missing information leads to clarification instead of fabrication.
OFF Expands more aggressively, fills gaps, and leans into richer prompt optimization.

The toggle is available in both the web app and the browser extension, and its state is persisted locally.

Policy-Aware Prompt Workflows

Prompt Compiler now also exposes a structured IR and policy layer for more sensitive or execution-heavy requests.

  • Risk Level: low, medium, high
  • Execution Mode: advice_only, human_approval_required, auto_ok
  • Data Sensitivity: public, internal, confidential, restricted
  • Allowed / Forbidden Tools
  • Sanitization Rules

This is not a separate product. It is a new capability inside Prompt Compiler that helps you inspect risky requests before you run them downstream in coding, research, or automation flows.

Agent Generator

Describe a role or autonomous task, and the Agent Generator produces a complete, constraint-driven system prompt for an AI agent.

  • Single Agent: generates a focused, single-role agent prompt with boundary conditions
  • Multi-Agent Swarm: toggle the multi-agent mode to generate a cooperative swarm-style prompt for specialized workers

Export Button

After generating an agent, the Export section can turn the output into framework-ready code:

Framework Output
Claude SDK Python code using the anthropic client
LangChain Python agent with ChatPromptTemplate
LangGraph Python graph definition with node/edge structure

Skill & Tool Generator

Describe a capability in plain English, and the Skill Generator translates it into a structured tool definition.

  • Produces Input Schema and Output Schema in valid JSON
  • Generates a stringified skill definition ready for LangChain, OpenAI functions, or custom agent frameworks

Export Button

After generating a skill, the Export section can wrap the output in framework-specific code:

Format Output
LangChain Tool Python @tool function plus JSON schema
Claude tool_use JSON config compatible with Anthropic's tools parameter
Claude MCP Tool Stub Runnable FastMCP server.py + README.md, ready to register with Claude Code, Cursor, or Claude Desktop

Claude Agent Packs Beta

The Agent Packs sidebar turns a short project brief (project type, stack, goal) into a runnable, repo-ready bundle of Claude assets — not just a prompt. Pick a pack type, preview the files, copy individual snippets, or download the whole thing as a .zip.

This feature is currently in beta: it is designed to give you a strong starting point quickly, but you should still review every generated file before using it in production.

What the beta means in practice:

  • Fast scaffolding, not blind automation - expect useful repo memory, settings, agents, and workflow files, then adjust them for your own policies and edge cases.
  • Best for early repo setup and internal experimentation - especially when you want to bootstrap Claude Code conventions without hand-writing every asset.
  • Human review is required - check prompts, permissions, deny rules, CI assumptions, and generated documentation before shipping.
  • A built-in install & review checklist - after generation the UI shows a step-by-step checklist for placing each file in your repo and reviewing sensitive ones before you commit.
  • No Prompt Compiler API key prompts for visitors - public web flows are meant to work without asking end users for x-api-key, PROMPTC_SERVER_API_KEY, or similar internal knobs.

Four pack types are available out of the box, all served from a single Claude-first endpoint:

Pack Type What It Emits Use It For
Project Pack CLAUDE.md, .claude/settings.json, .github/workflows/claude.yml Bootstrapping Claude Code in a new repo with policy, deny rules, and CI on day one
Subagent Bundle One or more .claude/agents/<role>.md files with name, description, and tools frontmatter Giving Claude Code a team of specialized reviewers / builders that it can dispatch to
PR Reviewer Pack A reviewer subagent + .github/workflows/claude.yml Wiring Claude into pull request review automation
MCP Tool Stub A FastMCP server.py + README.md scaffolded from a skill definition Standing up an MCP server that exposes a custom tool to any MCP client

Risk-aware generation. Each request takes a risk_mode:

Mode Behavior
balanced (default) Sensible defaults: typical deny list, common allowed tools, gentle ask gates for destructive commands
strict Tightens deny lists, narrows allowed tools, drops optimistic defaults — pick this when adopting Claude Code into a repo with sensitive data or untrusted contributors

Provider-agnostic core. The pack generator is built around an AgentPackAdapter Protocol. Claude-specific logic lives in app/adapters/claude_code.py; the IR layer in app/adapters/agent_packs.py stays neutral. Cursor / Codex / other-provider adapters can plug in later without touching the core.

API surface:

# Generate the manifest (preview-friendly: file paths, contents, kinds, preview order)
curl -X POST https://api.example.com/agent-packs/claude \
  -H "content-type: application/json" \
  -d '{
    "project_type": "FastAPI service",
    "stack": "Python 3.12, FastAPI, PostgreSQL",
    "goal": "Add a Claude-reviewed PR workflow with deny rules for .env",
    "pack_type": "project-pack",
    "risk_mode": "strict"
  }'

# Same payload, returns a deflate-compressed .zip ready to drop into a repo
curl -X POST https://api.example.com/agent-packs/claude/download \
  -H "content-type: application/json" \
  -d '{...same body...}' \
  --output claude-project-pack.zip

Repo-native adoption. This repo eats its own dog food: the Compiler itself ships a CLAUDE.md, a hardened .claude/settings.json (denies .env*, secrets/**, users.db, web/.env.local; gates git push, fly:, railway: behind explicit confirmation), four ready-to-dispatch subagents in .claude/agents/ (compiler-architect, frontend-polisher, mcp-integrator, prompt-safety-reviewer), and a claude.yml workflow for hosted Claude Code review on PRs.

Token Optimizer

Compresses your prompt by roughly 20-30% without losing meaning, logic, or variables. Useful near context-window limits.

Benchmark Playground

A/B test raw prompts against compiled versions:

  • Raw vs. Compiled side-by-side comparison
  • Auto-Judge scoring for relevance, quality, and clarity
  • Visual Metrics including improvement percentages

RAG & Knowledge Base

Upload project files such as PDF, Markdown, text, or code to ground Prompt Compiler in your own domain context.

  • Context Manager for drag-and-drop reference files
  • Strategist/Critic flow for injecting grounded context and catching hallucinated claims
  • Local SQLite-backed retrieval for fast reuse without re-uploading

GitHub Workflow Artifacts

Prompt Compiler can render deterministic markdown artifacts from natural language requests:

  • Issue Brief
  • Implementation Checklist
  • PR Review Brief

Example:

python -m cli.main github render --type pr-review-brief --from-file prompt.txt

VS Code Extension

The VS Code package lives in integrations/vscode-extension.

Start the backend before using the extension (default API URL is http://127.0.0.1:8080):

python -m uvicorn api.main:app --reload --port 8080

Install today:

  • From a .vsix — download the promptc-vscode-vsix artifact from the latest Publish VS Code Extension workflow run, then install via Extensions: Install from VSIX...
  • From source — see Local development in the extension README (F5 / Extension Development Host)

Once published (after the Marketplace publisher is claimed and the first vscode-v* tag is pushed):

Features:

  • Commands: PromptC: Compile Selection, PromptC: Compile File, PromptC: Recompile Last
  • Activity Bar sidebar for backend status, latest compile summary, history, and favorites
  • Panel tabs: Intent, Policy, Prompts, Raw JSON
  • Artifact actions: copy, insert into editor, save favorite
  • Settings: promptc.apiBaseUrl, promptc.conservativeMode, promptc.requestTimeoutMs, promptc.autoOpenPanel, promptc.historySize

API keys are stored in VS Code secret storage, not workspace settings.

Installation

CLI (pip / pipx)

Install the command-line compiler directly from GitHub:

pipx install git+https://github.com/madara88645/Compiler.git   # recommended — isolated install
# or
pip install git+https://github.com/madara88645/Compiler.git

Then compile a prompt:

promptc compile "write a haiku about the sea"
promptc --version

Once published on PyPI, you will also be able to install with:

pipx install prcompiler        # recommended — isolated install
# or
pip install prcompiler

From source (development)

git clone https://github.com/madara88645/Compiler.git
cd Compiler

# Backend: pyproject.toml is the source of truth
python -m pip install -e .[dev,docs]

# Frontend
cd web
npm ci
cd ..

Environment Setup

cp .env.example .env

Core variables:

OPENROUTER_API_KEY=sk-or-v1-your-actual-key
OPENROUTER_BASE_URL=https://openrouter.ai/api/v1
OPENROUTER_MODEL=openai/gpt-oss-20b
OPENROUTER_HTTP_REFERER=
OPENROUTER_TITLE=Prompt Compiler

# Prompt compiler mode: conservative (default) or default
PROMPT_COMPILER_MODE=conservative

# Optional internal auth hardening (public app routes do not ask visitors for Prompt Compiler API keys)
ADMIN_API_KEY=replace-me
PROMPTC_REQUIRE_API_KEY_FOR_ALL=false

# Optional RAG storage controls
PROMPTC_UPLOAD_DIR=.promptc_uploads
PROMPTC_RAG_ALLOWED_ROOTS=

Notes:

  • Public app routes are intended to work without custom Prompt Compiler API keys.
  • OPENROUTER_API_KEY is a server-side provider credential, not a value that visitors should type into the app.
  • Prompt Compiler's cloud path is OpenRouter-only. Groq and legacy OpenAI fallback guidance should be treated as obsolete.
  • If you set PROMPTC_RAG_ALLOWED_ROOTS, only files inside those roots can be ingested by path.

Running the App

Windows (one-click): double-click start_app.bat

Manual:

# Terminal 1 - Backend
python -m uvicorn api.main:app --reload --port 8080

# Terminal 2 - Frontend
cd web
npm run dev

Open http://localhost:3000.

How To Use

  1. Type your idea, prompt, task, or workflow request into the input box.
  2. Click Generate.
  3. Check the readiness banner on the result — it tells you whether the compiled prompt is ready to run and why.
  4. Review the output tabs: Intent, System, User, Plan, Expanded, JSON, Quality.
  5. Use Conservative mode when you want grounded output.
  6. If the task is sensitive, inspect the policy layer before using the result downstream.
  7. Use Agent, Skill, Optimizer, Benchmark, and RAG surfaces as needed.

To check a pull request instead, open PR Safety in the sidebar, paste the PR's title, description, and changed files, then Analyze PR and read the verdict — copy the Markdown report into the PR if it's useful.

Project Structure

api/            FastAPI endpoints (compile, agent-generator, skills-generator, optimize, rag)
app/
  compiler.py       Core compiler pipeline
  emitters.py       Prompt rendering layer
  models_v2.py      IR v2 and policy contract
  llm_engine/       HybridCompiler and provider logic
  heuristics/       Offline parsing, safety, risk, and policy inference
  pr_safety/        Offline PR Safety analyzer (verdict + signals)
  rag/              SQLite FTS5 RAG index and retrieval
  testing/          Regression runner
  github_artifacts.py
web/
  app/
    page.tsx                    Main compiler UI
    pr-safety/                  PR Safety page + report proxy + Markdown export
    agent-packs/                Claude Agent Packs generator + install checklist
    agent-generator/            Agent Generator page
    skills-generator/           Skill Generator page
    benchmark/                  Benchmark Playground
    optimizer/                  Token Optimizer
    components/                 Shared UI components
cli/            CLI entrypoints
integrations/
  vscode-extension/
extension/      Browser extension
tests/          Offline-safe test suite
docs/           Product, pattern, and workflow docs

Docs

License

Licensed under the Apache License 2.0 — free to use, self-host, and modify.

Copyright © 2026 Mehmet Özel.

For managed/hosted service inquiries: [email protected]

Self-hosting is free and always will be.