suraj-dhal/reverse-prompt-engineer

You have a great AI-generated report, post, code sample or image, but not the prompt that made it. Maybe an intern wrote it and left, maybe the chat is gone, maybe you want to turn a one-off result into something repeatable. Reverse Prompt Engineer studies the output and whatever context you have, and gives you back one copy-ready prompt that can reproduce it, together with an honest account of what is evidence, what is inference and what is unknown.

It will never pretend to recover "the exact prompt": nothing can do that from an output alone. It reconstructs a prompt you can actually use, and tells you how much to trust it.

Example

You: We're a small CA firm. Our intern wrote this LinkedIn post with ChatGPT and has since left. What prompt would make posts like this for our next events?

(pastes a post announcing a free GST workshop: emoji hook, three ✅ bullets, 📍 venue line, 🎟️ "only 40 seats", hashtags)

Reverse Prompt Engineer (abridged):

## Reconstructed Prompt

```text
ROLE
You are the social media writer for Mohanty & Associates, a small chartered accountancy firm in Bhubaneswar.
TASK
Write a LinkedIn post announcing our free workshop on {{TOPIC}}.
EVENT DETAILS (use exactly as given)
Date: {{DATE}}   Time: {{TIME}}   Venue: {{VENUE}}   Seats: {{SEAT_COUNT}}
OUTPUT FORMAT
1. A hook line that starts with one emoji and asks whether the reader's business is ready. …
3. "What you'll learn:" followed by exactly three ✅ bullets …
CONSTRAINTS
Under 100 words, excluding hashtags. Do not state legal deadlines unless they appear in the verified facts below …
```

## Reasoning

1. **Observable clues:** emoji-led hook; exactly three ✅ bullets; a 📍/🕓 logistics line; …
2. **Likely intent:** a reusable event-promotion post …
3. **Context preserved:** firm, city, audience, free format; the date and seats became placeholders …
4. **Prompt-design factors:** the numbered format reproduces the six-part structure …
5. **Confidence and limits:** Overall confidence: Moderate. The claim that e-invoicing "is now mandatory for more MSMEs" is unverified …

See the full version and two more examples (code, and an instruction audit) in worked-examples.md.

What it can do

Install

Claude Code (2.1.275+):

/plugin install reverse-prompt-engineer --marketplace Suraj-Dhal/reverse-prompt-engineer

OpenAI Codex:

codex plugin marketplace add Suraj-Dhal/reverse-prompt-engineer
codex plugin add reverse-prompt-engineer@suraj-dhal

Claude.ai / Claude apps: download reverse-prompt-engineer-<version>-skill.zip from Releases, then go to Settings → Capabilities → Skills → Upload skill.

Older Claude Code versions, other hosts, updating, pinning and uninstalling are covered in docs/installation.md.

Use

Paste an output and ask, in your own words:

• "What prompt could have produced this?"
• "I lost the prompt for this report. Rebuild it from the final document."
• "Rebuild the master prompt from our chat and these notes."
• "Here's my prompt and the result. Which instructions did it miss? Then give me a prompt that reliably produces this."

Every answer has the same shape: Reconstructed Prompt (one block; fill in any {{PLACEHOLDERS}}) and Reasoning (five labelled points ending in an overall confidence of High, Moderate or Low). See docs/usage.md for modes, tips and troubleshooting.

How it works

flowchart LR
    A[Your output<br/>+ context] --> B[Inventory evidence<br/>and source precedence]
    B --> C[Analyse the target]
    C --> D[Map each requirement:<br/>explicit · observed · inferred ·<br/>unknown · conflict]
    D --> E[Synthesise with the<br/>prompt skeleton]
    E --> F[Coverage pass,<br/>then compression pass]
    F --> G[Reconstructed Prompt<br/>+ Reasoning]

The skill loads progressively: a short description at startup, the core instructions when activated, and focused playbooks (domain checks, context preservation, evidence calibration, safety, advanced procedures) only when a situation calls for them. docs/how-it-works.md explains the workflow, the evidence model, the output contract, the safety model and the design decisions.

Quality

Prompt-based software regresses silently, so every promise above is backed by a check:

The latest live results, and how to reproduce them, are in docs/evaluation.md.

Repository layout

.
├── plugin.json                      Agent Plugins manifest (+ OpenAI listing metadata)
├── .codex-plugin/plugin.json        OpenAI Codex manifest
├── .claude-plugin/                  Claude Code manifest and marketplace
├── assets/                          Icon (PNG + SVG source)
├── skills/reverse-prompt-engineer/  ← the skill (the only part loaded at runtime)
│   ├── SKILL.md
│   ├── references/                  playbooks and worked examples
│   └── assets/                      prompt skeleton, evidence matrix
├── evals/                           behavioural suite (fixtures → generated cases)
├── scripts/                         validate, lint, build evals, package, bump version, release notes
├── tests/                           unit tests
├── docs/                            documentation
└── .github/                         CI, release and live-eval workflows; templates

Development

python3 scripts/validate.py                     # structural validation
python3 -m unittest discover -s tests           # unit tests
python3 scripts/build_evals.py                  # regenerate the eval suite after editing fixtures
claude plugin eval . --tag smoke --runs 1 --ablation none   # quick live check (costs a little)
python3 scripts/package.py                      # build reproducible release ZIPs

Python 3.9+ standard library only. See CONTRIBUTING.md.

Documentation

License

Released under the MIT License. Copyright © 2026 Suraj Dhal.

You may use, copy, modify, merge, publish, distribute, sublicense and sell this software, commercially or not, provided the copyright and licence notice is kept in all copies or substantial portions. It is provided "as is", without warranty. Read docs/licensing.md for a plain-language guide, including what the licence covers, your outputs, and trademarks.

This is an independent project. It is not affiliated with or endorsed by Anthropic or OpenAI. "Claude", "Codex" and related names are trademarks of their respective owners.

Author

Suraj Dhal: surajdhal.com · GitHub @Suraj-Dhal · surajbbsr2007@gmail.com