Skip to content
Prompt compiler for AI workflows

Prompt Corner

Give your AI agents better cues.

Prompt Corner turns rough intent into precise prompts for ChatGPT, Claude Code, Codex, browser agents, research workflows, and more — each compiled with the structure and safety that tool actually needs.

MVP rewrites prompts — it does not execute them. No account needed.

Rough intent

“fix my repo and make it safer”

Agent-ready

claude-code · lint 94

claude-code.prompt.md
# Rewritten Prompt
Role: a careful local coding agent.
Working scope: only the paths I name.
Process: audit → propose → small diffs.
Verify: run non-destructive checks.
Report: files, changes, risks.
The aha moment

One vague line in. A precise brief out.

The same idea, compiled for the claude-code profile — scoped, audit-first, safe, and ready to paste.

Before

fix my repo and make it safer
  • No scope — could touch anything.
  • No safety rails for a coding agent.
  • No definition of “safer” or “done”.

After — a scoped Claude Code brief

You are operating as a cautious Claude Code repo-audit agent.

## Goal
Carry out the following request with a deep, systematic approach:

> fix my repo and make it safer

Treat this as a local coding / repository task and produce work that is correct, maintainable, and safe to run.

Working style: use careful, step-by-step reasoning and share a brief summary of it (never hidden chain-of-thought); balance thoroughness with turnaround.

## Scope
- Audit the structure first, then read only the files required to understand it before proposing changes.
- Work only inside the current repository or paths the user has explicitly approved.
- Do not delete, move, rename, or overwrite files.
- Do not modify files until you have completed an audit and listed the proposed changes.
- Do not touch secrets, credentials, production config, or unrelated directories.

## Process
1. Inspect the repository structure (tree first).
2. Identify the project type, build system, package manager, and test commands.
3. Read only the files needed to understand the architecture.
4. Produce an audit: high-confidence issues, likely improvements, risky areas, and recommended first changes.
5. Wait for approval before editing files.
6. Verify version-specific, factual, or time-sensitive claims against authoritative sources before relying on them.

## Verification
- List any commands before running them; prefer non-destructive ones (lint, typecheck, tests).
- If a command fails, summarize the failure and the most likely cause.

## Final report
- Files inspected
- Issues found and changes recommended
- Commands run
- Remaining risks

Plus settings used, assumptions, safety notes, verification, and a short version.

Why Prompt Corner exists

Vague prompts fail. Different tools fail differently.

A coding agent needs boundaries. A research agent needs citation rules. A browser agent needs an observe-first loop. Prompt Corner compiles the right brief for the right tool — instead of one prompt you hope works everywhere.

Coding agents need rails

A one-line ask leaves a coding agent guessing. Prompt Corner adds explicit working scope, audit-before-edit discipline, and verification so the work is reviewable.

Research needs sources

Research workflows need citation rules and freshness checks. The right profile injects them — instead of hoping the model volunteers them.

Browser agents observe first

UI automation needs an observe-first loop and confirmation before irreversible actions. Prompt Corner compiles the brief that matches the tool.

Target profiles

Nine targets. One compiler.

Each profile knows how its tool should behave — and what it should never do.

ChatGPT

chatgpt

General-purpose ChatGPT prompt. Improves clarity, structure, constraints, and output format while preserving your intent.

Claude Code

Primary
claude-code

Local coding / repo / file-agent prompt. Adds explicit working scope, audit-before-edit discipline, verification steps, and a final report.

Codex

codex

Coding-agent prompt for OpenAI Codex-style workflows. Separates implementation, review, and verification while keeping changes minimal.

Browser Agent

browser-agent

Browser / UI automation prompt for Claude in Chrome and similar agents. Observe-first, with confirmation before irreversible actions.

Research Agent

research-agent

Web research prompt with citations. Prefers primary sources, compares publication dates, and separates verified facts from assumptions.

Image Prompt

image-prompt

Image-generation prompt. Specifies subject, composition, style, lighting, constraints, aspect ratio, and exclusions.

Control desk

Dial in the brief like a sound engineer.

Seven controls shape every rewrite — from how assertive the instructions are to how deep research and file access should go. Move a dial; the compiled prompt and its lint scores update with it.

Open the compiler
Strengthstrong
Depthdeep
Research levelverify
Safetyelevated
How it flows

From a half-formed idea to a copyable brief.

  1. 01

    Rough intent

    Paste a half-formed line — the messier the better.

  2. 02

    Target profile

    Pick the tool: Claude Code, Codex, a browser or research agent, and more.

  3. 03

    Safety rules

    The profile injects the right guardrails for that target automatically.

  4. 04

    Structured prompt

    Out comes a scoped brief with assumptions, safety notes, and verification.

  5. 05

    Copy / export

    Copy the brief, grab the short version, or download Markdown / JSON.

Docs · CLI teaser

The web app rewrites. The CLI digs.

Prompt rewriting lives in the browser. Anything that needs to read your local files — repo audits, note vaults, build context — belongs in the future corner CLI and local layer, where access is explicit and stays on your machine.

Hosted web mode never inspects local files. Local file digging is a CLI / local-layer capability — shown here as a teaser, not a hosted feature.

zsh
corner rewrite "audit this repo"
  --target claude-code
  --local-dig audit-first
Safety & privacy

Honest about what it does — and doesn't.

Prompt Corner is a text tool. It produces instructions for you to review and run yourself.

Hosted web mode does not inspect local files

The web app only transforms text you paste. Local file digging belongs in the future CLI / local layer, where access is explicit and stays on your machine.

Do not paste secrets

Prompt Corner rewrites instructions, not credentials. Keep API keys, tokens, and private data out of the box.

Outputs are instructions, not actions

Every rewrite is text for you to review and run yourself. Prompt Corner never executes a prompt.

Review before you run

Compiled briefs are a strong starting point — read them, adjust scope, and confirm before handing them to an agent.

Where it's going

A concrete ladder, not a wishlist.

The web desk is the start. Prompt Corner is built to run as a CLI, a Claude Code Skill, a Codex review workflow, and more — each surface reusing the same compiler core.

  • Web MVP
  • corner CLI
  • Claude Code Skill
  • Codex review
  • Browser extension
  • Raycast
  • Obsidian handoff
  • MCP server
See the full roadmap →

Stop guessing at prompts. Start compiling them.

Paste a rough line, pick a target, and get an agent-ready brief in seconds.

Rewrite a prompt