Rephrasy

Build me a local AI-text humanizer workbench. This is the honest consolation build: it makes AI text read naturally, it does NOT promise to beat AI detectors.

Stack: Node 22 + Express, vanilla JS frontend, SQLite via better-sqlite3. No build step.

- One page on localhost:5180: input textarea left, output pane right, a Humanize button, and a strength select (light touch / standard / heavy rewrite).
- Humanize calls an LLM (Anthropic or OpenAI, key in .env) with a fixed system prompt per strength level that: varies sentence length and rhythm, cuts hedging and filler ('delve', 'moreover', 'it's important to note'), swaps uniform paragraph shapes for uneven ones, and keeps meaning, facts, names, and numbers intact. Stream the output.
- A word-level diff view between input and output (the `diff` npm package) so I can see exactly what changed.
- A Re-roll button that re-humanizes the current output with a different seed phrase in the prompt, keeping the last 5 attempts switchable via tabs.
- Every run saved to SQLite: timestamp, strength, input, output. History page with the last 100 runs and copy buttons.
- Cmd+Enter runs humanize.
- .env.example with the key name; README states clearly: output is for readability, no detector-bypass guarantee, and text leaves the machine only for the LLM call.
- Out of scope, deliberately: detector score checking, browser extension, accounts, multi-language tuning, API access.

# AGENTS.md — Build guide for Rephrasy

## Project scope
Paste text, rewrite it with an LLM prompt tuned to cut AI tells and vary sentence rhythm, diff the result, and iterate until it reads naturally.

Catalogue verdict: no. You can one-shot a rewriting wrapper that makes text sound less robotic, and for that job the DIY build is genuinely fine. What you cannot one-shot is the actual product: custom fine-tuned models plus a continuous evaluation loop against detectors (GPTZero, Turnitin, Copyleaks, Pangram) that retrain specifically on LLM-rewritten text. A prompted rewrite moves detector scores inconsistently, and detectors drift monthly, so a static prompt that works today quietly stops working. If your bar is 'reads naturally', build it; if your bar is 'passes detectors reliably', the moat is the model and the eval treadmill, not the text box.
Use the implementation prompt below to define the deliverable. Complete each phase's acceptance checks before extending the scope.

## Working agreement
- Inspect the repository and its existing instructions before choosing paths, dependencies or commands. Keep one coherent stack and explain changes to the proposed architecture.
- Plan a vertical slice that accepts a real input and produces the useful output described below. Persist only the state the prompt calls for; respect memory-only and upstream-managed workflows. Use fixtures only when they are clearly labelled.
- After scaffolding, document the actual install, development, check and build commands in README and keep them synchronized with the package or project manifest. Do not report commands as successful unless they ran.
- Work in small steps. At handoff, list implemented flows, checks actually performed, remaining blockers, and any credentials or provider setup the owner must supply.
- Do not publish, spend money, contact customers, delete source data or run irreversible migrations without the project owner's authorization.

## Prerequisites
- OpenAI/Anthropic API key

## Stack and architecture
- Stack: Node 22 + Express, vanilla JS frontend, SQLite via better-sqlite3. No build step.
- Model source documents, editable pages or notes, internal links, revisions, and exports; retain source IDs and timestamps.
- Interface for this AI humanizer + detector bypass workflow: a local web page with input, progress, review, and export views.

## Security and data integrity
- If OpenAI/Anthropic API key is unavailable, keep the source record and show a recoverable error instead of a fabricated result.
- Keep API keys and OAuth secrets in .env; redact them from logs, generated files, and error responses.
- State the practical limit: fine-tuned models trained specifically to survive AI detectors.

## Agent implementation rules
- Project rule, data: Model source documents, editable pages or notes, internal links, revisions, and exports; retain source IDs and timestamps.
- Project rule, behavior: One page on localhost:5180: input textarea left, output pane right, a Humanize button, and a strength select (light touch / standard / heavy rewrite).
- Project rule, recovery: If OpenAI/Anthropic API key is unavailable, keep the source record and show a recoverable error instead of a fabricated result.

## Optional agent skills and references
No installable skill is listed for this guide; use the project rules and the stack documentation.

Read the linked SKILL.md and its dependencies before adding a skill. Select only the skills matching this project's runtime and task; their documentation does not supply API access, credentials or approval to perform external actions. Pin the reviewed revision where the tool supports it. Follow the chosen agent's documented project-level installation mechanism.

## Distribution ideas
These are optional planning notes. Obtain the owner's approval before publishing or contacting anyone.
- Demonstrate this AI humanizer + detector bypass operation with sample or user-owned data: One page on localhost:5180: input textarea left, output pane right, a Humanize button, and a strength select (light touch / standard / heavy rewrite).
- Show a concrete result or export: Every run saved to SQLite: timestamp, strength, input, output. History page with the last 100 runs and copy buttons.
- State the practical limit: fine-tuned models trained specifically to survive AI detectors.

## Engineering roadmap
1. Phase 1, architecture and data: Stack: Node 22 + Express, vanilla JS frontend, SQLite via better-sqlite3. No build step. Model source documents, editable pages or notes, internal links, revisions, and exports; retain source IDs and timestamps.
2. Phase 2, implement: One page on localhost:5180: input textarea left, output pane right, a Humanize button, and a strength select (light touch / standard / heavy rewrite).
3. Phase 3, implement: Every run saved to SQLite: timestamp, strength, input, output. History page with the last 100 runs and copy buttons.
4. Phase 4, review and output: Paste text, rewrite it with an LLM prompt tuned to cut AI tells and vary sentence rhythm, diff the result, and iterate until it reads naturally. Every run saved to SQLite: timestamp, strength, input, output. History page with the last 100 runs and copy buttons.
5. Phase 5, recovery and acceptance: If OpenAI/Anthropic API key is unavailable, keep the source record and show a recoverable error instead of a fabricated result. Verify this invariant with a saved fixture: A classifier score cannot be presented as proof of authorship; a failed detector run remains unknown and never overwrites source text. State the practical limit: fine-tuned models trained specifically to survive AI detectors.

## Paid-product capabilities outside this build
- fine-tuned models trained specifically to survive AI detectors
- continuous re-testing as GPTZero/Turnitin/Copyleaks/Pangram update
- built-in detector scoring to verify output before you use it
- 50+ language support and custom writing styles
- Chrome extension and REST API

## Implementation prompt
Build me a local AI-text humanizer workbench. This is the honest consolation build: it makes AI text read naturally, it does NOT promise to beat AI detectors.

Stack: Node 22 + Express, vanilla JS frontend, SQLite via better-sqlite3. No build step.

- One page on localhost:5180: input textarea left, output pane right, a Humanize button, and a strength select (light touch / standard / heavy rewrite).
- Humanize calls an LLM (Anthropic or OpenAI, key in .env) with a fixed system prompt per strength level that: varies sentence length and rhythm, cuts hedging and filler ('delve', 'moreover', 'it's important to note'), swaps uniform paragraph shapes for uneven ones, and keeps meaning, facts, names, and numbers intact. Stream the output.
- A word-level diff view between input and output (the `diff` npm package) so I can see exactly what changed.
- A Re-roll button that re-humanizes the current output with a different seed phrase in the prompt, keeping the last 5 attempts switchable via tabs.
- Every run saved to SQLite: timestamp, strength, input, output. History page with the last 100 runs and copy buttons.
- Cmd+Enter runs humanize.
- .env.example with the key name; README states clearly: output is for readability, no detector-bypass guarantee, and text leaves the machine only for the LLM call.
- Out of scope, deliberately: detector score checking, browser extension, accounts, multi-language tuning, API access.

## Completion evidence
Demonstrate the prompt's acceptance scenarios against the scoped workflow. Include setup from a clean checkout and failure recovery. Check persistence across restart and export/restore only for the state the prompt says to store; for memory-only tools, confirm that temporary content is discarded as specified. Record actual results and remaining limitations. A detailed plan alone does not establish a working replacement.