README

AI-assisted proofreading for TYPO3 page content. Detects issues like spelling, punctuation, grammar, gender-inclusive language, style. Results are shown in a dedicated report, where suggestions can be applied to the content with one click.

Status: alpha. Works and is in production use on TYPO3 11, but has not been widely tested. See "Limitations" below. If you are using this, please report any issues!

Limitations

This is an early exploratory version with quite some limitations:

German content only. The A.I. prompt was optimized over multiple iterations for German language content.
Single backend: OpenRouter The general idea is to support generic OpenAI-compatible backends, but I have not been able to test this with anything other than OpenRouter yet. The extension currently relies on a strict JSON output format.
Single page language only. So far, only L=0 is checked.
No workspace support. The extension reads and reports on live content only; workspaces are ignored and untested.
Only tested on Typo3 11. Targets Typo3 11-13, but 12-13 are thus untested.

How it works

Access the new module "KI-Lektorat" from any page:
Checks need to be scheduled ("Report erstellen"). Depending on settings the check can take multiple minutes.
When ready, the report is automatically shown:
Findings can be auto-applied (where possible), marked as manually applied or discarded.
When done, mark the report as "Geprüft" to update the header icon and statistics.

Installation

TYPO3: 11.5 / 12.4 / 13.4 · PHP: 8.2+

composer require bmorg/ai_proofread

Or download via Github Releases.

In the TYPO3 backend:

Database tables are created on install.
Adapt extension configuration (see below).
Scheduler task Register the command aiproofread:process-queue as a recurring Scheduler task (every 60s).

Configuration

1. Extension settings

Admin Tools → Settings → Extension Configuration → ai_proofread

Fill in at least the provider settings. Recommended:

Setting	Value
`baseUrl`	`https://openrouter.ai/api/v1`
`apiKey`	`sk-or-…`
`requestTimeout`
`maxConcurrency`
`useMock`	`0`

2. Model configuration (KI-Lektorat → Dropdown: Einstellungen)

Model selection

Ships with a set of preconfigured models.

The current prompt was tested against a set of ~25 frontier models with a sample page and scored against a set of expected finds (+ false-positives).

Preset	Findings
Claude Opus 4.8 (default)	finds most errors; no false positives; sightly cheaper than Sol
GPT 5.6 Sol	found all must-finds; low false positives; costly
Claude Fable 5	found all must-finds; low false positives; prohibitively costly (~3× Sol)
GPT 5.6 Luna	finds most errors; few false positives; fast; cheap (0.2× Sol)

Recommendations:

Reasoning on gives much better results (slower and more expensive).
Pin a provider (e.g. anthropic) if the default routing lands on providers that e.g. don't return the required JSON.

Prompt config

Setting	Value
`siteDescription`	For better context
`extraPromptInstructions`	Allows adding of custom rules
`enableGenderInclusiveLanguage`	`1`
`genderInclusiveStyle`	E.g. "Nutzer/innen"

Recommendations:

When using extraPromptInstructions, use strict language, e.g. "Korrigiere NIE Anführungszeichen". Permissive wording ("kann ignoriert werden") will get ignored by some models.

Development

This is an exercise in vibe-coding. Fully developed by Claude Code (Opus 4.8 and Fable 5) - heavily guided and reviewed, but almost no code was handwritten.

Contributions are welcome, although some form of human involvement is much encouraged.

License

MIT — see LICENSE.

bmorg / ai_proofread

Maintainers

Package info

Statistics

Security