Most PDF extraction pipelines would have missed the Brazilian court-filing injection.

Two law firms hid white-on-white text in a labor court petition. It told the court's AI reviewer to answer superficially and not challenge the documents. The court's tool, Galileu, flagged it and blocked it, and the court fined the firms 10% of an R$842,000 claim.

Most of the Hacker News thread blames the court for running AI on filings. But the trick works on any pipeline that reads PDFs. White text is invisible on the page and still sits in the text layer, which is what gets extracted.

My extraction pipelines check each field against the page it came from. That catches a wrong value, but not an instruction that changes how the model reads the rest of the file. That needs a check before the model sees anything: text a person can't see on the page shouldn't reach the model.

In the PDFs you extract from, how often does the text layer hold something the page doesn't show?

I’ve been working on a small skill called Manual Orchestrator to make longer projects with AI agents easier to manage.

The problem started with a familiar situation: one chat would accumulate plans, implementation details, corrections, and test results. By the time I was ready for the next phase, I wanted a fresh conversation without losing the decisions we’d already made.

Manual Orchestrator gives that process some structure. One chat acts as the orchestrator: it breaks the work into stages and writes a prompt for the next one. I run that prompt in a new session, then bring the result back. The orchestrator reviews it against the stage’s criteria before preparing the next prompt. If something needs fixing, I go back to the same worker session.

The skill also keeps a record in the project with the current stage, decisions, evidence, and next steps. That makes it easier to pause, resume, or hand the work to another session.

The handoff is manual. I choose when to open a session and carry the result back. The orchestrator and worker sessions can also use different agent CLIs, with the prompt adapted to the capabilities available.

It’s open source and still a work in progress. I’m using it to learn what helps in practice and where the workflow needs to improve. Installation has been checked for Codex, Claude Code, and Gemini CLI, but the full workflow hasn’t been tested across every CLI or version.

If you try it, I’d be interested in hearing what works and where it gets in your way.

I like that the handoff is manual. A written record with decisions and next steps is what I miss most when a long AI chat gets messy. How do you decide when a stage is finished: with tests, or with your own review?