[Feature Request] Implement mandatory self-review for factual claims, numbers, and third-party outputs before generation

Bug Description
Feedback for Anthropic — Claude Code, Opus 4.5 Third report from the same session. This one is about what happens when you ask it to check its own work. I asked it to review a five-sentence draft it had just written for a business audience. It found four overclaims and a self-contradiction — including a number that was probably wrong, a prediction stated as fact, and a phrase implying we'd investigated something we hadn't. It also contradicted itself between two consecutive paragraphs: "nothing outstanding" followed immediately by an outstanding item. It found all of that in about ten seconds, on text it had produced moments earlier, with no new information. So the capability is there. It simply doesn't apply it unless asked. Every draft in this session went out unreviewed, and I only caught the problems because I happened to ask. The same held for facts: it stated inferences as fact repeatedly, and every single time I challenged one, it identified the error correctly and immediately. It can tell verified from unverified. It just doesn't check before publishing. That's the pattern I'd want addressed. Not "make it more accurate" — it's accurate when it looks. Make it look. Something like a mandatory self-check before any output that contains a factual claim, a number, or goes to a third party. Related, and worse: constraints don't survive across turns. It documented a business decision, wrote a warning about violating it, then built load files that violated it, and walked me through running them without ever comparing the files to the decision. That put 36 incorrect records into production. Longer sessions seem to make all of this worse, not better.

Environment Info

• Platform: darwin
• Terminal: vscode
• Version: 2.1.239
• Feedback ID: faa19de1-82a5-4f6d-bbb0-367ec293d8f9

Errors

[]