[Bug]: AI Tailor consistently lowercases descriptions and rewrites only specific sections across multiple models

[mat926]

Describe the bug

The AI tailor keeps lowercasing each description , I tried using ollama with various models (gemma, deepseek, mistral) as well as gemini/openai. Every resume I tailor, it always makes these same changes in modifying the description , and I end up editing them manually.

In addition , it's always modifying the same 5 descriptions and summary . I never seen it modify the other descriptions, the skills or any other part of the resume.

To Reproduce

1. Setup AI providers
2. Create master resume
3. Create tailored resume with pasted job description

Expected behavior

I should see the descriptions not lowercased and every part of the resume should've been tailored.

Environment

• OS: Docker on Arch Linux
• Browser / Version: Brave
• Device:

Screenshots

Here are 3 screenshots for different software engineering job postings I tailored . In all of them you can see how it lowercases the descriptions. And it's always modifying those same descriptions.

[linear]

RES-254

[srbhr]

Both halves of this were worked on in c708b76 (PR #827), now on main — but I'm leaving this open, because only one half is structurally guaranteed and you're the person best placed to confirm the other.

"it's always modifying the same 5 descriptions" — fixed in code. The tailor's allow-list genuinely only permitted summary, work/project descriptions and skills; everything else was silently unmodifiable. It now also covers education[i].description, additional.languages, certificationsTraining and awards, with the diff field types needed to surface those in the preview modal (building them previously raised a validation error). Degree, institution and years remain intentionally locked. This part is enforced by code and covered by tests.

"consistently lowercases descriptions" — mitigated, not guaranteed. Worth being straight with you: there is no code path that lowercases your text. I checked, and every case-folding operation in the improver is comparison logic, not a write. The lowercasing was the model's own output, so the fix is a prompt rule instructing it to preserve original capitalization, particularly for proper nouns, acronyms and technical terms (REST, API, AWS).

That means it depends on the model actually following the instruction — and you reproduced this across gemma, deepseek, mistral, gemini and openai, which is exactly the spread where a prompt-level guardrail is least reliable.

So: could you pull latest main and re-run a tailor on the same resume/JD that produced those screenshots? If the casing holds now we'll close this. If it still lowercases — especially if you can say which model — that tells us a prompt rule isn't sufficient and it needs a code-level fix instead, which is useful to know.