Why do LLMs drop Turkish diacritics during text generation?

Large language models systematically replace Turkish-specific characters with ASCII equivalents during extended text generation. Systematic degradation is observed beyond 1500 words.

Can Turkish diacritics issues be fixed with prompt engineering?

No. System prompt instructions work for the first 500 words, but the model gradually disregards this instruction during long-form generation. A deterministic validation layer is required.

How does the Claude Code Turkish diacritics plugin work?

The plugin uses Claude Code's PostToolUse hook mechanism. It triggers automatically after every Edit or Write operation, analyzes file content with hunspell, and provides feedback via stderr.

What is the 4-layer hunspell detection system?

Layer 0 handles word deduplication, Layer 1 does hunspell suggestion matching, Layer 2 generates brute-force diacritics variants, and Layer 3 uses a 2,944-entry ambiguity lookup table.

What is the performance of the Turkish diacritics plugin?

Across 201 real blog posts, it averages 4.81 seconds per file with zero timeouts. On a reference file with 90 known errors, it achieved 90/90 detection with zero false positives.

Why LLMs Fail at Turkish Characters: A Deterministic Fix with Claude Code

Large language models systematically drop Turkish diacritics (ç, ğ, ı, ö, ş, ü) during long-form content generation. The development process behind a zero-token-cost Claude Code plugin that solves this.

TL;DR

LLMs drop Turkish diacritics (ç, ğ, ı, ö, ş, ü) during long-form generation. Prompt-level warnings are insufficient. I built a Claude Code plugin with a 4-layer hunspell-based PostToolUse hook that validates every edit automatically. Tested on 201 real posts, averaging 4.8s per file, zero timeouts.

View Subscription Sign In

Subscribers Only

The full article plus 21 more subscriber-only posts come with a subscription.

ai developer-tools

Subscribers Only

RELATED

Argus: Make Your AI Coding Assistant's Web Searches Visible

Which Files Do You Actually Read? Developer Interaction Tracking

Context Engineering for AI Coding Agents: From Static Documents to a Living Ecosystem