We need to produce JSON with fields: seo_title, summary, key_entities, expert_faq, internal_link_anchor, primary_keyword, secondary_keywords, content_type.

We must follow style: 180-250 words, 3-4 paragraphs, each 45-65 words. Must bold a few key details (3-5 total). Must not use banned words, etc.

We need to extract info from source: The source is a Hacker News post titled "Better Models: Worse Tools". The original text describes a problem where newer Claude models (Opus 4.8, Sonnet 5) produce malformed tool calls for Pi's edit tool, adding extra invented fields in the edits array, causing rejection. Older models didn't have this issue. The issue appears due to training on Claude Code harness which is forgiving, leading models to invent fields. The summary must capture that.

We need to bold key details: maybe Opus 4.8, Sonnet 5, Pi’s edit tool, Claude Code, grammar-aware decoding? But we need only 3-5 total across summary.

We must not invent facts. Only use info from source. The source mentions Anthropic models, Opus 4.8, Sonnet 5, Pi's edit tool, Claude Code, grammar-aware decoding, etc.

We need to include key_entities: companies (Anthropic), maybe Pi? Pi is a tool? It's likely a tool name; but it's not a company. Could be a project. We'll list companies: Anthropic. People: Petr Baudis (mentioned). Locations: none.

We need expert_faq: one insightful question and answer based on article.

Internal_link_anchor: 2-4 word topic phrase.

Primary_keyword: single most important 2-3 word keyword: maybe "Claude models" or "tool calling". Let's think: The article is about newer Claude models worsening tool call adherence. So primary keyword could be "Claude models". That's 2 words.

Secondary_keywords: list 3-5: "tool call schema", "Opus 4.8", "Sonnet 5", "Claude Code", "grammar-aware decoding".

Content_type: likely "news" (timely) or "article". Since it's a Hacker News post dated July 04, 2026, it's news. We'll set "news".

Now craft summary 180-