When Claude Code Meets Technical Debt: Harsh Lessons from Using Sonnet 4.5

Core Question: How Does Claude Code Differ from IDE‑Integrated Claude Calls?

Many newcomers wonder why Claude Code, which also invokes Claude 4.5, behaves differently from Claude accessed via VS Code Copilot or other platforms. The author explains that Claude Code is a dedicated AI coding system with a full SDK, specialized prompt tuning, and a model trained specifically for code generation, whereas IDE integrations rely on generic LLM capabilities and platform‑specific prompt engineering.

1. Claude Code vs. Generic IDE Calls

Claude Code : Provides a complete SDK kit, precise prompt tuning, and a specialized training model—essentially a brain trained for programming.

IDE Integration : Uses the same underlying Claude/GPT model but its performance depends on the host platform’s prompt engineering (e.g., Copilot, Windsurf, Cursor, Trae, Codebuddy) and is a general‑purpose LLM rather than a coding specialist.

Breakthrough in September: The World’s Strongest Coding Model Arrives

On 2025‑09‑29 Anthropic released Claude Sonnet 4.5 and launched a native VS Code extension for Claude Code. After subscribing to the Claude Max Plan, the author experienced:

“Precision off the charts” – most well‑specified tasks completed in a single pass.

“Sustained combat ability” – the model ran for hours, modifying dozens of files and thousands of lines of code.

“True point‑and‑click” – no repeated rework or new chaos introduced.

According to the SWE‑bench Verified benchmark, Sonnet 4.5 achieved a 77.2% score (up to 82% with parallel execution), which Anthropic claims makes it the "world’s best coding model".

Harsh Reality: Even the Strongest Tool Can’t Rescue a Decayed Codebase

New Project vs. Legacy Project – Heaven and Hell

The author built a lightweight AI application from architecture design to production deployment, describing the process as "silky smooth". When turning to a legacy project that had accumulated technical debt, numerous issues emerged:

Port conflicts across front‑end, back‑end, database, and Redis.

Chaotic environment variables (.env, .env.local, system variables).

Repeated database connection‑pool failures.

AI gateway failing to load correctly.

Docker orchestration configuration disorder.

Even after deploying Claude Code with Sonnet 4.5/Opus 4.1 and investing dozens of hours, progress stalled.

Technical Post‑mortem: Why the Top Model Still Struggles

Context loss : Despite a 200 K‑token window, a tangled codebase leads to retrieval of conflicting implementations.

Repair loop : Fixing issue A triggers issue B, fixing B re‑breaks A.

Decision dilemma : The model prefers minimal changes, which often only treat symptoms.

Lack of global view : It can read code but not runtime state, network topology, or actual data flow.

Key Insight: AI Is an Amplifier, Not a Savior

The author concludes that AI coding tools amplify developer ability but cannot eliminate deep architectural or configuration problems. Successful rapid MVP creation (1–2 weeks) is possible for fresh projects, but legacy systems require human‑level architectural decisions.

What Sonnet 4.5 Excels At

Bootstrapping new projects from 0 to 1 – fast MVP construction, including architecture suggestions, scaffolding, best‑practice code generation, and RESTful API/database schema creation.

Clear functional iteration – adding APIs, evolving database schemas, UI component tweaks, algorithm optimization, and performance improvements.

Precise local fixes – pinpointing errors from explicit messages, correcting logic, refactoring, and writing test cases.

What It Cannot Do

Understand and resolve architectural‑level issues.

Handle chaotic environment and configuration problems.

Replace the decision‑making authority of an architect.

Practical Methodology: Getting the Most Out of Sonnet 4.5

Method 1 – The Art of Requirement Expression

Core principle: Provide a complete evidence chain so the AI has all information needed for decision‑making.

Method 2 – Give the Model Space, Avoid Over‑Constraining Prompts

When the model is sufficiently capable, overly detailed prompts can restrict its creativity, increase token consumption, and cause conflicts between user rules and the model’s built‑in best practices.

Empirical observations:

Simple tasks: the model follows instructions faithfully.

Routine troubleshooting: it may skip steps it deems unnecessary.

Complex development: it ignores lengthy constraints and defaults to its internal best practices.

Sonnet 4.5’s own strengths (77.2% on SWE‑bench, ability to plan 30+ hours of work, strong context understanding) mean that a balanced prompt—rich enough to guide but not over‑specify—yields the best results.

Conclusion

In the AI era, developers can achieve tenfold speed and tenfold insight, but only when they treat AI as a powerful assistant that amplifies, not replaces, solid engineering fundamentals.