How I Built a Slack Bot That Answers Codebase Questions With Claude

Описание и применение

Бот в Slack, использующий Claude и инструменты поиска/чтения кода, быстро отвечает на структурные вопросы о коде, экономя инженерам время и повышая доверие за счёт проверки путей файлов.

Задача

Engineers spent excessive time manually searching code for answers to structural questions like "Where does X live?" or "How does Y work?", costing about 20 minutes per question and reducing productivity.

Результат

The bot answers most questions in 20–45 seconds, with a median latency of 20 seconds. API costs are negligible compared to the time saved, and the system improves trust by validating file paths before replying.

Рабочий процесс

• 1. User mentions bot in Slack with a question.
• 2. Bot receives mention and forwards to Claude agent loop.
• 3. Claude decides to use "search_code" tool to find regex matches in repo.
• 4. Bot runs ripgrep via subprocess, returns limited matches.
• 5. Claude reads matched files using "read_file" tool, follows imports and references.
• 6. After a few tool calls, Claude produces a final answer with file paths and line numbers.
• 7. Bot validates all referenced paths against actual repo tree to prevent hallucinations.
• 8. Bot posts answer in Slack thread.

Доказательства

Early on, the bot answered a question with a confident reference to utils/validation.py — a file that did not exist.

Источник

Median answer latency is around 20 seconds end-to-end, the slowest multi-hop investigations land near 45.

Источник

A month of real usage cost less than a single takeout lunch in API fees.

Источник