Skip to content

AI Generation

Once the collectors bundle your git metadata and GitHub activity, git-brief injects it into a highly constrained System Prompt.

Language models are verbose. If you pass a raw git log to an LLM, it will output a 3-paragraph essay about how “The developer significantly improved the authentication module by refactoring…”.

That is useless for a daily standup. Standups require punchy, action-oriented bullet points.

To enforce a high-signal, low-noise output, the embedded system prompt enforces strict rules on the LLM:

  • Action Verbs Only: Pronouns (“I”) are forbidden. Bullets must start with verbs like “Fixed”, “Reviewed”, or “Pushed”.
  • Extreme Specificity: It must extract exact PR numbers (e.g., PR #234), branch names, or file paths from the context bundle.
  • Timestamp Enforcement: It is strictly forbidden from placing a commit dated “Today” into the “Yesterday” bucket.
  • Length Limits: Maximum 3 bullet points per section. The total output is hard-capped at 100 words.

Because the prompt relies exclusively on metadata (commit messages, PR titles, uncommitted file paths) and never your actual source code, the context window is tiny.

A typical git brief run consumes ~500 input tokens and ~100 output tokens.

Provider Model Cost per Run
Google Gemini gemini-2.5-flash Free
Anthropic claude-3-5-haiku-latest ~$0.001
OpenAI gpt-4o-mini ~$0.001