AI Agent Use Cases
Views —

AI Agent Use Cases

Added today

2All 191

Prompt to Prepare a Shopping Cart from a Recipe Photo

The author reports that Grok Bot receives a recipe photo, organizes the ingredients and quantities, and creates a shopping list by excluding ingredients already available at home.

Grok Bot
Life & Hobbies

1–20 of 191

Dev

Operating a Team Where AI Handles Code Reviews and Humans Only Check High-Risk Changes

The author reports that the Anthropic Claude Code team delegates code reviews to AI, with humans only directly reviewing high-risk changes.

Claude Code

How it works

  1. After writing code, the author verifies the changes using executable tests, builds, or design comparison screenshots.
  2. They show the results of the code changes to Claude, and an internal sub-agent reviews the diff in the new context.
  3. The author documents project and review standards clearly in files like CLAUDE.md and REVIEW.md, so the AI reviewer can reference them.
  4. For every pull request (PR), Claude automatically conducts a code review. It checks for issues by considering the purpose of the change and the entire codebase, and multiple agents vote to verify the presence of bugs.
  5. Security reviews are also automated; if any bugs or security issues remain, merging is blocked. The author notes that open-source GitHub Actions or the /security-review command can also be used.
  6. Automated linting, type checks, tests, and builds must all be passed. The PR template specifies how to test and how to roll back changes.
  7. Depending on the risk level of the change, the author applies different levels of human approval. Low-risk changes are merged automatically, while medium or higher risks require approval from one or two people. Only trusted authors can use the automated path.
  8. All checks are rerun in the merge queue to ensure no issues after merging. New features are first released to internal staff via feature flags and then gradually rolled out.
  9. To quickly detect and roll back issues, the author operates feature flags, dashboards, and an on-call system.
  10. The review process and AI reviewer are evaluated daily using real incident cases, and recurring problems are addressed through rules, documentation, or skill improvements.

Results

The author describes operating a system where AI automatically reviews most code changes, and only high-risk changes require direct human approval. After passing automated reviews, security checks, linting, testing, and builds, the level of human approval is adjusted based on risk. The author reports that the number of merged PRs doubled, while the rollback rate dropped by over 40%. Recurring review feedback is improved through documentation and rules, and incident cases are reflected in daily evaluations to enhance the quality of the AI reviewer.

Notes

The author explains the division of roles between automated code reviews and human involvement based on their actual experience with the Claude Code team. Specific configuration values, publicly available code, or prompt examples are not included. Internal criteria for trust, risk classification, and approval policies should be adjusted according to each organization's situation.

Source

Anthropic Claude.dev 블로그 Original post ↗

Example prompt · Claude Code

Goal: Implement a team workflow where AI automatically reviews most code changes, and only high-risk changes require direct human approval.

Required inputs and environment:
- GitHub repository and pull request (PR)-based development environment
- Access to Claude Code (or a similar AI code review tool)
- Project-specific CLAUDE.md (project description, stack, rules, etc.) and REVIEW.md (review standards, failure examples, skip paths, etc.) documentation
- Automated CI pipeline with linting, type checks, automated tests, and builds
- GitHub Action or /security-review command for security reviews
- Risk classification criteria (e.g., low risk: documentation/internal features, medium risk: core features, high risk: payments/build/CI, etc.)
- List of trusted authors and approval policy

Implementation and setup steps:
1. Write CLAUDE.md and REVIEW.md in each repository so the AI reviewer can reference them.
2. Configure workflows to automatically run /code-review and /security-review when a PR is created.
3. Integrate automated checks for linting, type checks, tests, and builds into PRs.
4. Add fields to the PR template specifying test methods, risk level, and rollback procedures.
5. Implement risk classification logic (e.g., based on path, change content, author trust level) so that low-risk PRs are merged automatically, while medium or higher risks require approval from one or two people.
6. Allow only trusted authors to use the automated path; all others always require human approval.
7. Rerun all checks in the merge queue to ensure no issues after merging.
8. Release new features to internal staff first using feature flags, then roll out gradually.
9. Collect incident and rollback cases to use as evaluation sets for the AI reviewer, and improve recurring feedback through rules or documentation.

Deliverables:
- A workflow where most PRs are merged with only AI review and automated checks, and only high-risk changes require direct human approval
- A system for rapid rollback and root cause analysis in case of incidents, and for improving the review process

Verification steps:
- Check that automated reviews, security checks, linting/testing/builds work correctly for each PR.
- Test with sample PRs to ensure risk classification and approval policies are applied as intended.
- Confirm that rollback and evaluation systems function properly in the event of an incident.