Retro - Coding Session Retrospective and Environment Optimization Skill

Skill Overview

Retro is a coding session retrospective skill: it reads the log of the session you specify, identifies where the agent got stuck, took detours, or made mistakes during the run, and turns its findings into concrete recommendations for improving the coding environment—typically a navigation pointer in CLAUDE.md / AGENTS.md, a rule in CODING_STANDARDS.md, or an automated check that should be added.

Use Cases

1. A session took too long, and you want to understand where the time went
   When an agent repeatedly searches through files or takes too long to locate the correct entry point, it often indicates that a navigation pointer is missing. Retro highlights these time-consuming information-discovery points separately.

2. The same mistake keeps recurring
   Errors that should have been caught by linting, type checks, or tests—or rules that should have been enforced during review but were never communicated to reviewers—can be formalized as automated checks or coding standards after the retrospective.

3. Project configuration and documentation keep growing
   When AGENTS.md becomes too bloated for anyone to read completely, or contains “ineffective instructions” that do not affect behavior, Retro identifies which content should be moved into coding standards, converted into check scripts, or supplemented with additional access to information that the agent cannot obtain on its own.

Core Functions

1. Read first-hand session records
   Retro reviews the session specified by the user—or the current session by default—and performs its retrospective based on what actually happened rather than on impressions.

2. Identify improvement opportunities across seven areas
   Navigation (whether a navigation pointer should be added), automated checks (which errors could have been caught by a machine), coding standards (which review rules should be added, removed, or clarified), global AGENTS.md (which instructions should be relocated), tool economy (whether expensive tool calls can be streamlined), ineffective instructions (which instructions do not actually change the agent’s behavior), and information access (whether key information sources are missing, such as development server logs or read-only access to third-party services). Each area has a specific trigger condition rather than relying on vague generalities.

3. Present improvement candidates ordered by severity
   The retrospective produces a prioritized list of candidates for the user to decide which ones to adopt. The skill itself does not directly modify the project.

Frequently Asked Questions

Will Retro modify my code or configuration files?

No. Retro produces a list of improvement recommendations, presented in order of severity. The user decides which recommendations to adopt and how to implement them. It does not automatically rewrite code, CLAUDE.md, or any configuration files.

How does Retro differ from code review?

Code review focuses on a single change—it can be performed from the diff, without requiring the implementer to explore the codebase, so the review stage has minimal contextual burden and is best suited to enforcing coding standards. Retro focuses on a session process—it examines why the implementer could not find a file or why they encountered a pitfall, and then improves the environment itself. The two improvement directions are often complementary: the review stage is better suited for adding rules, while retrospectives are better suited for adding navigation and checks.

Which session does Retro review when no session is specified?

By default, it reviews the current session. If you want to review a historical session, you must specify it explicitly—which may involve searching the local session logs.