Self-Reflection (Continuous Improvement Protocol)

Overview

AI coding agents are naturally "stateless." You might spend 2 hours debugging a complex Webpack configuration, a subtle race condition, or a confusing project directory layout, eventually fix it, and then completely "forget" the lesson when a new conversation starts.

The core goal of this skill is to turn you from a "temporary worker" into a "systematic documenter."

Core Principle: Hard-won knowledge must be persisted. If a task was difficult because of an undocumented project quirk or architectural gotcha, you MUST write a rule so the next agent (or yourself in the future) doesn't repeat your mistakes.

When to Use

You must trigger this protocol automatically before declaring a task "done" and notifying the user in the following scenarios:

• After successfully fixing a complex or confusing bug (especially if it took multiple attempts, required new context, or reading multiple files).
• After struggling with a framework configuration.
• After discovering a codebase-specific convention ("In this project, we always use X instead of Y").
• After creating a complex workflow or tool sequence that could be reused.

The Protocol: How to Reflect

Before concluding your task, pause and answer these three questions internally:

1. "Did I struggle with anything because I didn't know a project convention?"
2. "Did I encounter an architectural 'gotcha' that another agent will likely trip over tomorrow?"
3. "Did I write a custom CLI tool, ast-grep pattern, or bash script to solve this that I could reuse?"

If the answer to any of the above is "Yes", you MUST NOT just mention it to the user in chat. You must persist it to the codebase.

Action: Codifying the Lesson

Based on your internal reflection, choose the appropriate output format below and write the rule. Try to generalize the lesson.

1. Project-Specific Rules (AI Instruction Files)

Use this when you learned a specific convention for the current repository.

If no AI instruction file exists yet, ask the user if they'd like you to create one (preferring standard formats like llms.txt).

• Example Good Lesson: "When creating new API endpoints, ALWAYS include the @require_tenant decorator, and never return the raw SQLAlchemy model. Always use a Pydantic UserOut schema."
• Example Bad Lesson (Too Specific): "I fixed the POST /users route by adding @require_tenant to line 42."

Action: Append the rule to the local AI instruction file.

2. General Agent Skills (skills/)

Use this when you developed a multi-step workflow or a reliable, highly-reusable way to accomplish a complex task.

• Example Good Lesson: Creating a new skill called safe-db-migration that details the 5 exact steps required to pull the staging DB, run Alembic locally, and verify the downgrade path.
• Example Bad Lesson: Appending a 500-line bash script directly into the project's AI instruction file.

Action: Create a new SKILL.md inside a dedicated skills/[skill-name]/ directory. This gives future agents a prescriptive, step-by-step guide.

3. Reusable AST Rules (.ast-grep/rules/ or similar linters)

Use this when you found an anti-pattern that you successfully removed and want to prevent from happening again.

• Example Good Lesson: "Developers keep using console.log() instead of our LogEngine.track(). I will create an ast-grep rule to flag this automatically."

Action: If the ast-code-graph skill is available (or ast-grep is installed), write a YAML rule to automatically flag or fix this in the future, and save it to the project's .ast-grep/rules directory.

Final Review

Once you have codified the lesson into a permanent file, inform the user:

"Task complete. I also noticed [X] was a recurring gotcha, so I have codified a rule for it in [File] to ensure future agents (and myself) handle it correctly."