backprop

What problem does it solve?

Propagate learnings to improve existing skills by feeding observations, corrections, gotchas, or new patterns to a backprop subagent that decides how to update skills.

Core Features & Use Cases

• Take a learning, correction, or observation from the current session
• Spawn a backprop subagent that searches existing skills for relevance
• The subagent edits the affected skills to incorporate the new knowledge
• Reports what changed and why

When to use me

Trigger on:

• "backprop this" / "propagate this back"
• "this skill is wrong about X"
• "remember that X actually works like Y"
• "add this gotcha to the relevant skill"
• Any request to update skills based on something discovered during work

How to invoke

The user provides a learning. It can be:

1. A correction: "ast-grep patterns don't work with JSX fragments unless you set language to tsx"
2. A new pattern: "found that combining has + inside with stopBy: end is the way to match deeply nested X inside Y"
3. A gotcha/anti-pattern: "never use regex on multiline nodes, it only sees the first line"
4. A missing concept: "the tree-sitter-queries skill doesn't cover the set! directive for injection.combined"

Workflow

When this skill loads, the main agent MUST:

1. Collect the learning from the user's message (and any relevant context from the current conversation)
2. Spawn the backprop subagent with a prompt containing:
   • The exact learning/correction/observation
   • Which topic area it relates to (if obvious)
   • Any code examples or evidence from the session
3. The subagent handles everything else: searching skills, deciding edits, applying them
4. Report the subagent's summary back to the user

Prompt template for the subagent

LEARNING:
{the observation, correction, or new pattern}

TOPIC HINT:
{topic area if known, or "unknown -- search broadly"}

EVIDENCE:
{any code snippets, error messages, or examples that support the learning}

The subagent will search ~/.agents/skills/, read relevant skills, and edit them.

What the subagent decides

For each relevant skill, the subagent picks one of:

• append: Add to an existing section (new example, new bullet point, new gotcha)
• correct: Fix incorrect information
• restructure: Move content between sections or add a new section
• skip: Skill is tangentially related but the learning doesn't belong there

The subagent does NOT:

• Create new skills (that's what docs-to-skills is for)
• Delete large sections
• Change skill frontmatter (name, description) unless the description is now inaccurate