dge-compound

DGE Compound

Use this skill after a node, track, or demand is completed.

Purpose

Convert execution knowledge into reusable learning for future DGE loops.

Core rules

1. Capture what future agents need, not a diary of everything that happened.
2. Include failed approaches when they prevent repeated mistakes.
3. Link to graph IDs, evidence, tracker records, and PRs.
4. Save learning under delivery-graph/learnings/.

Workflow

1. Read graph scope.
2. Read relevant evidence and review reports.
3. Extract durable patterns, decisions, gotchas, and validation lessons.
4. Write delivery-graph/learnings/<slug>.md.
5. Suggest whether the graph schema or skill instructions should change.

Learning format

Add a small YAML frontmatter block so dge learnings --about "<topic>" can match this learning by tag and related id (the read side of the compound loop that /dge-design and /dge-plan-graph run before scoping new work). The frontmatter is optional and legacy learnings without it still list — but new learnings should carry it so future demands actually find them.

---
title: <Learning title>
tags: [<topic>, <area>, <keyword>]
related: [NODE-###, DEM-###, REQ-###]
---

# <Learning title>

## Applies when

## Context

## Guidance

## Evidence

## Avoid

## Related graph ids

After writing a learning, confirm it surfaces: dge learnings --about "<a tag>" should list it. A learning no future demand can find is not compounding.

Output

Follow the shared output convention (see skills/README.md): lead the final reply with a bold one-line synthesis of what was captured (e.g. "Learning captured: " or "Nothing new to compound"), then any schema/skill-instruction suggestion from workflow step 5, then a ## Next block: /dge-design or /dge-plan-graph for the next demand, or a single "nothing pending" line when there is none.