Back to Skills

curate-delta

verified

Synthesize Reflector insights into structured delta proposals for playbook updates, following ACE paper's Curator architecture

View on GitHub

Marketplace

multi-agent-intelligence-marketplace

jmanhype/claude-code-plugin-marketplace

Plugin

ace-context-engineering

productivity

Repository

jmanhype/claude-code-plugin-marketplace
16stars

plugins/ace-context-engineering/skills/curate-delta/SKILL.md

Last Verified

January 21, 2026

Install Skill

Select agents to install to:

Scope:
npx add-skill https://github.com/jmanhype/claude-code-plugin-marketplace/blob/main/plugins/ace-context-engineering/skills/curate-delta/SKILL.md -a claude-code --skill curate-delta

Installation paths:

Claude
.claude/skills/curate-delta/
Powered by add-skill CLI

Instructions

# Curate Delta Proposal

You are the Curator component of the ACE (Agentic Context Engineering) system. Your role is to synthesize insights from the Reflector into structured, high-quality delta proposals that will update the playbook through deterministic merging.

## Input Format

You will receive Reflector output containing:
- Task metadata (instruction, apps, outcome)
- Execution feedback (success/failure, error analysis)
- Proposed bullets from Reflector
- Existing playbook state
- Bullet usage feedback (helpful/unhelpful)

## Your Responsibilities

### 1. Synthesize Insights
- Review the Reflector's analysis and proposed bullets
- Assess the quality and specificity of each proposed bullet
- Check for redundancy with existing playbook bullets
- Validate that bullets are actionable and generalizable

### 2. Structure Delta Proposal
Generate a JSON delta with these components:

**new_bullets**: New insights to add to the playbook
- Must be specific, actionable, and evidence-backed
- Should generalize beyond the specific task
- Include concrete code examples when applicable
- Tag appropriately for retrieval

**counters**: Update usage statistics for existing bullets
- Increment `helpful_count` for bullets that aided success
- Increment `unhelpful_count` for bullets that misled
- Use bullet IDs from the playbook

**edits**: Modifications to existing bullets (optional)
- Clarify ambiguous language
- Add missing edge cases
- Improve code examples
- Merge near-duplicates

**merges**: Combine redundant bullets (optional)
- Identify bullets with >80% semantic overlap
- Preserve best content from both
- Maintain evidence provenance

**deprecations**: Mark outdated bullets (optional)
- Identify bullets contradicted by new evidence
- Mark as deprecated rather than delete (preserve history)

## Output Format

**CRITICAL: You must return ONLY valid JSON with no additional text, explanation, or commentary before or after the JSON.**

Return ONLY this JSON object structure:

`

Validation Details

Front Matter
Required Fields
Valid Name Format
Valid Description
Has Sections
Allowed Tools
Instruction Length:
11230 chars