context-engineer

name

context-engineer

description

Single entry point for authoring and curating Copilot context — create/update skills, agents, instructions, prompts, and project memory; consolidate overlapping customizations; debug discovery; and build compact DIGEST/RAW knowledge. Use when: create or review SKILL.md, .agent.md, .instructions.md, .prompt.md, AGENTS.md, or .github/memory; design routing, handoffs, applyTo patterns, tool restrictions, progressive disclosure, prompt structure, context curation, or Hierarchical RAG.

Context Engineer

Consolidated skill for creating and curating Copilot context artifacts (skills, agents, instructions, prompts) and memory. Keep SKILL.md as the compact core; load one references/<domain>/ guide only when the task needs that depth.

Output style

Apply the caveman skill to every user-facing response (default full) unless the user says stop caveman / normal mode. Keep artifact content, YAML frontmatter, file paths, applyTo globs, and generated markdown in normal precise form — never caveman. Relax caveman for irreversible-action confirmations (e.g. deleting artifacts), then resume.

Capabilities

Discrete units of work this skill owns (map a spec/change to one):

1. skill-creator — create/update/review/package SKILL.md (structure, frontmatter, progressive disclosure, bundled references/assets/scripts).
2. agent-creator — create/update .agent.md (structure, workflow, domain knowledge, handoffs, tool/visibility flags).
3. instructions-creator — create/update *.instructions.md / copilot-instructions.md (applyTo patterns, instruction hierarchies).
4. prompt-creator — craft/improve/debug .prompt.md and LLM system prompts (few-shot, chain-of-thought, metaprompts).
5. context-curator — analyze-first curation of .github/.copilot customizations and memory (context rot, drift, overlap, stale entries).
6. skillopt-curator — SkillOpt-style bounded edit loop to tighten/consolidate .github after a session.
7. hierarchical-rag — DIGEST/RAW two-tier knowledge structure for context compression.

Orientation

• Analyze-first for curation: propose a plan and only execute destructive/broad changes on explicit confirmation.
• Metadata standard: every artifact carries a description with trigger keywords, plus source/date where useful; the folder name must match the frontmatter name (see metadata-standard.instructions.md, auto-loaded).
• Progressive disclosure: keep a compact SKILL.md core and push depth into references/; the same principle these guides teach applies to the guides themselves.
• Project memory is plain Markdown under the active project's .github/memory/; MCP servers are for external evidence only.

Lazy reference loading

Do not read references/ up front. Classify the task, then open the smallest matching guide (usually one). Each domain guide lives at references/<domain>/<domain>.md; larger domains keep detailed files in that folder's references/, scripts/, assets/, or agents/ subfolder.

Task signal	Load only
Create/update/review/package a SKILL.md skill	references/skill-creator/skill-creator.md
Create/update a .agent.md agent	references/agent-creator/agent-creator.md
Create/update *.instructions.md / copilot-instructions.md	references/instructions-creator/instructions-creator.md
Create/improve/debug a prompt or system prompt	references/prompt-creator/prompt-creator.md
Analyze/curate .github/.copilot customizations + memory	references/context-curator/context-curator.md
Tighten/consolidate .github after a session (SkillOpt)	references/skillopt-curator/skillopt-curator.md
Build/maintain a DIGEST/RAW Hierarchical RAG structure	references/hierarchical-rag/hierarchical-rag.md

Reference map: inside a domain guide, an instruction to "load the ai-<x> skill" now means read references/<x>/ in this skill (drop the ai- prefix, e.g. ai-skill-creator → references/skill-creator/).

Workflow

1. Understand — clarify which artifact scope is affected (skill / agent / instruction / prompt / curation / memory).
2. Classify — map the task to one capability above; load the matching guide only if needed.
3. Execute — follow the metadata standard; analyze-first for curation, apply broad/destructive changes only on explicit confirmation.
4. Validate — folder name matches name; description has trigger keywords; references resolve.
5. Align — present result; confirm before irreversible actions.