CO

context-engineering-advisor

Helps distinguish between volume-based context stuffing and structure-based context engineering.

Install

mkdir -p .claude/skills/context-engineering-advisor && curl -L -o skill.zip "https://agentskills.codes/api/skills/download/13814" && unzip -o skill.zip -d .claude/skills/context-engineering-advisor && rm skill.zip

Installs to .claude/skills/context-engineering-advisor

Activation

This is the description your AI agent reads to decide when to run this skill — the better it matches your request, the more reliably it fires.

Diagnose context stuffing vs. context engineering. Use when an AI workflow feels bloated, brittle, or hard to steer reliably.
125 chars✓ has a “when” trigger
Advanced

Key capabilities

  • Identify context boundaries in AI workflows
  • Diagnose context stuffing markers like reflexive context window expansion
  • Implement tactical practices like bounded domains and episodic retrieval
  • Apply the Research→Plan→Reset→Implement cycle to manage context rot
  • Evaluate context efficiency using a quantitative framework
  • Design information architecture for AI grounding

How it works

The skill guides product managers in diagnosing context stuffing versus context engineering by identifying markers, applying core principles, and using diagnostic questions. It proposes a memory architecture and a Research→Plan→Reset→Implement cycle to manage context.

Inputs & outputs

You give it
AI workflow feeling bloated, brittle, or hard to steer
You get back
Diagnosis of context stuffing vs. context engineering and actionable recommendations

When to use context-engineering-advisor

  • Fixing brittle AI workflows
  • Optimizing context usage
  • Architecting information structures

About this skill

Context Engineering Advisor

Use this skill when an AI workflow is expensive, brittle, vague, retry-heavy, or overloaded with persistent instructions and source material.

Core Contract

Treat model attention as scarce. Optimize completed-task quality per token, not prompt size in isolation.

  1. Define the decision or deliverable before selecting context.
  2. Persist only product identity, non-negotiable safety rules, durable preferences, and recurring constraints.
  3. Retrieve episodic evidence just in time.
  4. Load one workflow or skill directly; do not create a router-of-routers.
  5. Preserve raw evidence outside active context and reference it by stable path or ID.
  6. Compact only at completed phase boundaries and retain decisions, exact language, citations, unresolved risks, verification state, and the next action.
  7. Measure the whole trajectory, including turns, retries, cache traffic, tool payloads, and output.

Diagnostic

For every context element, ask:

  • What decision does it support?
  • Can retrieval replace persistence?
  • Who owns its inclusion boundary?
  • What concrete failure occurs if it is omitted?
  • Is more context masking conflicting or poorly structured source material?

Exclude an element when no concrete failure can be named. Reconcile conflicting sources instead of passing all versions downstream.

Architecture Pattern

Use four layers:

  • Working context: the current goal, direct workflow, and up to five relevant sources.
  • Addressable archive: full tool results, transcripts, and raw evidence retrievable by ID.
  • Compiled knowledge: current navigation pages that cite raw sources and hashes.
  • Durable state: decisions, priorities, workstreams, and open loops.

Compiled knowledge is not authoritative for exact quotes, commitments, legal language, security findings, or final citations; retrieve the raw source for those.

Phase Cycle

research → plan/checkpoint → reset or compact → implement → verify

Research may be noisy. The checkpoint must be dense and structured. Implementation should start from that checkpoint plus only the evidence needed for the next phase.

Response Profiles

  • compact_operator: terse progress and tool narration.
  • artifact: complete polished deliverable.
  • verbatim: exact source or stakeholder wording.

Use compact_operator during execution. Do not shorten final artifacts or exact-language outputs merely to save tokens.

Supporting Material

Read the full facilitation and diagnostic guide only when running the interactive workshop or when a detailed example is required. Use workshop-facilitation for interactive pacing.

When not to use it

  • When the task is about prompt writing
  • When the task is about choosing AI tools like Claude vs. ChatGPT

Limitations

  • The skill is not about prompt writing
  • The skill is not about choosing AI tools
  • The skill focuses on designing information architecture

How it compares

This skill focuses on designing the information architecture for AI workflows to prevent context hoarding and improve AI attention allocation, contrasting with simply increasing context window size or retrying until success.

Compared to similar skills

context-engineering-advisor side by side with the closest alternatives in the catalog.

SkillInstallsUpdatedSafetyDifficulty
context-engineering-advisor (this skill)03moNo flagsAdvanced
prompt-lookup77moNo flagsBeginner
thought-based-reasoning15moNo flagsAdvanced
profile-me01moReviewBeginner

Try saying

Example prompts that trigger this skill in your AI assistant.

Search skills

Search the agent skills registry