Context Map

Map what the model can see.

Context Map is a no-tool prompt pattern for asking an AI agent to show its current context as a readable tree. It helps users inspect what is actually loaded, what is protected, what is only summarized, what is metadata, what is only a pointer, and what is inferred from visible context.

It does not judge whether the context is good or enough. It exposes the current context stack so a human can decide.

Why

Agent behavior often looks mysterious because the visible conversation is not the whole context. A model may also carry:

• system and developer instructions
• tool schemas and runtime metadata
• memory summaries
• workspace instructions such as AGENTS.md or CLAUDE.md
• environment context
• prior conversation history
• tool results that already appeared in the current thread
• the latest user request

Before asking the agent to act, ask what it can see.

Quick Start

Option A: Send the link

Share this repo:

https://github.com/dxxbb/context-map

If the user's agent can read URLs or GitHub repositories, they can ask:

Open https://github.com/dxxbb/context-map and use Context Map.
Do not call tools for the audit itself.

This is convenient, but it depends on whether that agent can access the link.

Option B: Copy the prompt

The most reliable option is to copy the prompt into the chat:

prompts/context-map.md

This works even when the agent cannot browse GitHub or install skills.

Option C: Install as a skill

If the agent supports SKILL.md, install the skill folder:

codex-skill/context-map/

Then ask:

Use Context Map. Do not call tools. Show the context you can currently see.

Prompt

Use the prompt in prompts/context-map.md.

Codex Skill

A shareable Codex skill is available in:

codex-skill/context-map/

Install by copying that directory into your Codex skills folder, then trigger it with a request such as:

Use Context Map. Do not call tools. Show the context you can currently see.

Output Shape

CONTEXT MAP
Model-visible context

Legend
  [loaded]    body visible
  [protected] visible but not quotable
  [summary]   summary / compacted context
  [metadata]  schema / runtime / catalog / environment metadata
  [pointer]   name/path/URL only
  [inferred]  derived from visible context

└─ Context Stack
   ├─ 01 SYSTEM [protected]
   │  ├─ 01.1 Runtime Policies [protected]
   │  │    - browsing rules
   │  │    - copyright boundaries
   │  │    ↳ original text not quotable
   │  │
   │  └─ 01.2 System Tools [metadata]
   │       - web.search_query / web.open / web.find
   │
   └─ 02 CURRENT REQUEST [loaded]
      └─ 02.1 Explicit Ask [loaded]
         - latest user request

Status Labels

• [loaded]: the body is visible in the current context.
• [protected]: the body is visible but cannot be quoted verbatim.
• [summary]: the section is explicitly injected as a summary or compacted context.
• [metadata]: runtime, schema, catalog, environment, or configuration metadata.
• [pointer]: only a name, path, URL, title, or catalog entry is visible.
• [inferred]: derived from visible context rather than directly stated.

Design Principles

• No tools by default.
• Do not reveal protected system/developer instructions verbatim.
• Do not treat a path as loaded body.
• Do not treat a scoped instruction label as proof of a physical file.
• Keep the whole output as one continuous tree.
• Preserve indentation with a fenced text code block.
• Follow the user's language; keep tool names and technical identifiers intact.

Screenshot

For sharing, use a real Context Map screenshot and redact sensitive information. Do not use a generated illustration unless you are explaining the concept rather than showing the tool.