Skip to main content

Assistant Workflows

ContextCrumb is designed for agent context loading. The core habit is:

Compress prose-heavy context before placing it into the model context.

The skill instructions focus on file loading because that is the easiest behavior to teach an assistant with a SKILL.md. Developers can also use ContextCrumb deeper in the stack: prompt compression, history compression, subagent output compression, and tool-output compression.

Who This Is For

This section is for people using coding assistants, research agents, MCP clients, or editor-integrated tools. If you are adding ContextCrumb to an application or pipeline, start with the Python API or Local Service.

Use ContextCrumb for large natural-language files and supported code files:

contextcrumb load path/to/file.md
contextcrumb load path/to/file.py

For supported Python, JavaScript, TypeScript, JSX, TSX, Go, and Rust files, default auto mode preserves executable source exactly and compresses only comments/docstrings.

Do not use compressed output as the only source for exact syntax:

  • Unsupported source code, or exact source edits
  • Config files
  • Diffs
  • Schemas
  • Commands
  • Legal or policy text

For structured tool results, preserve the structure and compress only prose values. For example, compress a long body or description field, but keep ids, URLs, scores, timestamps, and schema fields raw.

Common Prompts

Ask an agent:

Use ContextCrumb to compress the project notes before you use them as context.

Or:

Before reading the full transcript into context, run contextcrumb load on it and work from the compressed output.

For safer code-adjacent use:

Use ContextCrumb for supported source files in code-comments mode. Load commands, configs, and unsupported code raw.

Agent Integration Choices

IntegrationUse when
Skill filesYour agent supports repo-owned or installed skills
MCP serverYour agent can call MCP tools
MCP shrink proxyTool catalogs are verbose before any tool call happens
Local serviceYou want one warm model shared by multiple agent calls

Why This Helps Agents

  • Reduces context spent on filler wording
  • Keeps the original sequence of the document
  • Avoids inventing summaries before the agent understands the task
  • Lets agents process more source material before hitting context pressure
  • Provides stats for how much context was saved