Skip to content
Get packs

Guide · Agent skill

How to use the Continuous Learning v2 agent skill

Updated · written with AI from the project's own documentation and checked against our catalogue data

Continuous Learning v2 is an agent skill for Claude Code. It watches your sessions through hooks, notices patterns you repeat, such as corrections and error fixes, and saves each one as a small 'instinct' with a confidence score. This guide shows you how to install it, turn it on, use its commands, and decide whether it fits your work, so you can review what it learned and reuse it in later sessions.

Continuous Learning v2affaan-m/ECC · MIT

Quick start

  1. 1Download and unzip the skill. Download the ZIP from this page and unzip it into ~/.claude/skills/ or a project's .claude/skills/ folder. In the Claude apps, you can instead upload it under Settings → Capabilities → Skills.
  2. 2Enable the observation hooks. With a plugin install on Claude Code v2.1+, nothing more is needed. For a manual install, add the PreToolUse and PostToolUse hooks pointing to observe.sh in ~/.claude/settings.json.
  3. 3Work in a git repository. Start sessions inside a git project so the skill can detect the project and scope instincts correctly. Without a detected project, instincts go to global scope.
  4. 4Check what was learned. Run /instinct-status to see the instincts recorded for the current project and the global set, with their confidence scores.
  5. 5Review and curate. Use /projects to see instinct counts per project, and /promote to move patterns that apply across repositories into global scope.
  6. 6Evolve or share. Run /evolve to turn related instincts into skills, commands or agents, or use /instinct-export and /instinct-import to move your library between machines or people.

What the skill makes Claude Code do

Normally each Claude Code session starts with no memory of how you like your code written. This skill records prompts and tool use through hooks, then a background agent (Haiku, according to the skill) reads those observations and looks for recurring patterns: your corrections, error resolutions and repeated workflows.

Each pattern becomes an atomic instinct: one trigger, one action, a confidence score between 0.3 (tentative) and 0.9 (almost certain), a domain label such as code-style or testing, and the evidence that created it. Mature instincts can be grouped with /evolve into skills, commands or agents.

  • Instincts are scoped to the current project by default, so React habits stay out of a Python project.
  • A pattern is promoted to global scope with /promote when it appears in 2+ projects.
  • Universal rules, such as always validating input, can be shared globally.

Install continuous-learning-v2 in Claude Code

Download the ZIP of the skill from this page and unzip it into one of two places: ~/.claude/skills/ for all your projects, or a project's .claude/skills/ folder for one project only.

If you use the Claude apps instead, upload the skill under Settings → Capabilities → Skills.

  • Requirements: Claude Code (v2.1+ for automatic plugin hooks), a bash shell, and git available for project detection.
  • Optional: XDG or HOME directories for data storage.

Turn on the observation hooks

If you installed the skill as a plugin, no extra hooks block is needed: Claude Code v2.1+ loads the plugin's hooks/hooks.json automatically, and observe.sh is already registered there.

If you installed it manually in ~/.claude/skills, add the PreToolUse and PostToolUse hooks below to ~/.claude/settings.json. Both point to the observe.sh script.

{"hooks":{"PreToolUse":[{"matcher":"*","hooks":[{"type":"command","command":"~/.claude/skills/continuous-learning-v2/hooks/observe.sh"}]}],"PostToolUse":[{"matcher":"*","hooks":[{"type":"command","command":"~/.claude/skills/continuous-learning-v2/hooks/observe.sh"}]}]}}

Use the instinct commands

The skill provides six commands. Use /instinct-status to see what has been learned for the current project and globally, /evolve to group related instincts into skills, commands or agents, and /projects to list known projects and their instinct counts.

  • /instinct-status: show learned instincts (project and global)
  • /evolve: group related instincts into skills or commands
  • /instinct-export: export instincts to a file
  • /instinct-import: import instincts from others
  • /promote: promote project instincts to global scope
  • /projects: list known projects and their instinct counts
/instinct-status # Show learned instincts (project + global) /evolve # Group related instincts into skills/commands /instinct-export # Export instincts to a file /instinct-import # Import instincts from others /promote # Promote project instincts to global scope /projects # List all known projects and their instinct counts

Trigger the skill with natural requests

You do not need special syntax for the learning itself: once the hooks run, observations are recorded during your normal work. To work with the system, ask in plain language. The skill's own list of triggers covers requests such as configuring learning from sessions, reviewing or exporting instinct libraries, tuning confidence thresholds, evolving instincts into skills, and promoting project instincts to global.

  • "Show me what you have learned about this project."
  • "Export my instincts so I can share them with the team."
  • "Promote the instincts that also apply to my other repos to global."
  • "Turn my related testing instincts into a skill."

Understand where the data is stored

Observations are written to a file named observations.jsonl inside a per-project folder. The data directory is chosen in this order: the CLV2_HOMUNCULUS_DIR path if set to an absolute path, then $XDG_DATA_HOME/ecc-homunculus, then $HOME/.local/share/ecc-homunculus. The data is kept outside ~/.claude so Claude Code's sensitive-path guard does not block background writes.

Project detection uses CLAUDE_PROJECT_DIR first, then the git remote URL (hashed, so the same repo gets the same ID on different machines), then the repository path, and finally global scope if no project is found. Each project gets a 12-character hash ID.

Upgrade from v2.0 data

If you already have data in ~/.claude/homunculus from v2.0, migrate it once with the included script. Run it from the directory where the skill's folder sits, so the relative path resolves.

bash skills/continuous-learning-v2/scripts/migrate-homunculus.sh

What people use it for

  • Capturing a team's React or Django conventions so each new session follows them without repeating corrections.
  • Keeping Python habits out of a React project while still sharing universal rules like input validation across both.
  • Turning a cluster of testing or debugging instincts into a reusable skill or command.
  • Exporting an instinct library from one machine and importing it on another.

Frequently asked questions

Is the continuous-learning-v2 skill free and what is its licence?

The skill is listed under the MIT licence. The source does not mention pricing, so check the page you downloaded from for any other terms.

What do I need to run continuous-learning-v2?

You need Claude Code (v2.1 or later for automatic plugin hooks), a bash shell, and git available so the skill can detect your project. Using the Claude apps requires the skill to be uploaded under Settings → Capabilities → Skills.

Does the skill send my data over the network?

The listing states it does not use the network. It does run scripts, including the observe.sh hook and the migration script, and it stores observations locally in observations.jsonl.

Will it record my prompts and tool outputs?

Yes. The hooks capture prompts and tool use, and observations.jsonl stores them locally. The source does not describe any redaction, so avoid this skill on codebases where stored session data could expose secrets, or where local recording is not allowed.

Why are my instincts not showing up in another project?

Instincts are scoped to the project they were learned in by default. Use /promote to move a pattern to global scope, which the skill does when a pattern is seen in 2+ projects. Also check that the repo has a git remote or path so the project was detected.

How does it compare with other ways of teaching Claude Code my habits?

The source describes this skill as creating instincts automatically from observed sessions, with confidence scores and project scoping. Manually written instructions are deterministic, while learned instincts are probabilistic and can be wrong if confidence is misjudged, so review them with /instinct-status.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

Choose this skill if you use Claude Code daily, repeat the same corrections across sessions, and want project-specific habits captured and reviewable. Skip it if you do not want prompts and tool outputs stored locally, or if you work on sensitive or regulated code, or if you need fully predictable behaviour.

Download the skill