continuous-learning
by affaan-mcontinuous-learning automatically extracts reusable patterns from Claude Code sessions and saves them as learned skills. It uses a Stop-hook flow, supports configurable thresholds in config.json, and is best for skill authors, repo maintainers, and power users who want a practical continuous-learning guide for session-end capture.
This skill scores 68/100, which means it is listable but best presented with clear caveats. It has enough real workflow substance for Agent Skills Finder users to decide whether they need a Stop-hook-based session-evaluation flow, but it is narrower than a general-purpose skill and now has a preferred v2 path, so users should install it only if they want the simpler legacy approach or compatibility with older learned-skill workflows.
- Explicit triggerability: it tells agents to activate for Stop-hook session evaluation and includes the hook command pattern in the script comments.
- Operational detail: config.json and evaluate-session.sh show concrete defaults, thresholds, pattern categories, and learned-skills output location.
- Good install-decision context: the README text clearly states v1 is supported but v2 is preferred, helping users understand fit before installing.
- Dependency and setup friction: the script expects jq for config parsing and assumes Claude Code hook setup in ~/.claude/settings.json.
- Narrow scope and legacy status: this is specifically a continuous-learning v1 Stop-hook flow, not the preferred new-install path.
Overview of continuous-learning skill
What the continuous-learning skill does
The continuous-learning skill turns Claude Code sessions into reusable learned skills. It is designed for users who want the assistant to detect repeated patterns, useful fixes, and project-specific techniques at the end of a session, then save them for future reuse. If you want a continuous-learning skill that helps Claude improve from your own work history, this is the right kind of automation.
Who it fits best
This skill is best for Claude Code users who regularly solve similar problems across sessions and want those solutions captured automatically. It is especially useful for skill authors, repo maintainers, and power users who care about preserving debugging patterns, workarounds, and project-specific conventions without manually writing every lesson down.
What makes it different
The core difference is that this continuous-learning skill is Stop-hook based, not prompt-only. That means it evaluates the session once at the end instead of trying to inspect every message in real time. It is simpler, lighter, and easier to reason about, but it is also narrower than newer approaches like continuous-learning-v2, which the repo treats as the preferred path for new installs.
How to Use continuous-learning skill
Install and place the hook
Install the continuous-learning skill into your Claude skills directory, then wire it to the Stop hook so it runs when a session ends. The repo’s script expects to live under ~/.claude/skills/continuous-learning/ and writes learned output to ~/.claude/skills/learned/. A typical install decision is less about “can I add it?” and more about whether you want session-end extraction as part of your normal Claude workflow.
Start with the right inputs
The skill works best when the session contains enough substance to extract a pattern. The default minimum session length is 10 messages, so very short chats usually will not produce useful learned skills. If you want strong continuous-learning usage, feed it sessions that include a real problem, a correction, a workaround, or a repeated technique—not just a single answer.
Read these files first
For practical setup, read SKILL.md first, then config.json, then evaluate-session.sh. That order tells you what the skill does, what can be tuned, and how the Stop hook is implemented. config.json is the key file if you want to change thresholds, output location, or the pattern categories it looks for.
Shape prompts for better extraction
If you are using this continuous-learning guide as part of your own workflow, make the session explicit enough that the hook can detect reusable behavior. Good input looks like: “I need a repeatable way to debug failed installs in this repo; capture the steps and the final fix as a reusable pattern.” Weak input looks like: “help me with this.” The first gives the evaluator something stable to learn from; the second often leaves no durable pattern.
continuous-learning skill FAQ
Is this the right skill if I want automatic learning?
Yes, if your goal is to automatically extract reusable patterns from Claude Code sessions and store them as learned skills. If you want a continuous-learning skill that quietly improves future sessions from past ones, this matches that job well. If you want a more proactive or instinct-based system, the repo itself points you toward continuous-learning-v2.
How does it compare with a normal prompt?
A normal prompt can describe what you want, but it does not persist anything after the session ends. This skill adds a workflow layer: it watches for reusable patterns, then saves them to the learned-skills directory. That makes it more useful for repeated team or repo-specific work than a one-off instruction.
Is it beginner-friendly?
Moderately. The logic is simple, but the setup requires understanding Claude Code hooks, session boundaries, and where learned skills are stored. Beginners can use it if they follow the provided files closely, but they should not expect it to behave well without the hook wiring and config in place.
When should I skip it?
Skip this continuous-learning install if you only want ad hoc help, if your sessions are usually short, or if you do not want automated extraction writing to your local skills folder. It is also a weaker fit if you already know you want the newer continuous-learning-v2 path.
How to Improve continuous-learning skill
Tune the config before judging results
The biggest lever is config.json. Raise or lower min_session_length based on how long your useful sessions really are, and adjust extraction_threshold if the skill is being too conservative or too eager. If you care about specific pattern types, keep patterns_to_detect focused on the ones that matter most to your work.
Give the hook clearer pattern signals
The skill improves when sessions contain explicit corrections, repeated debugging steps, or a named workaround. For example, “The first fix failed because the issue was actually path resolution; capture that distinction” is much better than a generic success message. Stronger inputs make the extracted learned skill more actionable and less generic.
Watch for common failure modes
The main failure mode is overlearning trivial changes such as typos, one-off fixes, or external API issues the skill should ignore. Another is saving patterns that are too project-specific to reuse elsewhere. If you see noisy output, reduce what you ask the model to treat as a reusable lesson and narrow the session to the actual decision or technique.
Iterate from the saved output
After the first run, review what appears in ~/.claude/skills/learned/ and ask whether each extracted item would help in a future session. If not, tighten the pattern criteria or change the way you describe problems during the session. That feedback loop is what makes continuous-learning for Skill Authoring actually useful: better source sessions produce better reusable skills.