prunegraft

Managing a growing library of Claude Code skills eventually leads to token bloat and context degradation. Prunegraft solves this by analyzing session history to identify which skills are earning their place and which are wasting resources. It automates the maintenance of your skill folder, ensuring every instruction set is either actively used, archived, or evolved based on actual developer behavior.

What it does

• Usage tracking identifies which skills are frequently invoked and which have sat idle across sessions.
• Automated archiving moves unused skills to a retired directory to save tokens without deleting code.
• Pattern recognition flags manual tasks repeated across multiple sessions as candidates for new skills.
• Dependency mapping detects when skills build on each other to prevent redundant instructions.
• Continuous improvement logs failures and "gotchas" to refine existing skill logic based on real-world errors.

How it works

1. Analyze sessions by running the internal script to scan Claude Code logs and generate a usage report.
2. Execute pruning by retiring idle skills or explicitly marking essential ones to be kept based on report findings.
3. Graft new skills when a manual workflow repeats three or more times and requires complex steps.
4. Refine instructions by updating logic when a skill fails or a new edge case is discovered in the session history.

Frameworks & tools

This skill uses a standalone Node.js script that interacts directly with ~/.claude/ session records. It requires no external npm packages and can be executed via standard input.

Why this beats prompting it yourself

Manual cleanup is tedious and prone to accidental deletion of edge-case tools. Prunegraft uses a data-driven approach to prove a skill's value before keeping it, and it enforces a "one procedure, one home" architecture that prevents context-window bloat.

Use cases

• Token optimization by retiring skills that haven't been touched in months.
• Workflow automation by turning a hard-won recurring fix into a permanent skill.
• Skill auditing to see exactly how your agentic tools are interacting.
• Knowledge recovery by restoring retired skills when they are requested again.

Known limitations

Does not delete files permanently. Requires local Node.js to run the analysis script. Only detects patterns within the stored Claude session history.