pivoshenko.ai

Overview

This repository is managed with Kasetto. It pulls personal and external skills, MCPs, and instructions from upstream repositories into one synced setup. The full source list lives in kasetto.yaml.

What's in here?

• Locally authored skills, see skills/
• Locally authored MCP definitions, see mcps/
• Locally authored instructions (CLAUDE.md / AGENTS.md / .cursor/rules ... rule fragments), see instructions/
• External skills, MCPs, and instructions, see kasetto.yaml
• Retired skills and instructions, kept for reference but no longer synced, see archive/

Main Principles

• Minimalism: keep only skills and MCPs that are used daily
• Consistency: same conventions across all locally authored skills
• Composability: skills are small, single-purpose, and chainable
• Source of truth: upstream skills are pulled, not forked, so updates stay free

Installation

1. Install Kasetto, see the installation guide
2. Sync skills and MCPs into your local Claude Code config. Either run:

kst sync --config https://github.com/pivoshenko/pivoshenko.ai/blob/main/kasetto.yaml

Or add the source to ~/.config/kasetto/config.yaml and then run kst sync:

source: https://github.com/pivoshenko/pivoshenko.ai/blob/main/kasetto.yaml

Rules

Reusable agent rules, the behavioral guardrails that aren't project-specific, live in instructions/, one Markdown file per rule. Kasetto distributes them as its instruction asset kind: each is transformed into the target agent's native instruction file (CLAUDE.md, AGENTS.md, .cursor/rules, ...) and merged in as a managed block, so hand edits and other rules survive a re-sync.

The instructions/ files are the source for my global rules. Kasetto syncs them into ~/.claude/CLAUDE.md as managed blocks. The old single CLAUDE.md in pivoshenko.dotfiles was dropped once these took over, so dotfiles no longer deploys it.

Archive

archive/ holds what used to be part of the synced setup and got retired. Kasetto only pulls from skills/, mcps/, and instructions/, so moving something here takes it out of the agent config while keeping it readable in git. The site lists it in a separate archived section instead of dropping it silently.

Currently parked:

• Personal wiki (second brain): an Obsidian vault maintained primarily by Claude Code, inspired by Karpathy's LLM-maintained notes. The workflow skills (wiki-capture, wiki-summarize, wiki-project, wiki-write, wiki-lint) and the hourly vault-backup script (vault-snapshot.sh) live here now. The vault and its own schema (CLAUDE.md / INDEX.md / LOG.md) are untouched. Only the automation is off
• Memory rule (memory.md): routed each ~/.claude/projects/<slug>/memory into the vault's 97 MEMORY/<project>/, so agent memory lived in the wiki. Retired together with the wiki
• Co-authored attribution rule (co-authored-attribution.md): banned Co-Authored-By trailers on commits and PRs. Retired as no longer needed
• spec-dispatch (spec-dispatch): ran an OpenSpec change's task queue as parallel agent waves. Replaced by git-spec-dispatch, which does the same wave planning against a GitHub issue tree instead of openspec/tasks.md. The openspec CLI was installed but no repo here ever adopted it, so the skill had never actually run; issues were already the tracking surface, and keeping a second one meant maintaining two copies of every task list

Reason for the wiki and the memory rule: they didn't visibly change how the agents worked day to day, so the upkeep wasn't paying for itself. The obsidian-markdown syntax reference stays active. It's useful for Markdown work with or without a vault.