@dzhechkov/skills-reasoning

Generic reasoning & code-quality skill pack for Claude Code — 4 stack-neutral agentic skills that improve how an agent thinks, investigates, and writes code. Depersonalized from a curated skills archive; zero product coupling (no vendor, framework, or repo specifics). Part of the dz-harness-hub ecosystem.

Install

# with the dz CLI (recommended)
dz init --target claude-code --preset reasoning
# or install one skill
dz init --target claude-code --select karpathy-guidelines

# or plain npm (skills are copied into your .claude/skills/)
npm install @dzhechkov/skills-reasoning

The 4 skills

| Skill | When to use | What it does | |-------|-------------|--------------| | investigate | You're asked to investigate/research/understand a problem without fixing it | Systematic search for root cause → a diagnosis document, not a diff. Enforces "INVESTIGATE ONLY", prizes live reproduction over code-reading. Investigation and implementation are separate activities. | | solid | Writing, refactoring, designing, reviewing, or testing code | Operate as a senior engineer: SOLID principles, TDD (Red-Green-Refactor, non-negotiable), clean-code practices. Includes a references/ library. | | karpathy-guidelines | Writing/reviewing/refactoring code, to avoid common LLM mistakes | Behavioral guardrails (from Karpathy's LLM-coding observations): think before coding, simplicity first, surgical changes, surface assumptions, define verifiable success criteria. | | agents-md-creator | You want per-repo AGENTS.md files for a multi-repo project | Generates the always-on architectural skeleton an agent reads every session: dependency direction, package responsibilities, forbidden patterns, invariants, change checklist. |

Looking for systematic-debugging? It already ships in @dzhechkov/skills-qe — we don't duplicate it here.

Why these are separate from the pipeline

These are installable reasoning skills — an agent invokes them by description or you name them. They are deliberately not wired into any pipeline stage, so they compose with any workflow (feature-adr, plain coding, review). The code-quality bar that is project-calibrated (a critic / implementer bar tied to your architecture/vision.md) lives in @dzhechkov/skills-feature-adr as the critic / impl-bar roles — this pack is the stack-neutral reasoning layer beneath it.

Provenance

Canonicalized 2026-07-15 from a curated agentic-skills archive and depersonalized (all vendor/framework/repo specifics removed — verified by a zero-coupling grep gate). Individual skills embody public methodologies (SOLID, TDD, Karpathy's observations). See sources.json.