agent-skills

Personal collection of agent skills, installable via the skills CLI.

Install

# List available skills
npx skills add Mark-Life/agent-skills --list

# Install all skills for Claude Code
npx skills add Mark-Life/agent-skills -a claude-code

# Install a specific skill
npx skills add Mark-Life/agent-skills --skill session-report

# Update later
npx skills update

Pin to a tag/branch:

npx skills add https://github.com/Mark-Life/agent-skills/tree/v1.0.0

Skills

Skills are grouped into catalog folders under skills/<group>/.

Coding Workflows

Skill Description
product Product taste for shaping what to build, grown one principle at a time. Currently: boil the ocean (scope the ambition to what agents made possible), the primitive is the product (name the capability others compose on; schema, errors and guarantees are the surface), the work announces itself (start from a signal, not a prompt; reversibility draws the autonomy boundary), and grow a second nervous system (declare capabilities for machines, discover them lazily, and expose a core beneath the experience — agents are how humans find you).
new-project Scaffold a brand-new project from the personal Next.js monorepo template (Mark-Life/netxjs-monorepo) and run the standard bootstrap.
observability Instrument a TypeScript service so production is answerable: one wide event per request, job or run, high-cardinality fields, closed outcome unions, two-tier span names, bounded metric tags, config-gated OTel export, and tail sampling. Owns logging, tracing and metrics decisions.
media-cdn Put a local file on a media-cdn-service deployment and hand back a public URL — one zero-dependency CLI for upload, list, publish/unpublish, download, and whoami, credentials from ~/.agents/.env.local or --env <path>.
pr-issue Write PR, issue, and ticket titles and bodies in Vim's area: summary style — Problem:/Solution:/Security Impact:/Testing: for PRs, Problem:/Reproduction:/Expected vs actual:/Proposed solution: for issues.

Communication

Skill Description
agent-to-human Write the final human-facing output — chat reply, summary, status update, PR/issue/commit message — to respect the reader's attention: answer first, cut to the decision, receipts only, plain words. Source of truth for output style.
human-to-agent Write instructions an agent will execute — prompts, skills, CLAUDE.md/AGENTS.md rules, tickets — for predictable behaviour: prompt the positive, leading words, cut no-ops, checkable completion criteria.
writing-for-readers Write published prose — articles, blog posts, course lessons, landing page copy — in a voice that reads as a person explaining something out loud: what human prose does (first person, admitted limits, named specifics), measured paragraph- and sentence-length variance, twelve named tells to cut (aphorism pair, negate-then-correct, sincerity words, rule of three, em dashes…), and a mechanical pass over the finished draft.

Inspect & Debug

Skill Description
session-report /session-report <id> — generate a self-contained HTML report debugging a Claude Code (or Codex) session's context window: token budget, thinking cost, dumb-zone cutoff, full history.
memory-view /memory-view [project] — open a self-contained, secret-redacted HTML explorer of Claude Code's per-project auto-memory. View, search, and browse only — no edits, no curation.
context-doctor /context-doctor — audit and shrink the fixed context loaded every session (tool/MCP definitions, plugins, skills, subagents, memory/rules) by pruning, gating, or routing what's loaded but unused.
harness-doctor /harness-doctor [days] — the fleet view across every project's transcripts, not just one session: a zero-dependency script mines the last N days into five fact tables (sessions, tools, bash, user messages, errors); the model digs into them and returns a ranked, copy-pasteable list of fixes for wasted wall-clock, tokens, and repeated corrections.

Repository layout

skills/<group>/<name>/  # SHIPPED. `skills add` copies the whole skill folder to the user.
tests/<name>/           # NOT shipped. Unit tests + fixtures.
tests/guard.test.ts     # CI guard: skills/ must hold only runtime artifacts.
skills.sh.json          # display grouping on skills.sh (no effect on the CLI)

Skills live in catalog layout (skills/<group>/<name>/SKILL.md). The skills CLI resolves --skill <name> by the name field in frontmatter, not by path, so the grouping folders don't affect installs.

npx skills add copies a skill's entire folder to the end user (it only strips .git, __pycache__, __pypackages__, metadata.json). So tests, fixtures, and dev tooling must live in tests/, never inside skills/<name>/. See CLAUDE.md for the full convention.

bun test        # run unit + guard tests

Adding a new skill

  1. mkdir -p skills/<group>/<name> && cd skills/<group>/<name>

  2. Create SKILL.md with frontmatter:

    ---
name: <name>
description: One sentence describing when the agent should use this skill.
version: 1.0.0
---

# <Name>

Instructions for the agent.

  3. Keep the folder lean — runtime files only. Put tests and fixtures in tests/<name>/, not in the skill folder.

  4. Add a row to the table above (under the matching group) and list it in skills.sh.json.

  5. Commit and push.

Credits

Author

Written by Andrey Markin (@Mark-Life).

License

MIT © Mark-Life Ltd and Andrey Markin.

The licence lives at the repo root and covers every skill here. npx skills add copies a skill folder into your project without it, so if you redistribute a skill, carry this notice with it.