You only need these agent skills

Most coding agents are capable. What they lack is a shared playbook — the same planning habits, research discipline, and debug trail every time. That is what agent skills are for: small, sharp instructions your agent loads when the moment fits.

This post is the map. Install the collection once, then reach for the skill that matches the job — and skip the ones that do not.

Install

Add the whole collection to your agent:

npx skills add ravid7000/skills

Prefer a versioned npm install?

skills install npm:@ravid7000/skills

You can also install a single skill with --skill <slug>. The skills hub has copy-ready commands and links into each detail page.

Why they exist

Without skills, every session re-negotiates process: how to plan, what “done” means, whether to guess from code or gather evidence, how to leave work for the next agent. Skills turn those into defaults the agent already knows how to follow — so you spend less time re-explaining and more time deciding.

Workflow

plan-with-me

Turns a vague request into an agreed, written plan by asking one grounded question at a time.

Reach for it when the request is thin, several designs are defensible, sequencing matters, or you want a plan / RFC before anyone writes code.

Skip it when “done” is already clear and you are ready to implement.

handing-off-work

Writes a handoff a fresh agent can resume from — decisions, dead ends, and verified code state.

Reach for it when the session is ending unfinished, context is about to compact, work is moving to another harness or person, or you are resuming from a handoff and need to know how far to trust it.

Skip it when the work is finished and you only need a PR description.

Research

finder

Answers research questions from current, cited sources instead of model memory.

Reach for it when you need researched, sourced, up-to-date answers — best practices, docs, community guidance, or an error you should not invent history for.

Skip it when the lookup is trivial, or you already supplied the context and only want code.

Diagnostics

debug-flow

Traces broken flows across any stack with temporary correlated logs, then tears them out.

Reach for it when a flow is broken (UI, API, CLI, job), layers are unclear, the agent is guessing from code, and you will reproduce locally for a temporary trail.

Skip it when you need permanent production logging, metrics, or tracing — that is observability, not this-bug instrumentation. Also skip if there is no planned repro.

instrumenting-for-observability

Adds logging, metrics, and tracing designed backwards from the questions an outage will ask.

Reach for it when a path is about to ship or change how it can fail, you are reviewing whether a diff is diagnosable at 3am, or a retro found the outage was invisible.

Skip it when you only need temporary logs for this bug — use debug-flow instead.

Meta

creating-agent-skills

Authors and validates new skills so they load reliably and survive review.

Reach for it when you are adding, editing, or reviewing a skill in this collection.

Skip it when you are only installing or consuming skills, or writing domain content without the packaging rules.

A simple path

1. Start vague work with plan-with-me.
2. Before the session dies, leave a handing-off-work document.
3. Need facts, not vibes? finder.
4. Broken flow with a repro? debug-flow. Shipping something that must be diagnosable later? instrumenting-for-observability.
5. Extending the collection itself? creating-agent-skills.

You do not need every skill on every turn. You need the right one for the moment.

Where next

Browse the full catalog and install panel on the skills page, or open the source on GitHub.