Agent Skills

A skill is a folder. Inside that folder is a SKILL.md with YAML frontmatter on top and plain markdown below. Optionally, it can ship scripts, references, and assets.

Skills solve a specific problem: how to give Claude deep procedural knowledge without burning context on every turn. They load in three levels.

Progressive disclosure

The trick: at session start, Claude knows every skill exists and what each one does, but has loaded almost nothing. When a user prompt matches a skill's description, Claude reaches for it. You pay the full token cost only when the skill is needed.

Where skills live

~/.claude/skills/              # your personal skills, every project
.claude/skills/                # shared with the team, committed to git
{plugin}/skills/               # skills bundled inside a plugin

Inside skills/ each skill is a folder containing at least SKILL.md.

~/.claude/skills/
└── sf-apex/
    ├── SKILL.md
    ├── references/
    │   └── rubric.md
    └── scripts/
        └── score.py

That is the exact shape Jag's sf-apex ships with: a SKILL.md body, a rubric in references/, and scoring scripts the skill invokes.

Install the library first

Two commands, two paths. Pick the one that matches your tool.

npx skills add Jaganpro/sf-skills

That gets you the 36 skill files for Cursor, Codex, Gemini CLI, or Claude Code without local hooks.

For the full Claude Code experience (skills + 7 specialist subagents + PostToolUse validation hooks + LSP auto-fix loops on Apex / LWC / Agent Script + Code Analyzer):

curl -sSL https://raw.githubusercontent.com/Jaganpro/sf-skills/main/tools/install.sh | bash

Restart Claude Code afterwards. See Using Jag's SF Skills for prerequisites, update commands, and the full skill-by-skill reference.

Slash commands — when and why to invoke each

Every skill in Jag's library is exposed as a slash command. Claude will pick one automatically when your prompt matches the skill's description, but typing /sf-apex (or whichever) forces it to load, which is what you want when you're about to do real work.

Type /sf- in Claude Code to see the full list. Below is the working mental model: which command to reach for, and why.

Development and query skills

These are the skills that run every day. Each one ships with a scoring rubric so output is reproducible across the team.

When to skip them: one-line refactors, throwaway console scripts, or anything where the rubric overhead is noise.

Quality and debugging skills

These pair with the development skills. Invoke them when you're closing out work, not starting it.

Foundation skills

The declarative layer. Invoke when you're touching XML or config rather than code.

Integration skills

Invoke when you're standing up auth or callouts to an external system.

AI and Agentforce skills

This cluster is the reason most SEs install the library right now. Pick based on which artefact you're touching.

Data 360 skills

Data 360 work is phased: Connect → Prepare → Harmonize → Segment → Act, with Retrieve cutting across. There is an orchestrator skill and one skill per phase. Use the orchestrator when the task spans phases; use a phase skill when it doesn't.

Industries (OmniStudio + CME) skills

Invoke the phase skill for authoring, the analyze skill before you refactor.

Planning, DevOps, and diagram skills

The connective tissue around delivery.

The decision tree

Before you invoke anything, ask three questions.

1. What file am I touching? The file type usually tells you which skill to pick. *.cls → /sf-apex. *.flow-meta.xml → /sf-flow. *.agent → /sf-ai-agentscript. The hook system routes automatically on Write/Edit, but typing the slash command up front gives Claude the rubric on turn one instead of turn three.
2. Am I authoring, reviewing, or analyzing? Authoring and reviewing use the same skill (the rubric cuts both ways). Analyzing existing work — especially in OmniStudio — uses the analyze skill first. /sf-industry-commoncore-omnistudio-analyze before you touch anything vlocity-namespaced.
3. Is this one artefact or a pipeline? For Data 360 and any multi-artefact job, use the orchestrator (/sf-datacloud) rather than chaining phase skills. The orchestrator understands the hand-offs.

A minimal SKILL.md (for when you do write your own)

Most of the time you won't need to. But if you want a team-specific skill on top of Jag's library, this is the shape.

---
name: acme-handoff
description: Our handoff doc conventions. Use when the user asks about docs/handoff.md, customer transition, or end-of-POC paperwork.
---

You generate or update docs/handoff.md for Acme POCs.

When invoked:
1. Read references/handoff-template.md for the required sections.
2. Cross-reference force-app/main/default/ for what was actually built.
3. Summarise the last 20 git commits as a changelog.
4. Write docs/handoff.md. Do not paraphrase the template sections; include them verbatim.

If the user asks for fixes outside the handoff doc, hand off to the relevant Jag skill (e.g. `/sf-apex` for code fixes).

YAML is for routing. The body is for Claude. The three rules that keep skills from aging badly:

1. SKILL.md under 500 lines. Longer than that, push detail into references/.
2. Scripts for anything deterministic. Don't let the LLM reinvent a sf command on every run.
3. One job per skill. "Apex review" and "Apex fix" are two skills, even if they share references.

Skills versus subagents

Subagents are workers. Skills are knowledge.

You often want both. Jag's library ships seven subagents (fde-strategist, fde-engineer, fde-experience-specialist, fde-qa-engineer, fde-release-engineer, ps-technical-architect, ps-solution-architect) that each know when to invoke which skills. The strategist orchestrates up to four parallel workers via the Task tool. See Subagents and Agent Teams for the pattern.

Model-invoked versus manual

By default, Claude decides when to load a skill based on the description. You can override that.

---
name: acme-handoff
description: Our handoff doc conventions
disable-model-invocation: true
---

With disable-model-invocation: true, the skill only runs when the user types /acme-handoff. Useful when you want something as a manual command rather than something Claude reaches for unprompted. None of Jag's skills ship with this flag — they're all designed to be auto-invoked when relevant.

Next

• Using Jag's SF Skills for the full install guide, prerequisites, and skill-by-skill reference.
• Plugins and Marketplaces to distribute a bundle of skills, subagents, and hooks to your team.
• Subagents for the worker pattern that pairs with skills.
• Hooks and Custom Commands for the deterministic layer below skills.