Claude Skills and CLAUDE.md: a practical 2026 guide for teams

Why standardise Claude now

Left to chance, AI outputs vary by person and prompt. Skills (folders with a SKILL.md) and CLAUDE.md give Claude durable guidance: consistent behaviours for non-dev work, and persistent engineering conventions for code. That means fewer “one-off” prompts and more repeatable, reviewable outcomes. Claude+1

What are Claude Skills (and SKILL.md)?

Agent Skills are modular capabilities that Claude can invoke automatically when your request matches a Skill’s description. Each Skill contains a SKILL.md with instructions and optional scripts/templates; Claude selects and applies relevant Skills without slash commands. Think “reusable playbooks” instead of ad-hoc prompting. Claude+1

Where they help

• Policy writing, brief creation, research triage, QA checklists (non-dev teams)

• Repetitive founder/ops flows (triage, outreach, competitive scans)

• Consistent UI/UX review criteria for product design

Anthropic’s public examples and docs show Skills packaging expertise so Claude can decide when to use them. Claude

What is CLAUDE.md (for engineering teams)?

CLAUDE.md is a file Claude Code reads automatically to learn your project structure, coding standards, testing rules, commit/style conventions and workflows. You can place it at the repo root (recommended) or where you run claude, and you can scope additional CLAUDE.md files inside sub-directories for local rules. Anthropic

Why it matters: it eliminates “cold start” prompting every session. Claude Code starts with your context and writes diffs/tests/docs to match the way your team works. Anthropic’s guidance and product blog outline how to structure these files for best effect. Anthropic+1

Where to put domain knowledge (Projects)

For broader context (brand voice, policies, playbooks, research), use Projects to ground Claude with files and references. Projects create self-contained workspaces with their own knowledge base and chat history—ideal for marketing, support, or client work. Anthropic+1

Prompting foundations (quick wins)

From Anthropic’s prompting best practices:

• State audience, objective, constraints and success criteria.

• Provide examples (good/bad) and data snippets for reference.

• Ask for verification steps and citations when appropriate.

• Prevent “over-engineering” by explicitly asking for minimal changes when coding. Claude

SKILL.md: a practical template

Use this skeleton to stabilise outputs for non-dev work (policy drafts, briefs, QA):

# SKILL.md — InfoSec Policy Drafting (v1.0)
## Purpose
Draft or update InfoSec policies aligned to ISO 27001; produce reviewer-ready outputs.

## Triggers (Claude decides when relevant)
- “Draft/update InfoSec policy …”
- “Create a procedure for …”
- “Map controls to ISO/SoA …”

## Inputs
- Policy topic, scope, audience
- Constraints (standards, legal refs)
- Source docs (URLs/files)

## Steps
1) Clarify scope (ask 3–5 questions if ambiguous).
2) Outline sections; confirm in 5 bullets.
3) Draft with headings, definitions, RACI.
4) Add ISO mappings + change log.
5) Produce summary and next actions.

## Outputs
- Policy (Markdown, headings, numbered sections)
- 200-word exec summary
- Review checklist

## Refusal & Escalation
- Refuse legal advice; flag to Legal/SME when regulatory claims are requested.

## Evaluation Checklist
- Factuality vs supplied sources ✓
- ISO mapping included ✓
- Reading age ≤ 11–12 ✓

Skills are model-invoked and built around SKILL.md + optional scripts/templates. Claude+1

CLAUDE.md: engineering starter

Drop this in your repo root and adapt per stack:

# CLAUDE.md — Web App Conventions (v1.0)
## Architecture
Monorepo: /apps/web (Next.js), /packages/ui (React), /packages/config.
Testing: Vitest + Playwright. State: Zustand. Styling: Tailwind.

## Coding Rules
- Prefer small diffs; no unrequested refactors.
- Add/modify tests for changed logic.
- Accessibility: label form controls; keyboard nav required.
- i18n: wrap user-visible strings.

## Git & Reviews
- Conventional commits; one issue per PR.
- Include: summary, rationale, screenshots, test notes.
- Don’t add dependencies without a clear reason.

## DX Commands
- `pnpm dev`, `pnpm test`, `pnpm e2e`

## Definition of Done
- Tests pass; coverage not lower.
- Lighthouse perf ≥ 90 on changed pages.

## Prompts for You (Claude)
- Before coding: propose plan in bullets; wait for “OK”.
- Keep solutions minimal and focused; avoid over-engineering

Official guidance notes Claude Code reads CLAUDE.md automatically and supports placing it at repo root or where you run claude. Anthropic

Design patterns (worked examples)

• Front-end review Skill: enforce accessibility, performance budgets, design tokens and copy tone; output a pass/fail checklist and annotated diffs. (Implement as SKILL.md + optional audit script.) Claude

• Founder ops Skill: triage inbound, enrich with web citations, generate outreach copy under brand voice; produce CSV for CRM import.

• Policy desk Skill: transform SME notes + standards into versioned SOPs with change logs and RACI.

Roll-out checklist (6 weeks)

1. Select use cases (2–3 per team) and build initial SKILL.md/CLAUDE.md.

2. Set guardrails (PII, client secrets, regulatory red-lines; approved sources; citation rules).

3. Pilot with 5–10 users per team; enable Projects for shared context. Claude Help Center

4. Measure: time-to-first-draft, review cycles, defect rate, % outputs with acceptable citations.

5. Harden: refine Skills, expand CLAUDE.md, add evaluation checklists.

6. Standardise: document in your playbook, train, publish to the org.

Tip: When coding, explicitly instruct minimal changes to reduce unnecessary abstractions. Claude

Governance & risk

• Traceability: require sources/citations on external-facing content.

• Change control: version SKILL.md/CLAUDE.md; PR review by SMEs.

• Access: separate Projects by client/department; apply least-privilege. Claude Help Center

  
  

FAQ

What’s the difference between Skills and prompts?
Skills are reusable, documented behaviours packaged with SKILL.md and optional resources; Claude can invoke them when relevant. Prompts are one-offs. Claude

Where should I put CLAUDE.md?
At the repo root (recommended) or wherever you run claude; use additional CLAUDE.md files in sub-directories for local rules if needed. Anthropic

How do we centralise knowledge for a team or client?
Create a Project and upload reference docs, style guides and past work; keep Skills small and focused, and point them at the Project’s materials. Anthropic+1