When handing a recurring task to an AI model—like assembling a sales proposal from client call notes—you usually see two options. Either you paste your rate card into chat every morning and catch made-up details by hand, or you call a developer to write a custom script for every template.
Both paths wear you down: in a long thread, instructions drift by the tenth prompt and the model invents discounts, while custom code for routine documents eats up too much effort. I dug into Anthropic's Claude Skills specification, and it threads the needle cleanly: a no-code folder locks down the role, data validation, and output format—let's see how such a package is structured and how to keep it from misfiring.
A Folder Instead of an Endless Prompt: How a SKILL.md Package Works
Why create folders instead of one giant system prompt? If you load a dozen SOPs into context at once, the model needlessly re-reads 20,000+ tokens at the start of every chat before your first question. And if all you need is to calculate an estimate after a call, extra rules waste your limit and scatter the algorithm's attention.
That is why skills use progressive disclosure—like a desk card index with thin title tabs while heavy folders stay on the shelf. In Claude.ai, Claude Code, and the Claude API, only an index stays in memory: each skill's name and description take roughly 100 tokens, so 10 skills use around 1,000 tokens. Only when a request matches the description does Claude load the full SOP.
The core of the folder is SKILL.md: a YAML header at the top with name and description fields, followed by steps, validation rules, and output formatting. Reference tables go into references/, Python or Bash scripts go into scripts/, and access to external APIs like a CRM or WordPress is granted via MCP.
Let's assemble a package for call-to-estimate quoting—we place rates in references/pricelist.md and write sales-proposal/SKILL.md:
---
name: sales-proposal
description: Builds a sales proposal from call notes. Trigger on requests to calculate an estimate or draft a proposal.
---
# Sales Proposal Assembly
1. Extract the client's goal, requested services, and timeline from the notes.
2. Check every rate strictly against `references/pricelist.md`.
3. Never invent discounts outside the price list; if inputs are missing, stop and ask.
4. Output a DRAFT following the checklist: goal, estimate, timeline, next step.
Before, a sales manager copied an old chat, forgot the updated price list, and had to catch a hallucinated discount before hitting send. Now they drop three rough paragraphs from a meeting into the window: Claude activates sales-proposal, checks prices against references/pricelist.md, asks clarifying questions if inputs are missing, and outputs a clean checklist draft in both the browser and the terminal.
Edge-Case Testing: How to Keep Skills from Mixing Up and Misfiring
With one folder everything runs smoothly, yet you might fairly ask: "Once I have a dozen skills, won't Claude start mixing them up and modifying live services without asking?"
The first failure point hides in a vague description field. Before activation, the model sees only this header: write something generic like "helps with writing," and across skills for proposals, blog posts, and newsletters the assistant will trigger the wrong SOP or pepper you with questions. Write the description as a sharp entry filter and trigger.
The second line of defense is explicit prohibitions inside SKILL.md. SEO practitioner Julian Goldie demonstrated a stack of 7 working skills: folders hunt for trending topics with links, analyze Google Search Console, format articles in WordPress—shrinking editorial-compliant article production from 2 hours down to 10 minutes—and prepare videos, carousels, LinkedIn comments, and emails from client questions. From his setup, the first thing I would adopt is hard guardrails: Search Console is read-only, articles and emails go strictly to drafts without overwriting live pages, and the video module waits for a "Yes" before publishing.
Rolling out a skill works cleanly in four steps: pick a frequent task, ask Claude to draft SKILL.md from a good example, test it on real notes, and whenever it slips, add a rule to the file instead of correcting it in chat. Just note that scripts/ require Python or Bash in the environment, and team skills are best synced via Git.
Once the role, the price list in references/, and the checklist live inside SKILL.md, the model holds the standard without guessing. Official starter templates live in Anthropic's repository, and to go deeper check our site breakdowns: our piece on moving from static files to dynamic prompts shows why dumping documents dilutes model focus, while the guide on self-learning skills in Claude Code teaches how to add automated tests and error memory to a skill folder.

