Lesson 2 of 5
Is this a tool or a skill?
Product pages list both “integrations” and “skills,” often in the same sentence, and the two do different jobs. One is a callable operation. The other is a procedure for using operations. Module 4 introduced the skill as a place to keep a procedure; this lesson puts it beside the connections from lesson 1 and shows, with one evaluation a vendor published, why having the operation is not the same as knowing how to use it.
Numbered links identify sources. Underlined terms offer a definition without leaving the lesson.
An operation and a procedure are different kinds of thing
Lesson 1 ended with a harness that could discover and call find_supplier. Module 3 defined a as an operation the harness performs when the model asks: something happens, on some host, under some access. Module 4 defined a as a folder of instructions for one recurring job, listed by name and description and loaded in full only when chosen. In the products this lesson cites, a skill's folder can also hold scripts the model may run and configuration particular to a product; the procedure itself, though, is text. When a skill is chosen, that text enters the context. Loading it performs no lookup. If the steps say “look the supplier up,” the lookup still needs a tool that can do it, and the harness's for that tool still applies.
Anthropic's training material for its own harness draws the same line for its products, describing MCP servers as a way to provide external tools and integrations, a different category from skills, which carry task-specific knowledge loaded when relevant. 13 OpenAI's documentation for its harness describes a skill as a folder with a file of instructions plus optional scripts and reference documents, of which only the name and description are visible until the skill is selected. 14 Both descriptions are of particular products on 29 September 2026. The distinction they share is the one this course needs: a connection gives the model something it can ask the harness to do; a skill gives the model an account of how to do a job with what is available.
The garden example makes the difference concrete. Here is an authored skill for the check the organizer wanted, written against the tool from lesson 1 and its matching rule:
name: check-supplier-approved
description: Use before relying on a supplier's quote for the garden. Confirms the supplier
is on the committee's approved list and records the approved-list row beside the quote.
Steps
1. Call find_supplier with the name as it appears on the quote. The tool matches any
row whose name contains that text, ignoring letter case.
2. If no row is returned, call it once more with the shortest distinctive word in the
name. If still no row, report: no match under the tool's matching rule. Do not
report the supplier as unapproved; a person checks spelling and the list.
3. If more than one row is returned, do not choose. Report all rows and stop.
4. If one row is returned, compare its full name and contact with the quote. A single
partial-name match does not establish identity. Ask the organizer to resolve any
mismatch; otherwise record the id and source location next to the quote.
Every step names a tool result and a decision. Suppose the harness has no other route to the table: no shell, no API connection, only this MCP server. Take find_supplier away and the skill can still be loaded, and step 1 describes a lookup that cannot happen. What the model does then is not settled by the skill: it may stop and say the tool is missing, it may generate a call the harness cannot run, or it may write a sentence about looking up that no observation supports. Keep the tool and remove the skill, and the lookup can happen, but this setup no longer supplies the committee’s procedure for two rows or none. Other instructions or learned patterns may suggest a response; they need not match the committee’s rule. That is the case where a model does what it thinks is reasonable, and module 3's trace showed what that can look like.
Having the operation is not knowing when to use it
Supabase, a hosted database provider, compared agents with and without its tools and skill in April 2026. Its employee Pedro Rodrigues reports six scenarios per condition, graded by a model. Agents with tools alone often omitted a required database security setting; all models tested applied it when the skill was available. This supports a limited distinction: having an operation available and being given the procedure for using it are different conditions. The vendor wrote the evaluation, the cases were few, and model grading can be wrong. It establishes neither a general success rate for skills nor the cause of every omission. 15 The site also collected the accompanying talk.
The garden version is the same shape. Connect the supplier server and the harness gains find_supplier and its description. The committee's rule, that every quote is matched to the approved list before use, is not in that description; it is the committee's rule, and it reaches the model only if something carries it: the skill above, a line in the brief, an instruction file, or the model's training, and the last of those is the one you cannot inspect.
What a procedure enforces, and what a connection adds
The check-supplier-approved skill says “do not choose” when two rows come back. That sentence is an in module 3's first layer: it makes the right stop more likely, and it prevents nothing. If the harness's rules allow a write and the model proposes one anyway, the write runs. A skill folder can contain more than prose; a script inside it runs as a program, under whatever the harness allows programs, and product configuration inside it may change what a tool does. Sort each part: the procedure is text the model reads, and any enforcement comes from the harness's rules or the boundary, which are elsewhere. A product that says an agent “has a skill for safely handling supplier data” has described, at least, text that will be loaded; what else it has described is a question for the product's own documentation.
A connection is different, and the difference is worth being exact about. Connecting the supplier server with the garden account's credential adds a capability the harness did not have: it can now reach whatever that credential is authorized for. That is a real grant, made when the connection was set up. What the server may then read or change is set by what it runs under, the file permissions of its process or the authorization attached to the credential it presents, and none of that is visible in the connection's name. Lesson 5 turns this into the last line of a tool's contract, because it is the line that product descriptions most often leave out.
Choosing between an instruction, an operation and a boundary
The module's checkpoint asks you to decide, for a given problem, whether the fix is better instructions, a new operation, or enforced access. Lessons 1 and 2 give a way to diagnose it from evidence. If the operation exists and ran, and the record shows it was used at the wrong moment or its result was handled wrongly, the missing piece is most likely a procedure: a skill, or a line in the brief. If no operation exists that can do what the task needs, no amount of instruction supplies one; the fix is a tool, reached through one of lesson 1's routes. If the operation exists and the problem is that it could touch something it must not, an instruction reduces the chance and a boundary removes it, and module 3's three layers say which you are looking at. These are diagnoses, and they can be wrong: a wrong choice that looks like a missing instruction may turn out, on a second look at the record, to be a tool returning too little to choose from.
Apply the test to three authored garden problems. The agent read a quote from a supplier not on the list: an operation existed for reading quotes, nothing told the model to check the list, so the fix is the procedure above. The agent could not check the list because the file was on the committee's service and the harness had no way to reach it: the fix is a connection, through the service's API or an MCP server in front of it. The agent, while checking, could also have rewritten the list, because the same credential allowed writes: the fix is a credential authorized to read and not to write, which is enforced, not asked for. The three problems look alike from the outside. They are fixed in three different places.
Carry this forward
- A tool is an operation the harness performs; a skill is a procedure, text loaded when its job comes up, sometimes with scripts and configuration beside it. Loading the procedure performs no operation and enforces none of its own steps.
- A connected server supplies operations and their descriptions, not the rules of the organization using them. One vendor's small evaluation found a required step omitted with the tools alone and applied with a skill that stated it; that shows the pattern for those tasks, not a general effect size.
- Connecting a server with a credential adds a real capability; the connection's name does not say how much. What the operation may read or change is set by what the server runs under.
- Diagnose a problem before fixing it: a wrong use of an existing operation points to a procedure; a missing capability points to a connection; an operation that could touch what it must not points to an enforced boundary. The diagnosis comes from the record and can be revised.
Lesson check
Choose an answer for each question.
This is a practice check. Answers reset when you leave this page. Results are not saved and do not place you in another module or award a qualification.
The skill above says “if more than one row is returned.” To write steps like that you need to know what a tool's results and errors look like. The next lesson reads the supplier table the way a program does.
Sources · Reviewed 2026-09-29
- 13. Anthropic Academy: Skills vs. other Claude Code features
- 14. OpenAI Codex documentation: Build skills
- 15. Supabase (Pedro Rodrigues, 9 April 2026): Supabase agent skills
- 16. VibeLeaderboard Intel: Combine skills and MCP to close the context gap (Pedro Rodrigues, Supabase)
Sources support nearby explanations. Product details can change. Scenarios and model results are authored illustrations unless a lesson explicitly identifies a recorded execution. Module 7 identifies its offline code checks.