Skills

An agent's system prompt describes its persona. Procedures — how to take a confirmation number, how to verify a caller, how to walk a return — either get pasted in there, where they lengthen the persona and have to be maintained agent by agent, or they do not exist.

A skill is the third option: a named, self-contained procedure you switch on per agent. The agent follows it whenever the situation it describes comes up, and it stays separate from the persona you write.

Skills live on the Skills row of the Tools & Skills group, on the Conversation step of the agent editor. The row shows which skills are on; clicking it opens the editor in a side panel.

Platform skills

Three skills ship with Tandem and can be switched on by name.

collect-code handles a spoken code: a confirmation number, order number, member ID, case number. It asks for the code on its own, asks for a word when a letter is ambiguous ("was that M as in Mary?"), reads the code back, and retries up to three times before falling back.

read-back-values covers the other direction — a value the agent says to the caller, which they have to write down or check against a screen. A voice reads 12345 as "twelve thousand three hundred forty five", so the skill has the agent write it as "one two three four five" instead, grouped with commas for the pauses. It covers codes and IDs, and also phone numbers, dates, times, and dollar amounts.

The skill governs speech only. The agent still passes the original, unspaced value to your tools, so a lookup on the code keeps working.

verify-identity confirms a caller before you disclose anything. It collects your configured identifying details one at a time, calls your own verification tool, allows three attempts, and then follows the failure path your system prompt names. It never tells the caller which detail was wrong.

verify-identity uses a tool you have already bound to the agent under Tools. It does not bring one of its own. Without a verification tool configured, the agent says it cannot confirm the account and follows the failure path.

In a website chat, collect-code and verify-identity use chat wording. The agent takes what the visitor typed as written: it does not ask them to spell a code or an answer, and it does not read a value back to check it. Each still allows three attempts and follows the same failure path. read-back-values applies to calls only, so it is skipped in a chat.

Each platform skill also has a View action, which shows its full instructions as the agent receives them on a call. Read them there before you switch a skill on — the description on the row is a one-line summary, and the instructions are what actually run.

Custom skills

A custom skill is a name, a description, and instructions.

Name — lowercase letters, numbers, and hyphens, up to 64 characters, and it must start with a letter or number (not a hyphen). It cannot be the name of a platform skill; use the Duplicate action if you want to start from one.

Description — up to 280 characters. This is what tells the agent when the skill applies, so write it as a trigger: "Use when the caller asks to return a part." Keep it to a sentence or two.

Instructions — up to 8,000 characters. This is the procedure itself, and it can be much longer than the description.

You can have up to ten platform skills enabled and ten custom skills.

Starting from a platform skill

Each platform skill has a Duplicate action. It copies the description and instructions into a new custom skill named <skill>-copy, and opens it for editing.

The copy is a fork, not a replacement: the platform skill stays in the list and stays independently switchable, and it keeps getting our updates while your copy does not. That is what the -copy suffix is for — a custom skill that takes a platform skill's exact name is ignored at call time rather than overriding it.

Writing skills for voice

These instructions are executed by an agent talking on a phone. The format does not imply that constraint, so it is the one most often missed.

No tables, no bullet lists, no code blocks. The agent will read them out loud.

Ask for one thing at a time. "Collect the customer's name, address, and order number" gets you a run-on question and a partial answer. "Ask for the order number. Wait for an answer, then ask for the postal code" gets you both.

State retry limits in words, and say what happens when they run out. There is no loop construct; write the repetition into the instructions. "You get three attempts. After the third, stop asking and offer to transfer."

Say what to do when the caller goes off-script, because on a phone call they will. Callers change the subject, do not have the number in front of them, and ask why you need it. Every one of those is a branch your instructions either cover or leave to chance.

What skills are not

A skill is instructions, not permissions. It cannot restrict which tools the agent may call, and an agent with a skill active still answers unrelated questions with unrelated tools, which is usually what you want.

To constrain what an agent may do, use Steering Rules, which are checked before every tool call.

Reserved names

skills is reserved — it is the name the agent uses to fetch a skill's instructions when it has to fetch them — so you cannot name a workflow or integration tool skills. Every other name the platform's own tools answer to is reserved the same way — end_call, search_knowledge_base, and each of the built-in tools. See Tools for the full list.

Over the API

Skills are part of an agent's configuration, under skills.builtins (platform skill names) and skills.custom. GET /v1/agent-media/skills lists the platform catalog with each skill's full instructions, which is what the View and Duplicate actions read. See the API reference.

Next steps