Prompts, agent skills, workflows, and Plays are four containers for agent knowledge, and each one preserves a different thing. A prompt preserves intent, a skill preserves guidance, and a workflow preserves a graph a person drew. A Play preserves the exact path an agent already executed successfully. The right container is the one that keeps what the next run cannot afford to lose.
Teams mix these up because all four look like text about a task. The difference shows on the second run.
A prompt and a skill make the model reason again. A workflow runs, but only the steps someone predicted. A Play runs the recorded steps, with new inputs, and stops when the world no longer matches the recording.
How are agent skills different from prompts?
A skill is a prompt with structure and a trigger. Skills commonly mix guidance and instructions: when to act, which tools to prefer, what to avoid. The harness loads the skill into context when the task matches, and the model interprets it.
That interpretation is the cost. A skill tells the agent how to think about a task. It does not tell the agent which API call worked last Tuesday, in what order, with which parameters. Two runs of the same skill can take two different paths, and both paths cost reasoning tokens.
How is a workflow different from a skill?
A workflow is a graph that a person drew before the work happened. Tools such as n8n, Zapier, and LangGraph execute the graph without reasoning at each step. That is cheaper and more predictable than a skill.
The cost moves to authorship. Someone predicted the steps, wired the tools, and now maintains the graph as APIs change. Nobody records a workflow from work that already happened; a person designs it in advance.
How is a Play different from all three?
Rote compiles a Play from a run that succeeded. It records every API call, browser step, and shell command the agent makes. It removes the failed attempts, turns hardcoded values into typed parameters, and packages the result with a declared read/write contract.
Nobody drew the graph. The agent found the path, and Rote kept it.
Rote turns a successful agent run into an inspectable, repeatable Play that can travel across harnesses, models, machines, and teams. In one measured run, discovery took about 15,000 tokens and 40 seconds; replaying the Play took about 300 tokens and 2 seconds.
The four containers side by side
The table compares what each container stores, who writes it, and what happens on the second run.
| Container | What it stores | Who writes it | Second run | Breaks when |
|---|---|---|---|---|
| Prompt | Intent in prose | A person | Model reasons from scratch | The model interprets it differently |
| Skill | Guidance and triggers | A person | Model reasons within guidance | The API changes and the prose does not |
| Workflow | A predicted step graph | A person, in advance | Executes the drawn steps | A step nobody predicted is needed |
| Play | The recorded successful path | Compiled from the agent's run | Executes the recorded steps | The API contract drifts; the Play stops and says why |
Each row above is a legitimate tool. The mistake is using a prompt or skill for work that repeats unchanged, or a workflow for work nobody has performed yet.
Skills tell the agent how to think; Plays are what worked
Skills and Plays are complementary. A skill is the right container for judgment: how to approach a class of problems, what to check, when to ask. A Play is the right container for a method: this call, then this one, with these inputs, producing this shape.
Rote itself ships as a set of skills, and the Play sidekick is one of them. Before work starts, the sidekick searches local Plays and then the registry:
rote play search "weekly pipeline report"
rote play search "weekly pipeline report" --source registryThe sidekick offers a full match for inspection and approval; it never runs one on sight. No match means the agent works normally, and after a verified run the sidekick offers to keep it. The skill carries the habit; the Play carries the result.
When each container is the right choice
Use a prompt when the task is new and you expect the model to explore. Use a skill when the task recurs but the inputs and path vary enough that judgment is needed each time. Use a workflow when a person already knows the exact steps and no agent needs to discover them. Use a Play when an agent has already done the work once and the next run should execute rather than rediscover.
Questions people ask
Are agent skills the same as prompts?
No. A skill adds structure, a trigger, and often tool guidance to a prompt. The model reads and interprets both on every run.
Can a Play replace a skill?
Only for the method part. A Play cannot carry judgment about when to run or what to do when the task is unfamiliar. A skill can point the agent at the right Play.
Does a Play have to come from a recorded run?
Yes. Rote compiles a Play from a trace that executed. That is what gives it evidence a workflow or skill does not have.
What happens to a Play when the API changes?
The Play carries a fingerprint of the API contract it was built against. When the contract drifts, the Play stops and reports the mismatch. The harness examines the failure, the agent re-adapts, and the new successful path is pushed as the next version of the same URI.
This essay is one of five on turning a run into a reusable procedure
Start with the overview, How a successful AI agent run becomes a reusable procedure. The other essays each take one question further:
- How to extract a reusable procedure from an agent trace: the five compile operations and the four conditions that block them.
- Agent procedure memory survives a model change only as a program: why a Play runs from any model and stops when the API drifts.
- Share an agent workflow with a teammate without sharing a credential: what travels with a Play URI and what stays on the runner's machine.
Install the Play sidekick and try this on your own agent
One command installs the Play sidekick, installs Rote when it is missing, and wires the skill into the agent apps it finds on your machine. It runs on macOS and Linux, and on Windows inside WSL2.
curl -fsSL https://getrote.dev/playoffs/install.sh | shRestart your agent app, type $play (or /play in Claude Code), and run the Hello Play. It uses public data, needs no credentials, and declares no writes.
Ask what the second run should cost
Before writing the next skill, ask whether the task will run unchanged next week. If it will, start it with $play explore <outcome>, let the agent perform it once, and settle the Play. Reserve the skill for the judgment around it.