Цю сторінку ще не перекладено.

Prompts and skills

This page is for developers who build on the MCP server or connect a host to it. It lists the 16 prompts a host can offer its user and the 11 skills a model can load, what each takes, and when to use which.

Prompts and skills, compared

Both tell a model how to do a job with the server's tools. Neither runs anything on its own, and neither gives a caller more than its token and role allow.

Prompts Skills
What it is A short task, filled in with arguments, sent to the model as one user message A longer guide to a whole job: steps, checkpoints, rules and how to recover from errors
Who starts it The person, from the host's prompt menu or slash commands The model, which reads it as a resource when a task needs it
How it is fetched prompts/list, then prompts/get with arguments resources/read on a skill:// address
Changes anything No. The tools it names do, under the usual contracts No

Prompts

The server lists every prompt to every signed-in caller. prompts/get answers one user message. For every prompt but follow, the message is the prompt's template with its arguments filled in. Every template opens with the server's short rules: drafts and versions, reading through resources, changing a draft with an expected revision, and never releasing without the person's say.

List the prompts
{"jsonrpc": "2.0", "id": 3, "method": "prompts/list", "params": {}}
Get one, with its arguments
{
  "jsonrpc": "2.0", "id": 4, "method": "prompts/get",
  "params": {
    "name": "add-guardrail",
    "arguments": { "uri": "instruction://acme/support-agent", "risk": "a refund over €500 without a person" }
  }
}
The answer (shortened)
{
  "messages": [
    {
      "role": "user",
      "content": {
        "type": "text",
        "text": "How instruction.md works, in short:\n- An instruction has one draft …\n\nTask: add a guardrail to instruction://acme/support-agent so that this never happens: a refund over €500 without a person.\n\n1. Read `instruction://acme/support-agent/draft`; find where it belongs. …"
      }
    }
  ]
}

An argument that is not required has a default the template names, such as "the person's personal workspace". A required argument left out doesn't fail the request. The message then says which arguments are missing and asks the model to get them from the person and run the prompt again. A prompt name the server doesn't have doesn't fail either: its one message is "No such prompt.".

PromptWhat it doesArguments
write-instruction
Write an instruction
Write a new instruction from what the person wants an agent to do, and save it as a draft for them to review.intent (required): what the instruction should make an agent do, in the person's words
workspaceId (optional): where to create it (default: the personal workspace)
propose-edit
Propose an edit
Change an instruction the governed way: read it, then make the change as suggestions (or directly, when you may and were asked to), never releasing it.uri (required): instruction://<id|owner/name> to change
intent (required): what should change, in the person's words
tighten-rules
Tighten the rules
Make an instruction's rules precise and testable: vague words out, conditions and limits in, as suggestions.uri (required): the instruction
focus (optional): which rules or section, if not all
explain-rule
Explain a rule
Explain what one rule of an instruction makes an agent do, and where it could be misread.uri (required): the instruction
rule (required): the rule's words or its name
write-section
Write a section
Write one new section of an instruction from a description, as a suggestion in place.uri (required): the instruction
about (required): what the section covers
after (optional): the heading it follows, if not at the end
add-guardrail
Add a guardrail
Add a NEVER rule for a risk the person names, as a suggestion.uri (required): the instruction
risk (required): what must not happen
make-example
Make an example
Write a worked example for an instruction: a situation, what the agent should do, and why, as a suggestion.uri (required): the instruction
case (optional): the situation the example shows
fix-problems
Fix problems
Fix what the language reader refuses or warns about in an instruction's draft, one problem at a time.uri (required): the instruction
convert-prompt
Convert a prompt
Turn a system prompt into an instruction: its rules become MUST, SHOULD and NEVER lines, saved as a new draft.prompt (required): the system prompt's text
title (optional): a name for the instruction
prepare-release
Prepare a release
Summarize what changed since the latest version into a version message, check the draft, and say what releasing would change for the agents that follow it.uri (required): the instruction
answer-comments
Answer comments
Read the open comment threads on an instruction and answer each one, changing the draft as suggestions where a comment asks for it.uri (required): the instruction
review-proposal
Review a proposal
Review a proposal (a whole text offered back) against the instruction it targets: correctness, validity, who it affects; recommend approve or request changes.instructionId (required): the instruction (ins_…)
proposalId (required): the proposal (prp_…)
explain-diff
Explain a diff
Explain what changed between two versions of an instruction and what it means for the agents that follow it.instructionId (required): the instruction (ins_…)
from (required): a version (v3, ver_…), a ref, or draft
to (required): a version, a ref, or draft
create-from-intent
Create from intent
Create a new instruction from a goal, starting from a template when one fits, and save it as a draft.intent (required): what the instruction should make an agent do
workspaceId (optional): where to create it
adapt-for-agent
Adapt for an agent
Adapt an instruction for a specific agent or host without changing the original: fork it, adjust the copy, keep its origin.uri (required): the instruction to adapt
agent (required): the target agent or host (agentd, Claude Code, Cursor, a CI job…)
follow
Follow an instruction
Take an instruction as this conversation's instructions: its delivered text (the version its ref names now, with its includes and its parameters' defaults), which you follow from here on.uri (required): the instruction to follow: instruction://owner/name, with @ref (stable, latest) or @v3 for one version

The prompts that change an existing instruction tell the model to make the change as suggestions. propose-edit lets it edit the draft directly only when the caller may and the person asked for that. The prompts that make a new instruction (write-instruction, convert-prompt, create-from-intent) save it as a draft, and adapt-for-agent forks one. None of them releases a version. prepare-release writes the version message and asks the person whether to release. The release itself is instructions_publish, which asks to confirm. See Contracts.

The follow prompt

follow is not filled from a template. The server reads the instruction as the caller, with the caller's permission, and answers its delivered text: the version its ref names now, with its includes and its parameters' defaults in. The read is recorded like any other read.

  • uri takes instruction://owner/name, owner/name, or either with @stable, @latest or @v3. Without a ref it is the instruction's default ref.
  • The text comes wrapped in an <instruction> element with its address, version and digest, the same digest the hooks and instruction://current give.
  • An instruction the caller can't read answers not_found, as a read does. An empty uri answers validation.
What follow answers
{
  "description": "acme/support-agent v7",
  "messages": [
    {
      "role": "user",
      "content": {
        "type": "text",
        "text": "These are the instructions you follow in this conversation, from instruction.md: acme/support-agent v7 (signed). Follow them from here on; they replace any earlier version of them.\n<instruction uri=\"instruction://acme/support-agent@stable\" version=\"v7\" digest=\"sha256:…\">\n# Support agent\n…\n</instruction>"
      }
    }
  ]
}

The prompt gives the conversation the version as it is when it runs. A host that doesn't subscribe to resources doesn't hear of a new version afterwards: run the prompt again, or set up following.

How hosts show prompts

  • Claude Code lists each prompt as a slash command named after the server it was added as: with claude mcp add instruction-md …, the follow prompt is /mcp__instruction-md__follow instruction://acme/support-agent. See Claude Code.
  • Codex reaches the server through the instruction mcp bridge, because it doesn't speak MCP 2026-07-28 by default. To follow an instruction in Codex, use its hooks: see Codex.
  • The stdio servers pass the prompts on. instruction mcp lists the server's prompts once you are signed in. With --follow, or a project's .instruction/follow.json, its own follow takes the place of the server's, and uri becomes optional: it defaults to the followed instruction. instruction-mcp lists the server's prompts as they are. See instruction-mcp.
  • Other hosts differ. A host with no prompt menu can still call the same tools; the prompts only save the person typing the task.

Skills

A skill is a Markdown file, SKILL.md, with front matter that says when to use it, and sometimes reference files beside it. The server serves each as a resource. The index lists them all with each file's digest and size:

Read the index
{"jsonrpc": "2.0", "id": 5, "method": "resources/read", "params": {"uri": "skill://instruction-md/index.json"}}
The index's shape
{
  "version": 1,
  "skills": [
    {
      "name": "instruction-md-basics",
      "description": "Start here before using any instruction.md tool. What instru…",
      "uri": "skill://instruction-md/instruction-md-basics/SKILL.md",
      "resources": [
        {
          "uri": "skill://instruction-md/instruction-md-basics/SKILL.md",
          "digest": "sha256:…",
          "size": 8772,
          "mimeType": "text/markdown"
        }
      ]
    },
    …
  ]
}
Read one skill
{"jsonrpc": "2.0", "id": 6, "method": "resources/read", "params": {"uri": "skill://instruction-md/releasing-and-rolling-back/SKILL.md"}}

The addresses are skill://instruction-md/<name>/SKILL.md and skill://instruction-md/<name>/reference/<file>. The server has no skills/list or skills/get method: read the index, then the files.

SkillWhen to load itForFiles
instruction-md-basicsStart here before using any instruction.md tool. What instruction.md is (versioned instructions for AI agents, served over MCP), its model (draft, suggestions, comments, checkpoints, versions, latest, roles), how to find tools and read through resources, the revision and confirmation rules every change follows, and what each error means.an assistant acting for a person, an agent that follows an instruction, an agent that writes instructions, a coding agent, a CI pipelineskill://instruction-md/instruction-md-basics/SKILL.md
writing-instructionsWrite a new instruction for an AI agent from what a person wants it to do — "write an instruction for our support agent", "set up rules for the coding agent". Interviews for purpose and limits, starts from a template, writes rules, guardrails and examples in the language publishing accepts, validates until clean, previews what the agent will receive, and saves the draft.an assistant acting for a person, an agent that writes instructions, a coding agentskill://instruction-md/writing-instructions/SKILL.md
skill://instruction-md/writing-instructions/reference/language.md
improving-an-instructionChange an existing instruction safely — "make the refund rules stricter", "add a rule about tone", "fix the problems", "rename it". Reads the draft and what others have pending, makes precise edits rather than rewriting, proposes them as suggestions when the person can't edit directly or others are working on it, validates, and shows the diff.an assistant acting for a person, an agent that writes instructions, a coding agentskill://instruction-md/improving-an-instruction/SKILL.md
releasing-and-rolling-backRelease a new version of an instruction to the agents that follow it, roll back to an earlier version, move a ref such as stable, or revoke a harmful version — "ship it", "publish the changes", "roll back to v6", "promote to stable", "v5 leaked a key, kill it". Checks for problems and pending suggestions first, writes the version message, asks the person to confirm before agents switch, and checks adoption after.an assistant acting for a person, a CI pipelineskill://instruction-md/releasing-and-rolling-back/SKILL.md
release-gates-in-ciRelease an instruction from a CI pipeline behind gates — validate the changed file, write it to the draft, run the checks, release it to canary, and let stable follow once it has stood and passed. "Publish instructions from GitHub Actions", "only ship checked versions", "canary then stable", "why didn't stable move?". Also sets up the gates on the refs, which a person does once.a CI pipeline, an assistant acting for a personskill://instruction-md/release-gates-in-ci/SKILL.md
following-instructionsFor an AI agent or agent runtime that runs with an instruction from instruction.md as its operating context — bind to a version or ref, read and verify it, apply it at a safe moment, report what you applied, stay current when a new version is released, and hold the last good version when something is wrong.an agent that follows an instruction, an agent that writes instructions, a coding agentskill://instruction-md/following-instructions/SKILL.md
converting-a-promptTurn an existing system prompt into an instruction — "make this prompt an instruction", "import our support bot's prompt", "convert this". Converts its rules into MUST, SHOULD and NEVER lines, shows the result before anything is created, creates the instruction from it, and checks it.an assistant acting for a person, an agent that writes instructions, a coding agentskill://instruction-md/converting-a-prompt/SKILL.md
reviewing-changesReview what others propose on an instruction — "what's waiting for me", "review the suggestions", "answer the comments", "accept Maria's change". Reads the pending suggestions and comment threads, explains each against the released text, and settles them only as the person decides.an assistant acting for a personskill://instruction-md/reviewing-changes/SKILL.md
sharing-and-accessShare an instruction or change who can use it — "share it with Priya", "let the support team edit it", "make it public", "who can see this?", "ask for edit access", "make my own copy". Reads who has access, changes grants and visibility with the server's confirmation shown to the person, answers access requests, and asks for access or forks when the person can only read.an assistant acting for a personskill://instruction-md/sharing-and-access/SKILL.md
connecting-an-agentConnect an agent so it follows an instruction — "hook our support bot up to the refund rules", "give the agent a credential", "is the agent on the latest version?", "pause that agent". Issues the agent's credential, shows how it signs in and binds to the instruction, and checks from the instruction's side that it follows and applies the latest version.an assistant acting for a person, a coding agentskill://instruction-md/connecting-an-agent/SKILL.md
administering-an-orgRun an organization — "invite Tom", "make Priya an admin", "create a workspace for the support team", "what's our policy on public instructions?", "add our GitHub skills repo", "only let instructions use our approved MCP servers", "who changed this and when?", "export the audit log". Manages members, invitations, workspaces and roles, the organization's policy and integrations, and its audit trail, asking first where the server does.an assistant acting for a personskill://instruction-md/administering-an-org/SKILL.md

Which skills to load first

session_bootstrap names the skills to load first in its skills field, by the kind of caller:

CallerSkills it names
Every callerinstruction-md-basics
An agent or a service whose credential has no write, publish or admin scope (and isn't a CI service)following-instructions
A person, an agent or service with a write, publish or admin scope, or a CI servicewriting-instructions, improving-an-instruction, releasing-and-rolling-back
A CI service (its client ID starts with ci- or names github)release-gates-in-ci

The lists add up: every caller gets instruction-md-basics, and a CI service gets the three for writers as well as release-gates-in-ci. Load the others when the task calls for them. capabilities_search finds skills and prompts too, by what you want to do in words.

What a skill can't do

A skill is guidance for the model, not a permission. A tool it names still checks the caller's scopes and role, and a tool that asks to confirm still asks. A skill never tells a model to do more than its person asked.

Generated from the code

The prompt and skill tables are built from the tool registry (packages/registry/registry.json) when the docs are built, and the "For" column from each skill's front matter. A prompt or skill added to the server appears here at the next build.

Last checked against the product:

Was this page helpful? Report a problem