prompt

name	prompt
description	Assemble a complete agent prompt by orchestrating the focused skills (promptrole, promptcontext, promptbehavior, prompttool, prompttask, promptverification) into one combined file. Use when drafting a new agent prompt or improving an existing one.
license	MIT

Prompt

Output: a single combined file — prompts/{slug}.md (Markdown, default) or prompts/{slug}.xml (XML).

Context

Assemble a complete agent prompt. This skill applies the focused skills in sequence — promptcontext, promptrole, promptbehavior, prompttool, prompttask, promptverification — and combines their outputs into a single file with the correct section order and cross-section coherence.

Tradeoff: Not every prompt needs all sections. A simple one-shot prompt may need only role and task. Invoke only what the prompt requires.

Scope: This skill assembles section files into a single combined prompt. Each section's content is produced by the focused skills, not here.

Section order in the combined file (Markdown H2 headings or XML tags):

1. ## Context / <context> — runtime environment
2. ## Role / <role> — agent identity
3. ## Behavior / <behavior> — operational rules (skip for one-shot)
4. ## Tools / <tools> — tool definitions (skip if none)
5. ## Task / <task> — the instruction
6. ## Verification / <verification> — observable pass/fail checks against the produced output

Task

Apply to the User's Input

The user invokes this skill with a description of the agent, its purpose, or an existing draft to improve. Specialize every step below to that input.

• What domain is the agent operating in?
• What tools does it have access to (if any)?
• Is it a one-shot prompt, a tool-using prompt, or a fully agentic prompt?
• If the input does not supply enough signal for at least role and task, ask one clarifying question before drafting.
• If the user says "xml", "as xml", or "xml format", produce XML output; otherwise default to Markdown.

Part A: Classify the Prompt Type

Identify the type before writing any section. The type determines which sections are required.

• One-shot: single question or instruction, no tools, no loop → needs role + task only.
• Tool-using: agent has access to specific tools → needs role + behavior + tools + task.
• Agentic: agent runs more than once and may delegate → needs all sections; behavior must include a self-establishing iteration block.
• Retrieval-grounded: agent answers from provided documents → needs role + task with grounding constraints; behavior optional.

Part B: Draft Context

Apply promptcontext principles. Skip if the agent has no runtime environment dependency (e.g. a pure reasoning prompt with no file access, commands, or time sensitivity).

• Derive which fields are needed from what the agent will do: file operations → working directory; OS-sensitive commands → platform; time reasoning → date; version control → git state.
• Include only fields the agent will act on; omit everything it can infer or ignore.
• Static values as literals; dynamic values as {placeholder} names.

Part C: Draft Role

Apply promptrole principles.

• One identity sentence: "You are a [seniority] [domain] [role] who [behavioral expectation]".
• Narrow the domain: specific beats broad.
• One role only: secondary concerns become constraints, not second roles.
• Add the fallback clause: "If you cannot answer confidently, say so."

Omit everything operational (rules, prohibitions, loop directives) — those go in behavior.

Part D: Draft Behavior

Apply promptbehavior principles. Skip for simple one-shot prompts.

• Classify the behavioral context: read-only, write-capable, or agentic.
• Define operational directives (what to do before acting).
• Define prohibition rules (NEVER + concrete action + failure mode).
• Define decision patterns (reversible-first, simplest-first, parallel vs. sequential).
• Define communication style (answer-first, omit preamble, length default, jargon level).
• For agentic roles: open with a loop declaration that names the iteration steps, then add iteration budget, escalation trigger, termination criteria.

Part E: Draft Tools

Apply prompttool principles. Omit if the agent has no tools.

• List every tool with its exact API name and one-sentence purpose.
• For each tool: trigger condition, skip condition, parameters, preference ordering (when overlap exists).
• One usage example per tool.
• Failure handling per tool.

Part F: Draft Task

Apply prompttask principles.

• Open with a precise action verb (Write, Extract, Classify, Compare — never "help me with").
• Frame the problem before the solution; embed constraints upfront.
• Scope the work: what is in, what is out.
• Provide concrete context (file paths, function names, exact values).
• Specify output format, length, and observable success criteria using ### Part subsections.
• Name edge cases explicitly.

Part G: Draft Verification

Apply promptverification principles. Skip only if the agent produces nothing observable (e.g. a pure conversational reply with no artifact and no side-effects).

• List three to six pass/fail checks against the artifact the task produces.
• Each item is a single observable condition, declarative voice, present tense.
• Cover the output, not the agent's process.
• Do not restate task MUST/NEVER directives verbatim — verification names a measurement, not a directive.

Part H: Review Cross-Section Coherence

Before writing the file, check that sections do not conflict.

• Does behavior define a communication style that contradicts the audience named in role?
• Does task reference a tool not listed in tools?
• Does behavior include a loop but role defines a read-only analyst? (Mismatch in autonomy level.)
• Does verification cover what task produces? (Every artifact named in task should have at least one check.)
• Are signal markers (MUST, NEVER, IMPORTANT) consistent across sections — same marker means the same urgency level everywhere?

Fix any conflict before proceeding.

Part I: Write the File

Markdown (default) — write a single prompts/{slug}.md with H1 title and H2 sections in canonical order:

# {Agent Name}

## Context

Working directory: {working_directory}
Platform: macOS

## Role

You are a senior backend engineer who reviews distributed systems with a focus on correctness and performance. If you cannot answer confidently, say so.

## Behavior

- MUST read the file before proposing edits.
- NEVER predict tool results before they return — not as prose, not as structured output.

## Tools

### read_file

Read the contents of a file at a given path. Use before editing any file.

## Task

### Part A: Scope

Extract all API endpoint definitions from `src/routes/` — report file path, HTTP method, route pattern, and handler function name. Skip test files.

### Part B: Output requirements

MUST output a markdown table with columns: File, Method, Route, Handler.
NEVER include functions that are not HTTP endpoint handlers.

## Verification

1. Table contains one row per endpoint.
2. No test file endpoints included.
3. All four columns are populated for every row.

XML — write a single prompts/{slug}.xml with XML-tagged sections in canonical order:

<context>
Working directory: {working_directory}
Platform: macOS
</context>
<role>
You are a senior backend engineer who reviews distributed systems…
</role>
<behavior>
- MUST read the file before proposing edits.
</behavior>
<task>
Extract all API endpoint definitions from `src/routes/`…
MUST output a markdown table.
</task>
<verification>
1. Table contains one row per endpoint.
2. No test file endpoints included.
3. All four columns are populated for every row.
</verification>

Verification

1. File follows canonical section order: Context → Role → Behavior → Tools → Task → Verification (Markdown) or context → role → behavior → tools → task → verification (XML).
2. Each section present matches what this prompt type requires; no extra sections added.
3. No operational rules hiding in role; no identity claims hiding in behavior.
4. No section exceeds one screen — overloaded sections produce ambiguous agent behavior.
5. Signal markers (MUST, NEVER, IMPORTANT) are used consistently across all sections.
6. Format matches user's request: .md file with H1 + H2 sections for Markdown, .xml file with tags for XML.

Derive a slug: kebab-case of the 2–3 most distinctive words in the agent description. Create the prompts/ directory if it does not exist. Write the single combined file. List the file path to the user. No other output.