| name | prompt-engineer |
| description | Generate self-contained, agent-ready prompts for delegating work to fresh Claude sessions or sub-agents. Use whenever you need to hand off a task (fix, feature, debug, refactor) to another agent that has no prior context. Also use when crafting sub-agent prompts within workflows, or when any skill/workflow needs a high-quality delegation prompt. Trigger on: "generate a prompt for", "write a prompt to", "hand this off to another agent", "create a fix prompt", or any context where work must be delegated to a fresh agent with zero shared context. |
Prompt Engineer
Generate prompts that let a fresh agent act without back-and-forth. Every prompt you produce must be self-contained — the receiving agent has zero context from this conversation.
Which mode applies
Check first: is there already a structured task file behind this delegation (e.g. from /feature-plan's task decomposition)? If yes, use Task-file mode below — it's cheaper and more accurate than reasoning from scratch. If no structured task file exists — an ad-hoc delegation, a fast-path task, or a request from outside the feature workflow entirely — use the From-scratch mode that follows it.
Task-file mode
When a task file already exists, most of the five core questions are already answered in it:
| Question | Where it already lives in the task file |
|---|
| What | "What to do" section |
| Where | "Files" section |
| Why | The feature's CONTEXT.md, referenced by the task |
| Constraints | "Constraints" section |
| Verification | "Verification" section |
Your job shrinks to three things:
- Transcribe the task file's content into the standard dispatch prompt structure below — don't re-derive or re-word what's already precise; copy exact paths, exact commands, exact interface names as written.
- Select the model tier from the task file's own
Model tier field, then resolve it to an actual model per the Model Tier Mapping below. Never omit this and never let it default to inheriting the parent session's model — an omitted model silently uses the most capable and most expensive option, which defeats the entire point of tiering. If the task file says cheap, the task's "What to do" should already contain complete, specific code — the implementer's job is transcription plus verification, not judgment. If mid, the implementer needs to apply judgment to a prose description. Reserve quality for tasks the planner explicitly flagged as architecturally significant.
- Add anything assumed but not restated. If the task file relies on a project-wide convention from the feature's
CONTEXT.md (a Global Constraints line, an established pattern) without repeating it verbatim, add it explicitly here — the receiving subagent won't see CONTEXT.md, only what you hand it.
Do not re-litigate the task's scope or re-interview for missing information in this mode — if the task file is genuinely ambiguous, that's a planning-quality problem to flag back to /feature-plan, not something to paper over with your own guesses here.
From-scratch mode
Use this when no structured task file exists. A good delegation prompt answers five questions for the receiving agent:
- What — the specific outcome expected
- Where — exact file paths, line numbers, or components involved
- Why — enough context to make judgment calls (not just follow instructions)
- Constraints — what must NOT change, what to avoid, boundaries
- Verification — how the agent confirms it worked
If any of these are missing, the receiving agent will either ask questions (wasting time) or guess wrong (creating bugs).
Model Tier Mapping
This is the canonical tier-to-model mapping — defined here, and nowhere else in the workflow. Every dispatch prompt resolves its tier through this table before it ships:
| Tier | Model |
|---|
cheap | haiku |
mid | sonnet |
quality | opus |
These are the Agent tool's model parameter alias names, not dated model IDs, so the mapping keeps working across model releases without edits. Never resolve to fable, and never leave model omitted or inherited — an omitted model silently pulls the parent session's (usually most expensive) model, which defeats the entire point of tiering.
Prompt Structure
Use this structure for all delegation prompts, in either mode. Adapt section depth to task complexity — a one-line label fix needs less than a navigation flow rewrite.
## Context
[1-3 sentences: what the project is, what area of code this touches, why this matters]
## The Problem
[What's wrong or what's needed. Be specific and observable — "the category label is missing below supplement icons in StackView" not "labels are broken"]
## Files Involved
[Exact paths. If you know line numbers, include them. If uncertain, say which file to search in and what to search for]
## What to Do
[Step-by-step if multi-step, or a clear single instruction. Include the expected end state]
## Constraints
[What NOT to change. Adjacent features to leave alone. Patterns to follow. Design system tokens to use. Architecture to respect]
## Verification
[How to confirm the fix works — build command, what to check in simulator, expected behavior]
## Model
[Explicit tier or model ID — tier resolved to a model per Model Tier Mapping above; never left to inherit]
## Attachments
[Note any screenshots, error logs, or reference files being provided alongside this prompt]
Calibration by Task Size (from-scratch mode)
Trivial Fix (missing label, wrong color, typo)
- Context: 1 sentence
- Problem: 1 sentence with exact location
- Files: 1-2 paths
- What to Do: single clear instruction
- Constraints: 1 line ("don't change layout")
- Verification: "build and confirm X appears"
- Model: cheap tier
- Total: ~8-12 lines
Medium Fix (add tap navigation, wire up data binding, fix a flow)
- Context: 2-3 sentences covering the feature area
- Problem: describe current vs expected behavior
- Files: 2-5 paths with relationships explained
- What to Do: numbered steps
- Constraints: adjacent features, architectural patterns
- Verification: specific user flow to test
- Model: mid tier unless the steps already fully specify the code
- Total: ~20-40 lines
Large Task (new feature, significant refactor)
- Context: full paragraph on the system and motivation
- Problem: detailed gap analysis
- Files: file tree of relevant area + key patterns to follow
- What to Do: phased approach with intermediate checkpoints
- Constraints: comprehensive (design system, architecture, existing tests)
- Verification: multiple scenarios including edge cases
- Model: quality tier, or split into smaller task-file-mode delegations instead — a task this large is usually a sign it should go through
/feature-plan rather than a single from-scratch prompt
- Total: ~50-80 lines
Sub-Agent Prompts
When crafting prompts for sub-agents spawned within a workflow (not fresh sessions):
- Sub-agents inherit some context from parent but not conversation history
- Be explicit about what the sub-agent should return (format, length, structure)
- Include the specific question or task — don't rely on ambient context
- Specify whether the sub-agent should write code or just research/analyze
- Always specify the model explicitly, as a resolved model per the Model Tier Mapping above — see Task-file mode above for how the tier itself is decided when a task file exists
Anti-Patterns
Avoid these — they produce prompts that fail:
| Anti-Pattern | Why It Fails | Fix |
|---|
| "Fix the bug" | No specifics — agent has to explore blindly | Name the exact symptom and location |
| Assuming shared context | Fresh agent knows nothing about prior work | State everything explicitly |
| Vague verification | "Make sure it works" — how? | Specific observable outcome |
| Over-specifying implementation | Leaves no room for judgment on approach | Describe the outcome, not every keystep |
| Missing constraints | Agent "fixes" one thing and breaks another | Name what must not change |
| No file paths | Agent wastes time finding the right files | Anchor with exact paths |
| Omitted model tier | Silently inherits the most expensive model | Always state it explicitly, from the task file if one exists |
