Prompt Engineer - Context-Aware Prompt Crafting

Craft detailed, context-rich prompts for beads issues by first gathering codebase context via parallel exploration.

Context efficient: Explore agents via Task tool run as subagents and return only summaries. The full exploration happens out of your context - only the findings come back. This skill loads prompting guidelines, then you execute the workflow below.

Workflow

1. Receive issue(s) to prepare prompts for
2. Spawn parallel Explore agents (haiku) to gather context per issue
3. Synthesize findings into monster prompts
4. Return prompts ready for worker spawning

Phase 1: Parallel Context Gathering

For each issue, spawn an Explore agent to gather:

Task(
  subagent_type="Explore",
  model="haiku",
  prompt="For issue [ISSUE-ID]: [title]

  Find:
  - Key files that will need changes (with line numbers for relevant sections)
  - Existing patterns to follow (show examples)
  - Dependencies and imports
  - Related test files
  - Any similar past implementations to reference

  Return structured context, not recommendations."
)

Spawn these in parallel - one per issue. Haiku is fast and cheap.

Phase 2: Prompt Synthesis

Using gathered context, craft prompts following these principles:

Structure

## Task: [ISSUE-ID] - [Title]
[Explicit, actionable description of what to do - be specific about the work]

## Context
[Background and WHY this matters - gathered from exploration]

## Key Files
[Specific file paths with line numbers from exploration]
- `/path/to/file.ts:45-60` - [what's relevant here]
- `/path/to/other.ts:120` - [pattern to follow]

## Patterns to Follow
[Code examples from exploration showing existing patterns]

## Approach
[Implementation guidance based on codebase patterns]

## When Done
Run `/conductor:worker-done [ISSUE-ID]`

DO NOT include /skill:name commands in prompts (except worker-done). Skills auto-activate via UserPromptSubmit hook based on task description. Just describe WHAT needs to be done clearly - Claude will load relevant skills.

Prompt Principles

Be explicit and specific:

Less Effective	More Effective
Fix the bug	Fix the null reference on line 45 of Terminal.tsx when resize fires before init
Improve the UI	Add loading state to button with disabled styling and spinner
Add error handling	Wrap WebSocket connection in try-catch to handle backend disconnections

Add context (explain WHY):

Add try-catch around WebSocket connection to gracefully handle backend
disconnections. Currently users see a cryptic error when the backend restarts.

Reference existing patterns:

Follow the same error handling pattern used in useTerminalSessions.ts
(lines 45-60) for consistency.

Avoid aggressive language - Opus 4.5 follows instructions precisely:

Avoid	Use Instead
CRITICAL: You MUST	Use this when...
NEVER do X	Prefer Y because...
ALWAYS do Y	Default to Y for...

Phase 3: Return Prompts

Return all crafted prompts in a structured format:

## Prepared Prompts

### [ISSUE-ID-1]
```prompt
[Full prompt content]

[ISSUE-ID-2]

[Full prompt content]

## Anti-Patterns

| Problem | Solution |
|---------|----------|
| Vague instructions | Gather specific file paths and line numbers via exploration |
| Missing context | Explain WHY behind each task |
| No patterns referenced | Find and include existing code patterns |
| Over-scoped prompts | Focus on the specific issue, not surrounding cleanup |
| Skill/agent matching | Let auto-hook handle this - don't manually specify skills |

## Model Selection

This skill runs as Opus (orchestrator) but spawns Haiku for exploration:
- **Haiku**: Fast parallel exploration, cheap context gathering
- **Opus**: Synthesis and prompt crafting requiring judgment

---

**Execute this workflow now.** Get the issue IDs from the args or ask for them, then:
1. Spawn parallel Explore agents
2. Synthesize findings into prompts
3. Return the formatted prompts