build-prompt-chain

Transform an implementation prompt into a phased execution chain: $ARGUMENTS

Purpose

This skill transforms monolithic implementation prompts (typically from /handoff-prompt-to) into a sequence of phases designed for sustained-context execution via /super-agent.

The output is:

• An orchestration prompt (00-orchestrate.md) that a fresh Claude instance follows
• Phase files (01-*.md, 02-*.md, etc.) that the orchestrator feeds to a /super-agent worker session

Why chain? Complex tasks benefit from checkpoints. Discovery before implementation. Review before verification. The orchestrator uses /super-agent session-resume to maintain full context across phases while giving you control between each checkpoint.

Input Resolution

• File path provided: Read the prompt from that file
• "above" or empty: Look for the most recent implementation prompt in this conversation:
  1. If /handoff-prompt-to just ran and announced "Saved to [path]" → read that file
  2. If a prompt was displayed inline → use that content
  3. If a file path was mentioned → read that file

If no recognizable prompt is found, ask: "I don't see a structured implementation prompt in this conversation. Please provide a file path or run /handoff-prompt-to first."

---

Analysis Phase

Before generating, analyze the source prompt:

1. Size Check

If the source prompt is small (single focused task, <50 lines of actionable content), respond:

"This prompt is compact enough to run monolithically. Decomposition would add overhead without benefit."

Only proceed if decomposition adds value.

2. Ambiguity Resolution

Scan for phrases that introduce optionality:

• "Optionally...", "If time permits...", "Consider adding...", "You may want to..."

First: Check if the source prompt itself resolves the ambiguity:

Pattern in Source	Resolution
"Optionally (but recommended)"	Include — recommended means do it
"optional... off by default" or "behind a flag"	Skip — low priority, not core
"If X, then Y"	Check if X is true in context
"Consider adding... for [benefit]"	Include if benefit aligns with stated goals
Detailed spec follows the "optional" mention	Include — effort was spent specifying it
Mentioned in Definition of Done	Include — it's a success criterion

Only ask the user if genuinely unresolvable after checking the source. Most well-written prompts contain their own answers.

Document any scope decisions (resolved or asked) in the orchestration file's "Scope Decisions" section.

3. Phase Detection

Identify natural breakpoints in the work:

Signal in Source	Suggests Phase
"Read/understand/map" language	Discovery phase
Multiple distinct file groups	Separate implementation phases
"Then wire up / connect / integrate"	Integration phase
Test commands, verification steps	Verification phase
Migration + cleanup	Two-phase implementation
Sub-agent spawning instructions	Integrate into relevant phase

Determine the right number and type of phases (typically 2-4). Don't force a fixed structure.

4. Coverage Balance Check

Compare detail levels across sections. If one area (e.g., backend) has 3x more detail than another (e.g., frontend), flag it:

"Note: Backend has detailed specs; frontend section is sparse. The generated phases will reflect this imbalance."

5. Extract Critical Details

Before writing phases, extract and set aside:

• Code snippets/skeletons — preserve exactly, don't paraphrase
• Protocol specs — headers, formats, event structures
• Exact commands — test commands, build commands
• Exceptions/nuances — "keep this one", "don't do X here", edge cases

These MUST appear verbatim in the relevant phase, not summarized.

---

Output Structure

Directory: .ai-reference/prompts/<task-name>-chain/

Derive <task-name> from the source prompt's title or main goal (e.g., "add-auth-flow", "refactor-parser"). Use kebab-case, keep it short.

00-orchestrate.md       # THE prompt to paste into fresh agent
01-<phase-name>.md      # First phase (e.g., 01-discovery.md)
02-<phase-name>.md      # Second phase (e.g., 02-implement-backend.md)
...

Phase names should be descriptive: 01-discovery, 02-implement-api, 03-wire-frontend, 04-verify.

---

Orchestration Prompt Format (00-orchestrate.md)

The orchestration file is a prompt for an orchestrator — paste it into a fresh Claude window. That Claude instance will use /super-agent to spawn a worker session and drive it through each phase, maintaining context across the entire chain.

Execution model:

Fresh Claude (orchestrator)
    ↓ reads 00-orchestrate.md
    ↓ reads 01-discovery.md (to understand it)
    ↓ runs: /super-agent "Execute <path>/01-discovery.md"
    ↓ captures session-id
    ↓ worker reads file, executes, returns results
    ↓ orchestrator reviews results
    ↓ reads 02-backend.md, optionally edits based on Phase 1 findings
    ↓ runs: /super-agent --resume <session-id> "Execute <path>/02-backend.md"
    ↓ worker continues with full context
    ↓ ... repeat, reviewing and adjusting files between phases

# [Task Name] — Phased Implementation

You are the **orchestrator**. Use `/super-agent` to execute this implementation as a sequence of phases, maintaining a single worker session across all phases.

## Overview
[One sentence: what we're building and why]

## Scope Decisions
[If any ambiguous items were resolved, document the decisions here]

## Phase Sequence

| Phase | File | Goal | Done When |
|-------|------|------|-----------|
| 1 | `01-discovery.md` | Map code, create plan | Plan output |
| 2 | `02-implement.md` | Execute the plan | Files changed |
| 3 | `03-verify.md` | Prove correctness | Tests pass |

All phase files are in: `.ai-reference/prompts/<task-name>-chain/`

## Key Constraints
[Critical constraints that apply across all phases — preserve from source]

## Execution Instructions

### Phase 1: Start the Chain

1. Read `01-discovery.md` to understand what it asks
2. Run `/super-agent` with the file path:

/super-agent "Execute .ai-reference/prompts/<task-name>-chain/01-discovery.md"

3. **Save the session-id** from the response

### Subsequent Phases

When a phase completes:

1. **Review** the super-agent's output
2. **Read** the next phase file (e.g., `02-backend.md`)
3. **Edit the file** if needed based on prior phase results
4. **Resume** with the file path:

/super-agent --resume <session-id> "Execute .ai-reference/prompts/<task-name>-chain/02-backend.md"

The worker retains full context — it knows what happened in prior phases.

### If a Phase Fails

Resume the same session to fix:

/super-agent --resume <session-id> "The previous phase had issues: [describe]. Fix them and re-verify."

Do NOT start a new session — context is critical for correct fixes.

## Begin

Read `01-discovery.md`, then run `/super-agent` with its path.

The orchestration prompt is self-contained — the orchestrator Claude knows:

• How to spawn and resume the worker session
• What files to pass for each phase
• When to pause for review
• How to recover from failures while preserving context

---

Phase Design Principles

1. Sustained Context

Each phase assumes the agent remembers everything from prior phases.

• Do NOT repeat file contents or patterns already loaded
• Reference prior work: "Using the plan from Phase 1..."
• Keep phases lean - they're continuations, not fresh starts

2. Checkpoints

Every phase except the last ends with:

## Checkpoint
[What to output/confirm before proceeding]
Do not continue to the next phase. Wait for the next prompt.

3. Clear Scope

Each phase has ONE clear goal. If you're combining "understand + implement + test" in one phase, it's too big.

4. Observable Completion

Each phase must produce observable output so the orchestrator knows it's done:

• Discovery → outputs a plan
• Implementation → outputs file changes summary
• Verification → outputs test results

5. Preserve Technical Fidelity

Critical: Do not summarize away technical details. If the source has:

• Protocol specs (headers, formats) → include verbatim in relevant phase
• Code skeletons → include verbatim, not paraphrased
• Exact error handling patterns → preserve exactly
• Exceptions ("keep this refresh call") → call out explicitly

---

Phase Content Guidelines

Discovery Phase (when needed)

# Phase 1: Discovery

## Goal
[Understand X / Map Y / Create plan for Z]

## Read These Files
[List - be specific]

## Sub-Agent Tasks (if applicable)
[Spawn Explore agents for specific discovery tasks - be explicit about what to spawn]

## Output Required
[Plan format / Analysis format / etc.]

## Checkpoint
Output your [plan/analysis]. Do not implement yet. Wait for the next prompt.

Implementation Phase(s)

# Phase N: [Specific Implementation Goal]

## Goal
[What to build/change - reference prior phases]

## Scope
[Specific files/components - NOT everything]

## Technical Spec
[Include any protocol details, code skeletons, exact formats from source - VERBATIM]

## Exceptions & Nuances
[Call out any "keep this", "don't change that", edge cases]

## Guidelines
[Patterns, constraints relevant to THIS phase]

## Checkpoint
Summarize changes made. Do not proceed to [next thing]. Wait for the next prompt.

Verification Phase (when needed)

# Phase N: Verification

## Goal
Prove the implementation is correct.

## Steps
1. [Build/compile command - exact]
2. [Test command - exact]
3. [Behavioral verification - specific scenario]

## Sub-Agent Tasks
[If source specifies browser-qa or other verification agents, include explicit spawn instructions]

Example:
Spawn a `browser-qa` agent with this prompt:
> Navigate to http://localhost:8765. Verify [specific behavior]. Capture screenshot.

## Success Criteria
[What "done" looks like]

## If Issues Found
Fix and re-verify. Document what you fixed.

---

Transformation Heuristics

What goes where:

Source Content	Target
Task description, context, "why"	Orchestrate overview + Phase 1
Files to read/reference	Phase 1 (Discovery)
Files to create/modify	Implementation phase(s)
Execution steps 1-N	Split across implementation phases by natural grouping
Code snippets/skeletons	Relevant implementation phase — VERBATIM
Protocol specs	Relevant implementation phase — VERBATIM
Edge cases / exceptions	Relevant implementation phase (explicit section)
Sub-agent strategy	Integrate as spawn instructions in relevant phases
Test commands, verification	Final phase
Definition of done	Runbook "Done When" + final phase

Splitting implementation:

• By component (backend → frontend)
• By operation (create → wire up → configure)
• By risk (core logic → integration → polish)

Choose the split that minimizes cross-phase dependencies.

Handling Sub-Agent Instructions

If the source prompt has a "Sub-Agent Strategy" or similar section:

1. Don't put it in a separate advisory section
2. Do integrate spawn instructions directly into the phases where they apply
3. Use explicit Task tool syntax:

## Sub-Agent Tasks

Spawn these agents in parallel:

1. **Explore agent**: "Find all calls to refresh() and map the call graph"
2. **Explore agent**: "Identify threading patterns (locks, events, conditions)"

Wait for results before proceeding.

---

Requirements

1. Lean phases: Each phase ≤40% of original prompt length
2. No redundancy: Trust sustained context - reference, don't repeat
3. Executable runbook: Someone unfamiliar can follow it
4. Observable checkpoints: Each phase produces confirmable output
5. Technical fidelity: Code snippets, specs, commands preserved verbatim
6. Explicit scope: Ambiguities resolved before generation
7. Integrated sub-agents: Spawn instructions woven into phases, not advisory

---

After Generating: Provide Next Steps

After writing all chain files, output:

Created prompt chain in `.ai-reference/prompts/<task-name>-chain/`:

  00-orchestrate.md  ← Paste into a fresh Claude window to start
  01-discovery.md
  02-implement.md
  03-verify.md

## To Execute

Paste `00-orchestrate.md` into a fresh Claude Code window. It will:
1. Use /super-agent to spawn a worker session
2. Pass each phase file to the worker in sequence
3. Maintain context across all phases via session-id resume

The orchestrator controls pacing — review results between phases.

Scope decisions made:
- [any ambiguity resolutions]

Technical specs preserved:
- [list of verbatim sections carried forward]

Keep it simple. The orchestration file IS the deliverable — it contains the /super-agent invocations needed to execute the chain.