compose

Compose (Phase 6)

Final output generation stage. Combines the structured and compressed context from build with Phase 1 success criteria to assemble the appropriate output.

---

Prerequisites

Before execution, verify build artifacts:

Phase 1 analysis results (purpose / constraints / success criteria / Role / scope-declaration)
Structured and compressed context blocks (Key Facts / Constraints / Decisions / Notes)

Verification order:

1. Do the above artifacts exist in the conversation context? → Use them if present
2. If not, check whether _phase1-result.md exists → Read it to restore Phase 1 results if found
3. If Phase 1 results are restored but build artifacts (structured/compressed blocks) are missing:

   "Please run /context-engineering:build first. (Structured/compressed blocks are required)" Stop.

4. If both are missing:

   "Please run /context-engineering:gather first." Stop.

---

Output Budget Guide

Target size for Phase 6 output:

Output Format	Target Size	If Exceeded
A. Execution Instruction	200–800 words	Condense Notes to one-line summaries → if still exceeded, split artifacts
B. KB Entry	20–80 lines	Split the entry (see entry-template guidelines)
C. Project Artifacts	CLAUDE.md 100–150 lines	If over 150 lines, extract to a separate docs/ file

Immutable rule: Constraints and Decisions must not be deleted or condensed even under output budget pressure. These items were already guaranteed to be preserved in build Phase 5.

---

Phase 6. Execution Instruction Generation

Purpose: Automatically determine the output format based on the Phase 1 success criteria signal and generate the final output.

Automatic Output Format Determination

Read the signal from the Phase 1 success criteria answer and determine the format:

success criteria signal	Output Format
"지시해줘" (give me instructions), "프롬프트 만들어줘" (make me a prompt), "어떻게 써야 해" (how should I write it)	A. Execution Instruction
"저장해줘" (save it), "기억해줘" (remember it), "노트해줘" (note it)	B. KB Entry
"프로젝트 시작" (start a project), "개발하고 싶어" (I want to develop), "코드베이스 분석" (analyze codebase)	C. Project Artifacts

After automatic determination, confirm exactly once:

"Output format: {decided format name} ({A/B/C}). Enter A/B/C to change."

• If the user responds affirmatively ("yes", "proceed", Enter, etc.) → proceed with the determined format
• If A/B/C is entered → switch to that format

If the signal does not match any row in the table, present choices:

"Please select output format: (A) Execution Instruction (B) KB Entry (C) Project Artifacts"

---

G6 Gate

After compose completes, the AI automatically evaluates:

Criteria	Evaluation
success criteria match	Does the output match the result form defined in Phase 1?
No omissions	Are all Phase 1 purpose/constraints reflected in the output?
No conflicts	Are there no contradictions within the output?
No speculation	Does the output contain no unsupported claims?

G6 Pass: Complete — present the output. Delete _phase1-result.md if it exists.

G6 Fail: Automatically invoke Phase 7

G6 unmet: {criteria} — {reason}
Running verify for detailed validation.

---

Refer to the sections below for the detailed structure of each format.

---

A. Execution Instruction (System Prompt)

Assemble using Role + Task + Constraints + Output Format structure:

## Role
{AI role inferred from Phase 1 answers and context}

## Task
{Phase 1 purpose — describe the problem to be solved concretely}

## Constraints
{All Constraints items from Phase 5}
{Additional constraints derived from Phase 1 constraints}

## Output Format
{Form of the result defined in Phase 1 success criteria}

Writing principles:

• Role: One sentence, described as a specific expert role
• Task: Described at an actionable level without ambiguity
• Constraints: List form; each item understandable on its own
• Output Format: Specify format, length, and structure

Tool Context (Execution Instruction only)

If the Phase 1 purpose requires tool usage, add a ## Tools section to the instruction:

## Tools

| Tool | Purpose | Usage Example |
|------|---------|---------------|
| {tool name} | {purpose for this task} | {call example or pattern} |

MCP Tools (if applicable):

• context7: Library documentation lookup — resolve-library-id → get-library-docs
• playwright: Web page verification — browser_navigate → browser_snapshot

Inclusion criteria (based on Phase 1 purpose signals):

• "code", "implement", "test" signals → include Bash, Read, Write, Agent
• "documentation", "library" signals → include context7 MCP
• "web", "browser", "UI" signals → include playwright MCP
• No tool-usage signal → omit ## Tools section

Context Retention Instruction (Execution Instruction only)

If the execution instruction (Format A) directs a task with 3 or more steps, add the following block at the bottom of the instruction:

## Context Checkpoints

If this task has 3 or more steps:
1. After each major step completes, record in `_context-session.md`:
   - What has been completed so far
   - Key facts discovered
   - Information needed for the next step
2. Before starting the next step, re-read `_context-session.md`
3. After all steps are complete, delete `_context-session.md`

Reference: Context Session Template

---

B. KB Entry

Save in standard knowledge entry format:

---
title: {descriptive title}
domain: {domain-slug}
type: fact | decision | constraint | procedure | reference
source: user | document:{filename} | url:{https://...}
date: {YYYY-MM-DD}
reliability: high | medium | low
tags: [tag1, tag2]
related: []
---

# {Title}

> {one-line summary}

## Key Facts
{Key Facts items from build artifacts}

## Constraints
{Constraints items from build artifacts}

## Decisions
{Decisions items from build artifacts}

## Notes
{Notes items from build artifacts}

Save path: {KNOWLEDGE_PATH}/{domain}/{slugified-title}.md

After saving, update index.md:

• Add a row to the relevant domain table
• If the domain section does not exist, create a new section in alphabetical order
• Recalculate header counts (Total entries, Domains)
• Maintain descending date sort order

Reference: Entry Template | Index Template

---

C. Project Artifacts

Generate three documents for project development:

CLAUDE.md (AI behavior policy — Persistent System Prompt):

# {PROJECT_NAME}

This file provides guidance to Claude Code when working with code in this repository.

> {one-line project description}

## Architecture
{Architecture description based on Phase 4 Key Facts + Decisions}
{Layer structure table}

## Build & Test
{Build and test commands}

## Core Constraints
{Phase 4 Constraints items — up to 5, TC/BL prioritized}

## Non-obvious Gotchas
{Easy-to-miss pitfalls from Phase 4 Notes}

## References
- [Knowledge Base]({OUTPUT_PATH}/knowledge-base.md)

Reference: CLAUDE.md Template | Knowledge Base Template

spec.md (Technical Specification):

# {PROJECT_NAME} Spec

## Architecture Decisions
| Decision | Alternatives | Choice | Rationale |

## Package Structure
{Package structure by layer}

## Implementation Plan
| # | Item | REQ | Depends | Done Criteria |

## Requirements Traceability
| REQ-ID | Feature | Implementation | Status |

Reference: Spec Template

Implementation Plan: Describe the first implementation item at a level where work can begin immediately.