build

Build (Phase 4-5)

The stage for Context Structuring and Context Compression. Receives the context selected by gather, organizes it into a model-readable form, and retains only the essentials.

---

Prerequisites

Before running, verify gather Artifacts:

Phase 1 analysis result (purpose / constraints / success criteria / Role / scope-declaration)
Phase 3 selection result (list of Keep items)

Verification order:

1. Do the above Artifacts exist in the conversation context? → Use them if present
2. If not, check whether the _phase1-result.md file exists → If present, read it to restore the Phase 1 result
3. If the Phase 1 result has been restored but the Phase 3 selection result (Keep items) is missing:

   "Please run /context-engineering:gather first. (Phase 3 selection result is required)" Stop.

4. If both are missing:

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

---

Phase 4. Context Structuring

Purpose: Organize the selected context into a section structure that is easy for the model to read.

Structure

## Key Facts
- {key fact} [source: {source name}]

## Constraints
- {constraint} [source: {source name}]

## Decisions
- {decision and rationale} [source: {source name}]

## Notes
- {supplementary information} [source: {source name}]

[source: ...] tag rules:

• The source name is the identifier of the source collected in Phase 2 (filename, URL, "user", "KB:{entry-title}")
• Items derived from multiple sources: [source: file.md, user]
• Items derived from general knowledge: [source: general]
• During Phase 7 speculation checks, [source: general] items are treated as primary review targets

Structuring Rules

• Each item expresses a single fact / constraint / decision
• Empty sections are omitted entirely
• No Omission between items — if the same content appears in two sections, consolidate it into the more appropriate section
• Key Facts includes only items directly related to the Phase 1 purpose
• Attach [source: ...] tags to all items — they serve as the basis for Phase 7 speculation verification

G4 Gate

The AI automatically evaluates the following criteria:

Criterion	Evaluation
No ambiguous items	Does each item express a single clear fact / constraint / decision?
No Conflict between sections	Does the same content not exist in two sections?
Purpose reflected	Do Key Facts connect to the Phase 1 purpose?
Source tags attached	Do all items have a [source: ...] tag?

Auto-pass: Criteria clearly met → proceed to Phase 5

User confirmation (when ambiguous or unmet):

Phase 4 result: {summary}
  Unmet items: {item name} — {reason}
  Please revise or approve to continue.

---

Phase 5. Context Compression

Purpose: Remove unnecessary parts, retaining only essential facts, Constraints, and Decisions.

Token Budget Estimation

Measure the context size before compression:

1. Measure: Estimate the total word count of Keep items using wc -w

2. Budget decision:

   Task scale	Word count basis	Target compression ratio
   Small (single file / question)	< 500 words	No compression needed
   Medium (module / feature unit)	500–2,000 words	Compress to 50% or less
   Large (entire project)	2,000+ words	Compress to 30% or less

3. Strategy when over budget: Summarize Notes section first → condense Key Facts → condense Constraints last

Budget Example Scenarios

Scenario	Input	Target	After compression	Main method
Single API question	300 words	—	300 words	No compression needed
Module refactoring	1,200 words	≤50%	~550 words	Summarize Notes + remove duplicates
Full project analysis	4,000 words	≤30%	~1,100 words	Remove Notes + condense Key Facts

Incremental Loading

When a source is large (2,000+ words):

• Do not read the full text at once
• First scan only the file list / table of contents / headers
• Selectively read only the sections relevant to the Phase 1 purpose
• Record the path and line range of each section read

Compression Rules

• Remove duplicates: If items with the same meaning appear multiple times, merge them into one
• Summarize Tangential information: Content indirectly related to the Phase 1 purpose → replace with a one-line summary
• Preserve essentials: Constraints / Decisions items retain original text — these are not subject to compression
• Key Facts priority: Facts directly related to the purpose retain original text

G5 Gate

The AI automatically evaluates the following criteria:

Criterion	Evaluation
Constraints preserved	Are all Constraints still present after compression?
Decisions preserved	Are Decisions and their rationale retained without loss?
Over-compression prevention	Has information needed to achieve the Phase 1 purpose not been removed?
Budget met	Is the post-compression word count within the target budget?

Auto-pass: Criteria clearly met → pass to compose

User confirmation (when ambiguous or unmet):

Phase 5 result: {summary}
  Unmet items: {item name} — {reason}
  Please revise or approve to continue.

---

After Completion

Take the build Artifacts (structured and compressed context block) and proceed to the next stage:

/context-engineering:compose   — Run Phase 6 execution instruction generation