gather

Gather (Phase 1-3)

The starting point of context engineering. Clarify user intent, collect candidates from relevant sources, and select only what is needed.

---

Phase 1. Problem Definition

Purpose: Clarify what the user wants from the AI. This becomes the baseline for all subsequent Phases.

Query Approach

Ask in purpose → constraints → success criteria order, one at a time, using multiple-choice format when possible.

Question 1 (purpose — confirming the Task):

"What is the problem you are trying to solve right now?"

Question 2 (constraints — confirming Phase 2 sources + Constraints):

"What background information would the AI need to do this task well? (e.g., existing code, documentation, prior decisions)"

Question 3 (success criteria — confirming Output Format):

"What form should the output take? (e.g., execution instruction, saved note, project documentation)"

Additional question: Permitted only once if ambiguity remains unresolved after the 3 questions. No more than 4 questions total.

Scope Reduction

A processing step automatically performed by the AI after collecting answers to the 3 questions (not an additional question):

A. Problem scope reduction

• Analyze the purpose answer to identify the "minimum problem to solve in this session"
• Already small enough → auto-pass
• Can be decomposed → present 2-3 options, let the user choose

B. Source scope declaration

• Classify expected sources from the constraints answer:
  • Must-have: Sources directly mentioned or clearly needed — collect first in Phase 2
  • Nice-to-have: Sources that may indirectly help — collect after Must-have if budget allows

C. Output Session Scope declaration block

--- Session Scope ---
Minimum problem: {reduced purpose}
Must-have sources: {list}
Nice-to-have sources: {list}
Excluded: {explicitly out of scope}

Role Auto-Inference

Automatically infer the role the AI should take from the three answers. Do not ask the user directly about the role.

G1 Gate

The AI automatically evaluates the following criteria:

Criterion	Evaluation
purpose clear	Is the problem to be solved specifically defined?
constraints clear	Is the scope of background information needed understood?
success criteria clear	Is the form of the output confirmed?
Role inferable	Can the AI's role be inferred from the three answers?
scope reduction complete	Has the Session Scope declaration block been generated?

Auto-pass: Criteria clearly met → proceed to Phase 2

User confirmation (when ambiguous or unmet):

Phase 1 result: {summary}
  Unmet items: {item name} — {reason}
  Please revise or approve to continue.

---

Phase 1 Result Preservation Against Compaction

After G1 passes, record in _phase1-result.md to prevent loss during context compaction:

--- Phase 1 Result ---
purpose: {confirmed purpose}
constraints: {confirmed constraints}
success criteria: {confirmed success criteria}
Role: {inferred role}
scope-declaration: {session scope declaration block}

At the start of Phase 2, if Phase 1 results are absent from the conversation context, restore from _phase1-result.md. Delete after pipeline completion (G6 pass or G7 pass).

---

Phase 2. Context Candidate Collection

Purpose: Determine sources based on Phase 1 answers and collect candidate context.

Automatic Source Determination

Collect Must-have sources from Phase 1 Session Scope declaration first, then add Nice-to-have sources when budget allows.

Determine sources from the constraints answer signals in Phase 1:

Phase 1 Signal	Collection Source
Document / file / URL provided	Read full text of that file / fetch URL
Project / codebase mentioned	Explore codebase, existing spec, CLAUDE.md
"저장해줘" (save it) / user data input	Full text of the input
Question / task request	Scan KB index.md → read matching entries
Composite (multiple signals mixed)	Combine above sources — run sub-agents in parallel

Strategy classification (RAG / Memory / Tool Result / System Prompt) → Context Source Strategy

KB Search Enhancement

Check KB existence:

1. Search for the index.md file path (use the configured KB path if set, otherwise scan common locations)
2. If index.md does not exist:
   • Skip all 4 KB search steps below
   • Notify: "KB is not configured. Proceeding without KB." then continue collection using only other sources from the Phase 1 signal table
   • Reference: Failure Cases — Scenario 7
3. If index.md exists → execute the 4 KB search steps below:

When finding candidates in KB index.md, search in the following order:

1. Keyword matching: Scan the index.md table using core words from Phase 1 purpose
2. Tag expansion: Read the tags field of matched entries and add other entries with the same tags as candidates
3. Related entry tracking: Follow the related field of matched entries to add linked entries as candidates
4. Reliability priority: When multiple entries cover the same topic, prioritize reliability: high

If search results exceed 5, filter in Phase 3.

Dynamic Context Injection

For project/codebase sources, dynamically check the latest state at collection time:

• Recent changes: git log --oneline -5
• Project structure: find . -name "*.md" -maxdepth 2 -not -path "./.git/*"
• Build file status: head -20 package.json or head -20 build.gradle (if the file exists)

Include results in candidate context to use as a basis for Phase 3 Keep/Skip decisions.

Collection Precautions

• If a source exceeds 10,000 words, check whether to split by section
• If 0 candidates found in KB, notify immediately: "No relevant entries found in stored knowledge. Proceeding with general knowledge."
• For large sources, use progressive loading: scan table of contents/headers first → selectively collect only purpose-relevant sections
• Record the size (word count) of each collected source — used for build Phase 5 budget estimation

G2 Gate

The AI automatically evaluates the following criteria:

Criterion	Evaluation
Source completeness	Have all sources needed to achieve Phase 1 purpose been collected?
No obvious omissions	Are all materials mentioned in the constraints answer included?

Auto-pass: Criteria clearly met → proceed to Phase 3

User confirmation (when omissions are suspected):

Phase 2 result: {summary}
  Collected sources: {source list}
  Unmet items: {item} — Is this source also needed?
  Please revise or approve to continue.

---

Phase 3. Context Selection

Purpose: Retain only what is actually needed for Phase 1 purpose from the collected candidates.

Evaluate Against 3 Criteria

Criterion	Description
Relevance	Is it directly connected to Phase 1 purpose?
Recency	If more recent information exists, replace older content
Reliability/Confidence	Quality of the source (direct statement > official docs > indirect reference)

Classification

Classify each candidate item:

• Keep: Meets Relevance, Recency, and Reliability criteria → pass to next stage
• Skip: Does not meet criteria — specify reason
• Merge: 70%+ semantic overlap with an existing Keep item → consolidate into the higher-confidence item

G3 Gate

The AI automatically evaluates the following criteria:

Criterion	Evaluation
Criteria applied consistently	Were Relevance, Recency, and Reliability applied equally to all items?
Skip reasons clear	Does each skipped item have a stated reason?
Keep items sufficient	Does Keep contain the information needed to achieve Phase 1 purpose?

Auto-pass: Criteria clearly met → pass to build

User confirmation (when ambiguous or unmet):

Phase 3 result: {summary}
  Keep: {N}  Skip: {N}  Merge: {N}
  Unmet items: {item name} — {reason}
  Please revise or approve to continue.

---

After Completion

With gather artifacts (Phase 1 analysis + selected context set), proceed to the next stage:

/context-engineering:build   — Phase 4-5 structuring and compression