context-engineering

Context Engineering

A 7-Phase pipeline for systematically building context to send to AI. Handles knowledge accumulation, prompt assembly, and project document generation in a single flow.

---

Overall Flow

[input]
  ↓
Phase 1 →[G1]→ Phase 2 →[G2]→ Phase 3 →[G3]
                                          ↓
                         Phase 4 →[G4]→ Phase 5 →[G5]
                                                   ↓
                                         Phase 6 →[G6]→ [output]
                                                          ↓ (G6 fail)
                                                        Phase 7

---

Gate Behavior

After each Phase completes, the AI automatically evaluates the criteria:

[Auto-pass]      Criteria clearly met → proceed to next Phase without notifying the user

[User confirmation] Ambiguous or criteria unmet →
  "Phase N result: {summary}
   Unmet items: {content}
   Please revise or approve to continue."

---

Sub-skill Entry Points

Each phase can be run individually:

Command	Responsible Phase	Purpose
/context-engineering:gather	Phase 1-3	Problem Definition → Candidate Collection → Selection
/context-engineering:build	Phase 4-5	Structuring → Compression
/context-engineering:compose	Phase 6	Execution Instruction Generation
/context-engineering:verify	Phase 7	Final Verification (re-runnable)

Mid-entry condition: Only allowed when the previous sub-skill's artifacts exist. If artifacts are missing, the user is guided to run the prerequisite sub-skill first.

---

Full Execution (/context-engineering)

Calls sub-skills in order:

Step 1: Run gather

/context-engineering:gather — Phase 1-3 + G1-G3

Completion criteria:

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

Step 2: Run build

/context-engineering:build — Phase 4-5 + G4-G5

Completion criteria:

• Structured and compressed context block (Key Facts / Constraints / Decisions / Notes (if applicable))

Step 3: Run compose

/context-engineering:compose — Phase 6 + G6

Completion criteria:

• G6 pass: final output (instruction / KB entry / project artifacts)
• G6 fail: automatically move to Phase 7

Step 4: Run verify (when G6 fails)

/context-engineering:verify — Phase 7

Completion criteria:

• Confidence indicator (H/M/L)
• Feedback Loop guidance by issue type
• Self-scoring rubric score (0-100)
• Results recorded in _verify-log.md

---

Output Types

Automatically determined based on Phase 1 success criteria:

Purpose	Output
Execution instruction	Role + Task + Constraints + Output Format
Knowledge storage	KB entry + index.md update
Project development	CLAUDE.md + spec.md + implementation plan

Automatically determined in Phase 6, then confirmed by the user. See compose SKILL.md for details.

---

Pipeline failure scenarios + recovery protocols → Failure Cases & Recovery

Track overall Phase progress → Phase Checklist

Skill Modification Guidelines

1. Phase consistency: Maintain the gather → build → compose → verify flow
2. Gate format: Follow "Phase N result: {summary}\n Unmet items: {content}\n Please revise or approve to continue." format
3. Template sync: Always keep references/ templates and SKILL.md inline structures in sync
4. Compaction Survival: Place each SKILL.md's core behavioral contract within the first 5,000 tokens. Put supplementary information after — it may be truncated during context compaction without affecting core behavior.

Verification

After modifying a skill, run /context-engineering @<path> in a real project to confirm Phase 1 queries start correctly.