원클릭으로
context-engineering
Context window management techniques for maintaining efficiency and preventing context bloat
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
메뉴
Context window management techniques for maintaining efficiency and preventing context bloat
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Interactively elicit, capture, and maintain software/system requirements as a living, traceable repository. Use whenever the user wants to gather, write, refine, or organize requirements — including producing or updating an SRS (software/system requirements specification), writing ADRs / architecture decision records, building a requirements traceability graph (DAG), or keeping a decision ledger. Trigger proactively when the working folder already holds requirements artifacts (a `requirements/` folder of YAML, `decisions/` ADRs, an `srs.md`, or `ledger.md`), or when the user shares documents that are clearly specs, RFCs, or feature briefs and wants them structured. Also use for phrases like "let's spec this out", "interview me about what we're building", "capture these requirements", "write a decision record", or "turn these notes into a spec". Not for project management, code review, application code, dependency manifests like requirements.txt, or hardware "system requirements" (RAM/CPU).
Use this skill whenever the user wants to create a bash script that drives an autonomous coding loop — scripts that run Claude Code on a schedule (cron, GitHub Actions, systemd timer) or on-demand to make repository improvements, execute plans, triage issues, or otherwise do agentic work with minimal human intervention. Triggers include phrases like "scout script", "autonomous script", "agent harness", "script that runs on a schedule", "script that implements a plan", ".scripts/*.sh", or any request for a bash script whose body is "run Claude Code with a prompt and commit the results". Always use this skill before writing such a script — it encodes the interview flow, the adversarial-reviewer pattern, and the locking/ledger conventions.
Task structure and atomic commit patterns for granular, verifiable work units
Conflict identification and resolution patterns for requirements, decisions, and plans
Exploration map management for tracking discussed areas and uncharted territory during DISCUSS phase
Phase coordination, agent handoffs, and workflow state machine management
| name | context-engineering |
| description | Context window management techniques for maintaining efficiency and preventing context bloat |
| triggers | ["context management","fresh context","context window","context bloat"] |
This skill provides techniques for managing context effectively throughout the workflow to prevent bloat and maintain efficiency.
| Role | Context Budget | Reason |
|---|---|---|
| Orchestrator | <40% | Needs room for coordination |
| Interviewer | <70% | Conversation history accumulates |
| Planner | <60% | Needs to see requirements + codebase patterns |
| Executor | <80% | Can focus on single task deeply |
| Verifier | <60% | Needs to cross-reference requirements and code |
Instead of loading full files:
BAD: Read entire codebase
GOOD: Read specific files needed for current task
BEST: Read summaries, then specific sections
Load context in layers:
Load more context only when needed:
1. Start with minimal context
2. Identify what's missing
3. Load specific missing pieces
4. Repeat as needed
Spawn a fresh agent when:
When spawning fresh agent:
Include in handoff:
- Task ID and definition
- Key decisions from ITEM-XXX.md
- Relevant file paths
- Success criteria
Exclude:
- Full conversation history
- Files not related to task
- Exploration paths not taken
Natural points to reset context:
Watch for:
Periodically verify:
Use state files as external memory:
Task: Understand how authentication works in this codebase
Instead of:
- Loading all auth-related files into orchestrator context
Do:
- Spawn researcher agent with specific question
- Receive summary findings
- Keep summary in orchestrator context
Task: Implement TASK-005
Instead of:
- Loading full codebase patterns and task into orchestrator
Do:
- Spawn executor with task definition + relevant files only
- Receive completion status
- Update state files
After deep work, compress to summary:
Before compression: [Full exploration transcript]
After compression:
**Exploration Result**: Authentication
- Uses JWT tokens
- Stored in localStorage
- Refresh flow in AuthService.refresh()
- Decision needed: token expiry duration
BAD: Include full task XML in every message
GOOD: "Executing TASK-005 (see ITEM-XXX.md Tasks section)"
At checkpoints, snapshot key state:
### CHECKPOINT: EXECUTE-001
**Context snapshot**:
- Phase: EXECUTE
- Progress: 5/10 tasks
- Current: TASK-006
- Blocked: None
- Key decisions: Use JWT (DECISION-003)