메뉴
Use when preparing context for subagent dispatch or managing token budgets. Triggers: 'prepare context for', 'assemble context', 'token budget', 'context package', 'what context does the subagent need'. Also invoked by develop during planning and execution phases.
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
Use when deciding whether to dispatch subagents, when to stay in main context, or when facing 2+ independent parallel tasks. Triggers: 'should I use a subagent', 'parallelize', 'multiple independent tasks', 'run these at the same time', 'split this up', 'do both at once', 'dispatch template', 'context minimization'.
Use when writing or de-slopping outward-facing prose: newsletters, blog posts, announcements, marketing or landing copy, emails, release notes, doc intros. Triggers: 'make this sound human', 'this reads like AI', 'de-slop this', 'remove the AI tells', 'make it less AI', 'why does this sound like ChatGPT', 'help me draft this piece', 'write a newsletter', 'write a blog post', 'polish this copy', 'tighten this'. Composes with a personal voice skill (for example my-voice-skill) that supplies identity and register; this skill is voice-neutral and supplies the writing process plus the anti-slop catalog. NOT for: code logic or code review (use code-review), project doc structure (use documenting-projects), READMEs (use write-readme / polish-repo), or a specific person's voice (use their voice skill).
Use when building, creating, modifying, or planning any code change. Triggers: "implement X", "build Y", "add feature Z", "create X", "change how X works", "modify Y", "update the Z", "refactor X", "rework Y", "restructure Z", "make X do Y", "let's plan how to", "plan the implementation", "how should we implement", "how would you build", "what's the best way to implement", "I want to...", "We need...", "Would be great to...", "Can we add...", "Let's add...", "Let's build...", "Let's make...", "start a new project". Also for: new projects, repos, templates, greenfield development, refactoring, migrations, multi-file modifications, any code change requiring planning. PREFER THIS OVER plan mode or ad-hoc implementation for ANY substantive code change. NOT for: bug fixes (use debugging), pure research (use deep-research), questions about existing code without intent to change it, or test-only fixes (use fixing-tests).
Algorithmically estimate engineering tickets, JIRA cards, test cards, or feature requests. Produces well-scoped, AI-productivity-adjusted, PERT-buffered estimates with 80/90/95% confidence intervals for single and multi-engineer execution. Triggers - "estimate this ticket", "estimate this card", "estimate this test card", "size this", "story point this", "/estimate", "how long will this take", "T-shirt this", "point this story", "planning poker", "PERT estimate", "give me a timeline", "scope this work".
Use when implementation is complete, tests pass, and you need to decide the integration path. Also use when asked to prepare a branch for release: 'update changelog', 'bump version', 'bump patch version', 'make sure changelog is correct', 'make sure version is correct'. Triggers: 'done with this branch', 'ready to merge', 'ship it', 'wrap this up', 'how should I integrate this', 'what next after implementation'. NOT for: PR creation mechanics (use creating-issues-and-pull-requests) or deciding whether to merge (use finishing-a-development-branch first).
Use when an operator decision belongs on a browser page rather than a terminal prompt. Triggers: "present a decision in the browser", "ask via a canvas page", "interactive decision page", "let me decide in the browser", "render the choice visually". Also invoked by develop when the decision-surface preference is `canvas` (feature-config Question 8). NOT for: quick yes/no with no context (use terminal `AskUserQuestion`), read-only presentation (use `canvas`), or subagent-internal choices with no operator (decide in-context).
| name | assembling-context |
| description | Use when preparing context for subagent dispatch or managing token budgets. Triggers: 'prepare context for', 'assemble context', 'token budget', 'context package', 'what context does the subagent need'. Also invoked by develop during planning and execution phases. |
| Input | Required | Description |
|---|---|---|
purpose | Yes | design, implementation, review, handoff, subagent |
token_budget | Yes | Maximum tokens available |
source_context | Yes | Raw context to select from |
| Output | Description |
|---|---|
context_package | Tiered context ready for injection |
truncation_report | What was excluded and why |
Over budget: remove Tier 3 first, then Tier 2. Never remove Tier 1.
| Tier | Budget | Content | Examples |
|---|---|---|---|
| 1: Essential | 40-60% | Active instructions, user decisions, current artifact, interface contracts, blocking issues | Task spec, APIs, unresolved feedback |
| 2: Supporting | 20-35% | Recent learnings, patterns, prior feedback, success criteria | Last 2-3 iterations, codebase patterns |
| 3: Reference | 10-20% | Historical context, rejected alternatives, verbose docs | Early iterations, full docs (summarize instead) |
| Purpose | Tier 1 Focus | Budget Split | Use With |
|---|---|---|---|
| Design | Requirements, decisions, constraints, integration points | 50/30/20 | design-exploration, writing-plans |
| Implementation | Task spec, acceptance criteria, interfaces, test expectations | 60/25/15 | test-driven-development, executing-plans |
| Review | Code diff, requirements traced, test results | 55/30/15 | code-review, fact-checking |
| Handoff | Current position, pending work, active decisions, blocking issues | 70/20/10 | session boundaries, compaction |
| Subagent | Task, constraints, expected output format | 65/25/10 | dispatching-parallel-agents |
Estimation: tokens ≈ chars / 4 (conservative)
Available: context_window - system_prompt - response_reserve - tool_overhead
Example: 200000 - 8000 - 4000 - 2000 = 186000
Smart Truncation: Never blind head/tail. Preserve structure: intro (30%) + conclusion (20%), mark omitted middle.
| Action | Items |
|---|---|
| Persist | User decisions, validated assumptions, glossary, blocking issues |
| Regenerate | File contents, test results, code patterns (may have changed) |
| Discard | Exploration paths, rejected alternatives, verbose logs |
Handoff format: Position → Pending work → Active decisions → Key learnings → Verification commands
<FINAL_EMPHASIS> Context assembly is invisible infrastructure. Calculate budget. Prioritize by tier. Truncate intelligently. Every token earns its place. </FINAL_EMPHASIS>
