메뉴
Update .context/ (STATE.md, DECISIONS.md, ROADMAP.md) to reflect current project state. Run at natural stopping points, or anytime mid-task.
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
| name | update-context |
| description | Update .context/ (STATE.md, DECISIONS.md, ROADMAP.md) to reflect current project state. Run at natural stopping points, or anytime mid-task. |
| disable-model-invocation | true |
| allowed-tools | Bash, Agent |
<runtime_note>
This skill is explicit-invocation only (/update-context). It never auto-triggers — disable-model-invocation: true in the frontmatter enforces that, matching /init-project: updating .context/ should be a deliberate action the user asks for, not something that fires on an ambiguous cue.
</runtime_note>
This project uses a .context/ directory (created by /init-project) to
carry state across sessions. Delegate the full update to the
context-updater subagent (defined in ~/.claude/agents/) to bring it up
to date now.
if [ ! -f .context/STATE.md ]; then
echo "NOT_INITIALIZED"
fi
If .context/ doesn't exist in this project at all, say so plainly and stop
— don't create it. .context/ is only created by /init-project; this
skill updates it, it doesn't initialize it.
Spawn the context-updater subagent. Do not read git history or the
.context/ files yourself in this conversation — the whole point of
delegating is to keep this session's context window from absorbing the git
log and four files. The subagent reads and writes STATE.md, DECISIONS.md,
and ROADMAP.md (if present) directly.
Report the subagent's summary back to the user as-is. That summary may include a feature-status line (e.g. a feature in flight or recently passed review) alongside the usual STATE.md/DECISIONS.md/ROADMAP.md lines — pass it through unchanged.
Capture implementation decisions for a feature before planning begins — libraries, error handling, edge cases, scope boundary. First step of the full feature loop.
Run the task files produced by /feature-plan, one fresh-context executor subagent per task (or wave), one commit per completed task. Third step of the full feature loop.
Research (conditionally) and decompose an approved feature into per-task files. Second step of the full feature loop. Supports --thorough for a fuller research pass on unfamiliar or high-risk work.
Fast path for a small, well-understood feature or fix. Skips discuss/plan/verify ceremony but still writes one task file and still requires evidence before claiming completion.
Entry point for building a feature. Classifies scope and routes to the fast path or the full discuss/plan/execute/verify loop. Also reports status on an in-flight feature.
Single merged review pass (spec compliance + code quality + goal alignment) plus a hard evidence gate before the feature can be marked done. Fourth step of the full feature loop.