원클릭으로
slow-is-fast
// Research → Plan → Code workflow. Enforces evidence-based reasoning, codebase research, and structured planning before code modifications. Applied automatically for moderate+ tasks.
// Research → Plan → Code workflow. Enforces evidence-based reasoning, codebase research, and structured planning before code modifications. Applied automatically for moderate+ tasks.
| name | slow-is-fast |
| description | Research → Plan → Code workflow. Enforces evidence-based reasoning, codebase research, and structured planning before code modifications. Applied automatically for moderate+ tasks. |
| user-invocable | false |
Core rule: Never modify code based on assumptions alone. Ground every decision in evidence.
| Complexity | Characteristics | Workflow |
|---|---|---|
| Trivial | Typo, formatting, <10 lines, single API | Answer directly, quick grep for context |
| Moderate | Non-trivial logic in single file, local refactoring | Standard: all phases below |
| Complex | Cross-module design, concurrency, new architecture | Deep: use Explore agent, WebSearch, context7 |
Research can be minimal when:
Before any action, apply these internally (do not output unless requested).
Analyze in this order:
When encountering problems:
Execute before entering Plan mode. Summarize findings before proposing any approach.
For moderate+ tasks, use Explore agent to run research in an isolated context.
Understand what already exists. Do not reinvent or contradict existing patterns.
Verify API behavior, library usage, and framework conventions against official sources.
resolve-library-id → query-docs) for libraries/frameworks involvedApply the reuse decision matrix (see references/research-patterns.md).
Once conditions met → enter Code mode directly.
See also: references/research-patterns.md for anti-rationalization table and reuse decision matrix.
Use when the user explicitly asks for a deep local code review, PR review, bug hunt, pre-merge review, or Ultra Review style review with high confidence findings.
Use when the user asks to triage GitHub issues, prioritize backlog work, decide which issues are still valid, or generate delegation-ready issue briefs for coding agents.
Create and run a post-PR review loop for a ready-for-review GitHub pull request. Use when the user asks Codex or Claude Code to watch, loop on, monitor, periodically check, or automatically handle PR review feedback after a PR has been opened. In Codex, the default behavior is to create a 10-minute heartbeat automation first, not to perform an ordinary one-off fix; only execute the review/fix iteration when the skill is invoked by that automation or the user explicitly asks for one immediate pass.
Use when discussing early product ideas, AI or Agent product directions, market observations, MVP scope, wedge markets, non-consensus opportunities, or when the user wants direct red-team critique before architecture.
Route image generation to the local image2-with-codex HTTP service at http://127.0.0.1:4312, which delegates to the user's Codex CLI. Use only when the user explicitly asks for this route. Default behavior is raw prompt pass-through; do not substitute HTML layout, infographic tooling, or other image tools.
Systematic root-cause debugging. Use when: diagnosing bugs, investigating failures, debugging flaky tests, tracing unexpected behavior. Triggers on: '/investigate', 'debug this', 'why is this failing', 'find the root cause', 'flaky test'.