메뉴
Use when building software with AI agents and you need a serious end-to-end workflow instead of vibe coding. Covers harness choice, context discipline, cleanup, review loops, launch pressure, and security basics.
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 | agentic-engineering-workflow |
| description | Use when building software with AI agents and you need a serious end-to-end workflow instead of vibe coding. Covers harness choice, context discipline, cleanup, review loops, launch pressure, and security basics. |
| version | 1.0.0 |
| author | David Ondrej / Michael Shimeles interview notes |
| license | MIT |
| metadata | {"hermes":{"tags":["agentic-engineering","ai-coding","workflow","micky-podcast"],"related_skills":["source-code-context","code-structure-cleanup","grep-loop-review-workflow"]}} |
Use this as the high-level operating system for building with AI agents. The core idea is simple: stay in charge, keep the agent's context focused, and give it tight feedback loops.
This is not "ask the AI to build everything and hope." It is a workflow where the human decides the outcome, the agent does the mechanical work, and tests/reviews keep the result honest.
Do not use this for one-off tiny edits where a normal direct prompt is enough.
Pick the strongest harness/model you can access. The harness is the wrapper around the model: file search, terminal, browser, tools, system prompt, and project memory. The model matters, but the harness determines what the model can actually do.
Keep the task small. Ask for one feature, one fix, or one reviewable unit at a time. If a plan is too large, ask the agent to split it into smaller PR-sized chunks.
Give source code as context when docs are not enough. If you are using a package, SDK, framework, or open-source tool, put its source in a reference folder and tell the agent to search it before coding.
Build the minimal feature first. Do not refactor the whole app while building the feature. Get the smallest working version running.
Run a cleanup pass. After the feature works, ask the agent to find duplicated runtime mechanics and move them into reusable service-layer modules.
Run a review-fix loop. Use tests, typechecks, and AI/human review. Feed review feedback back into the coding agent. Keep fixing until the PR is clean or a human decision is needed.
Launch earlier than feels comfortable. Do not hide forever behind "one more feature." A semi-functional MVP with feedback beats a perfect private project.
Apply security guardrails. Use 2FA, a password manager, avoid young packages, and ask your agent to check whether your project is exposed when a package/security issue trends.
We are going to build this using an agentic engineering workflow.
Rules:
1. Keep the change small and reviewable.
2. Search the existing code before creating new abstractions.
3. If using a package/framework, reference its local source or official repo before guessing APIs.
4. Build the minimal working version first.
5. After it works, run a code-structure cleanup pass.
6. Run relevant tests/typechecks.
7. Summarize what changed, what was tested, and what still needs human judgment.
Task:
<describe the feature or fix here>
Use after an AI-built feature works but the code has duplicated mechanics, repeated API calls, or messy structure. Guides a cleanup pass that extracts reusable service-layer modules without changing behavior.
Use when you have a small PR or feature and want an agent to repeatedly fix review feedback until tests pass and the PR is merge-ready. Works with AI reviewers, Greptile-style review, or human feedback.
Use when an AI coding agent is guessing API names or framework behavior from incomplete docs. Shows how to provide local source-code references so the agent searches the real implementation before coding.