Daily Forge

AI-native engineer daily loop

๋Ÿฌ๋‹๋จธ์‹  ์œ„์—์„œ ๊ทธ๋Œ€๋กœ ๋”ฐ๋ผ๊ฐ€๋Š” 3์‹œ๊ฐ„ ๋ฃจํ‹ด์ด๋‹ค. ์ˆœ์„œ๋Š” ๊ณ ์ •ํ•œ๋‹ค: ์ž… ํ’€๊ธฐ, ๊ธฐ๋ณธ ๊ฐœ๋…, ์‹œ์Šคํ…œ๋””์ž์ธ, ์ฝ”๋“œ/LLD, AI agent, ์˜์–ด ์ถœ๋ ฅ, ์ŠคํŒŒ๋ง, ํšŒ๊ณ . ๋ชจ๋ฅด๋ฉด ์นด๋“œ๋กœ ๋“ค์–ด๊ฐ€๊ณ , ์•Œ๋ฉด ๋ง๋กœ ๋ฐฉ์–ดํ•œ๋‹ค.

Rule for the day

ํŽ˜์ด์ง€๋ฅผ ๋งŽ์ด ์—ฌ๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ๊ฐ€ ์•„๋‹ˆ๋‹ค. ๊ฐ ๋ธ”๋ก๋งˆ๋‹ค ํ•˜๋‚˜๋ฅผ ๋ณด๊ณ , ํ•˜๋‚˜๋ฅผ ๋งํ•˜๊ณ , ํ•˜๋‚˜๋ฅผ ๊ณต๊ฒฉ๋ฐ›๋Š”๋‹ค.

๋ชจ๋ฅด๋ฉด ๊ฒ€์ƒ‰ํ•˜์ง€ ๋ง๊ณ  ๋จผ์ € ์—ฐ๊ฒฐ๋œ ์นด๋“œ๋กœ ๋“ค์–ด๊ฐ„๋‹ค. ์นด๋“œ์—์„œ raw English๋ฅผ ๋งํ•˜๊ณ , ๋‹ค์‹œ ์ด ํŽ˜์ด์ง€๋กœ ๋Œ์•„์˜จ๋‹ค.

Today's weak axis

์˜ค๋Š˜ ๋ฌด๋„ˆ์ง€๊ธฐ ์‰ฌ์šด ์ถ•์„ ํ•˜๋‚˜ ๊ณ ๋ฅธ๋‹ค. ๊ณ ๋ฅธ ์ถ•์ด ์ŠคํŒŒ๋ง์˜ ๊ณต๊ฒฉ ๋ฐฉํ–ฅ์ด๋‹ค.

Start sparring
1

Boot the brain

8 minutes

์˜ค๋Š˜์˜ ์‚ฌ๊ณ ๋ฐฉ์‹๊ณผ ์ž…์„ ๋จผ์ € ์ผ ๋‹ค.

์ฒ˜์Œ 8๋ถ„์€ ๊ณต๋ถ€๊ฐ€ ์•„๋‹ˆ๋ผ ์ƒํƒœ ์ „ํ™˜์ด๋‹ค. ๊ฑท๊ธฐ ์‹œ์ž‘ํ•˜์ž๋งˆ์ž ์˜์–ด ๋ฌธ์žฅ๊ณผ ํŒ๋‹จ ํ”„๋ ˆ์ž„์„ ๊บผ๋‚ด์•ผ ๋’ค์—์„œ ๋ฉˆ์ถ”์ง€ ์•Š๋Š”๋‹ค.

Do this
  1. Read one move out loud.
  2. Say three natural phrases without translating.
  3. Pick one weak axis for today: system, code, AI, English, or behavior.
Say out loud

Today I am not memorizing terms. I am going to explain the constraint, prove the behavior, and defend the trade-off.

2

Foundation concepts

22 minutes

์šฉ์–ด๋ฅผ ์™ธ์šฐ์ง€ ๋ง๊ณ , ์นด๋“œ ํ•˜๋‚˜๋ฅผ ๋™๋ฃŒ์—๊ฒŒ ์„ค๋ช…ํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋งŒ๋“ ๋‹ค.

AI ์‹œ๋Œ€์—๋„ ๋ฌด๋„ˆ์ง€๋Š” ์‚ฌ๋žŒ์€ ๋ณดํ†ต ๊ธฐ๋ณธ ๋‹จ์–ด๋ฅผ ๋ชป ์žก์•„์„œ ๋ฌด๋„ˆ์ง„๋‹ค. API, runtime, thread, transaction, p99 ๊ฐ™์€ ๋‹จ์–ด๋ฅผ ์ฆ‰์‹œ ์„ค๋ช…ํ•ด์•ผ ํ•œ๋‹ค.

Do this
  1. Open /concepts in path order.
  2. Pick 3 cards only: one known, one shaky, one unknown.
  3. For each card: Korean meaning -> raw English -> failure mode -> self-attack answer.
Say out loud

The key concept is not the term itself; it is what breaks when I misunderstand it.

3

System design with failure

35 minutes

์„œ๋น„์Šค ํ•˜๋‚˜๋ฅผ ์š”๊ตฌ์‚ฌํ•ญ, ๋ณ‘๋ชฉ, ์žฅ์• , ์ง€ํ‘œ, ๋กค์•„์›ƒ๊นŒ์ง€ ๋งํ•œ๋‹ค.

์‹œ์Šคํ…œ๋””์ž์ธ์€ ์˜ˆ์œ ๋ฐ•์Šค ๊ทธ๋ฆฌ๊ธฐ๊ฐ€ ์•„๋‹ˆ๋‹ค. p99, retry, idempotency, backpressure, cache staleness, DLQ, rollback์„ ์‹ค์ œ ์งˆ๋ฌธ์œผ๋กœ ๊ฒ€์ฆํ•ด์•ผ ํ•œ๋‹ค.

Do this
  1. Start from /sdcamp layer order or one /sd walkthrough.
  2. Ask the universal service checks: usability, security, retry, idempotency, observability, rollout.
  3. Name one metric that proves the fix worked.
Say out loud

I will start from the user-visible path, then identify the invariant, then attack the failure modes.

4

Code, LLD, and bug sense

30 minutes

์ฝ”๋“œ๋ฅผ ์ฝ๊ณ , ์ž‘์€ ์„ค๊ณ„๋ฅผ ๋งŒ๋“ค๊ณ , AI๊ฐ€ ๋งŒ๋“  ๋ฒ„๊ทธ๋ฅผ ์žก๋Š”๋‹ค.

๋ฐ”์ด๋ธŒ์ฝ”๋”ฉ์œผ๋กœ ๋นจ๋ผ์ ธ๋„ ์ฝ”๋“œ ์†Œ์œ ๊ถŒ์€ ์‚ฌ๋ผ์ง€์ง€ ์•Š๋Š”๋‹ค. ์ž๋ฃŒ๊ตฌ์กฐ, ๊ฐ์ฒด ๊ฒฝ๊ณ„, ํ…Œ์ŠคํŠธ, ์˜ˆ์™ธ, ๋™์‹œ์„ฑ ๋ฒ„๊ทธ๋ฅผ ์ง์ ‘ ์„ค๋ช…ํ•ด์•ผ ํ•œ๋‹ค.

Do this
  1. Do one AI-off coding or Java kata.
  2. Then open one LLD or bug card.
  3. Explain the invariant, edge case, and test before reading the answer.
Say out loud

Before I write code, I want to pin down the invariant and the edge cases that would break it.

5

AI-native engineering

25 minutes

LLM/agent/MCP/RAG๋ฅผ ์ œํ’ˆ๊ณผ ์šด์˜ ๊ด€์ ์œผ๋กœ ์„ค๋ช…ํ•œ๋‹ค.

AI๋ฅผ ๊ทธ๋ƒฅ ์“ฐ๋Š” ์‚ฌ๋žŒ๊ณผ AI-native engineer์˜ ์ฐจ์ด๋Š” ๊ฒ€์ฆ์ด๋‹ค. token, context, tool call, trace, eval, permission, run ledger๋ฅผ ๊ตฌ์กฐ๋กœ ์„ค๋ช…ํ•ด์•ผ ํ•œ๋‹ค.

Do this
  1. Open /ai in path order.
  2. Pick 2 fundamentals and 1 runtime/safety card.
  3. Use /aiaudit once: let AI propose, then attack its assumptions.
Say out loud

A model output is not a result. It becomes a result only after grounding, tool validation, trace, and verification.

6

English output

25 minutes

๋‚ด ์ƒ๊ฐ, ์—…๋ฌด์ƒํ™ฉ, ๋ฉด์ ‘ ๋‹ต๋ณ€์„ ์ž์—ฐ ์˜์–ด๋กœ ๊บผ๋‚ธ๋‹ค.

๊ธฐ์ˆ ์„ ์•Œ์•„๋„ ๋ง์ด ์•ˆ ๋‚˜์˜ค๋ฉด ๋ฉด์ ‘๊ณผ ํšŒ์‚ฌ ๋Œ€ํ™”์—์„œ ๋ชป ์“ด๋‹ค. ์‰ฌ์šด ํ‘œํ˜„, collocation, pushback, clarification์„ ๋งค์ผ ์ž…์— ๋ถ™์—ฌ์•ผ ํ•œ๋‹ค.

Do this
  1. Use /myvoice for one raw Korean thought.
  2. Drill 5 phrase cards: everyday, connector, technical, pushback, uncertainty.
  3. Do one listen or talk round if the mouth is warm.
Say out loud

Let me phrase this clearly. The point I am trying to make is that the risk is not the feature; it is the failure mode.

7

Sparring and defense

25 minutes

์ฝ์€ ๋‚ด์šฉ์„ ์‹ค์ œ ์••๋ฐ• ์งˆ๋ฌธ์œผ๋กœ ๋ฐฉ์–ดํ•œ๋‹ค.

์นด๋“œ๋ฅผ ๋ณธ ๊ฒƒ์€ ์ž…๋ ฅ์ผ ๋ฟ์ด๋‹ค. ๋ฉด์ ‘์€ ๊ณต๊ฒฉ๋ฐ›๊ณ ๋„ ๊ตฌ์กฐ๋ฅผ ์œ ์ง€ํ•˜๋Š” ์ถœ๋ ฅ ๊ฒŒ์ž„์ด๋‹ค. ์—ฌ๊ธฐ์„œ ํ‹€๋ฆฐ ๋ถ€๋ถ„์ด ๋‚ด์ผ์˜ ํ•™์Šต ์ˆœ์„œ๋ฅผ ์ •ํ•œ๋‹ค.

Do this
  1. Start one sparring round tied to today's weak axis.
  2. Do not answer vaguely: state assumption, constraint, trade-off, metric.
  3. After feedback, say the corrected answer once more.
Say out loud

My assumption is this. If that assumption changes, the design changes here. The metric I would watch is this.

8

Close the loop

10 minutes

์˜ค๋Š˜์˜ ๋นˆํ‹ˆ ํ•˜๋‚˜์™€ ๋‚ด์ผ์˜ ์ฒซ ํด๋ฆญ์„ ๋‚จ๊ธด๋‹ค.

ํ•˜๋ฃจ์ข…์ผ ๋งŽ์ด ๋ณธ ๊ฒƒ๋ณด๋‹ค ์ค‘์š”ํ•œ ๊ฑด ๋‹ค์Œ ๋ฃจํ”„์˜ ์‹œ์ž‘์ ์„ ์žƒ์ง€ ์•Š๋Š” ๊ฒƒ์ด๋‹ค. ์•ฝ์  ํ•˜๋‚˜๋ฅผ ๋ช…๋ช…ํ•ด์•ผ ๋‹ค์Œ๋‚  ํ—ค๋งค์ง€ ์•Š๋Š”๋‹ค.

Do this
  1. Write one sentence: what I could not explain.
  2. Choose tomorrow's first page.
  3. Mark whether the gap was concept, code, AI, English, or pressure.
Say out loud

The gap I found today is specific: I could not clearly explain this failure mode under pressure.

If you only have 60 minutes

์ค„์—ฌ์•ผ ํ•˜๋ฉด ์ˆœ์„œ๋ฅผ ๋ฐ”๊พธ์ง€ ๋ง๊ณ  ์••์ถ•ํ•œ๋‹ค: Boot 5m, Concepts 10m, System 15m, Code 10m, AI 10m, Sparring 10m.

If you walk all day

3์‹œ๊ฐ„ ๋ฃจํ”„๋ฅผ ๋ฐ˜๋ณตํ•œ๋‹ค. ๋‘ ๋ฒˆ์งธ ๋ฃจํ”„๋ถ€ํ„ฐ๋Š” ์ƒˆ ์นด๋“œ๋ฅผ ๋Š˜๋ฆฌ์ง€ ๋ง๊ณ , ์ฒซ ๋ฃจํ”„์—์„œ ํ‹€๋ฆฐ ์นด๋“œ์™€ ์ŠคํŒŒ๋ง ๋‹ต๋ณ€๋งŒ ๋‹ค์‹œ ์นœ๋‹ค.

Close-out note