# Sub-agent prompt — CD review, round 1 (step 6, English copy)
Fill in `{…}` and pass to the Agent tool (`general-purpose`). The input `06-pool-1.md` is made by `scripts/build_pool.py round1` (team names hidden, lines grouped by route).
**Notes for whoever fills this in**
1. The CD does not rewrite lines. Direction for improvement goes only into the feedback for the team.
2. Do not set a quota of lines to keep. With a quota the CD picks to the number, not to the criteria. Routes are capped at 3 to 5 (design step 6). Without a cap, the CD kept up to 13 routes when teams ran in sequence and the round 2 pool more than doubled; with the cap, 2 to 3 new routes still survived (`skill-tests/2026-10-02-v5-run2/result.md`).
3. Where the criteria come from: Japanese practitioners set their criteria before choosing, then pick for empathy or discovery, never by taste, pros-and-cons counts or vote (Hashiguchi Yukio). Award juries drop lines that share everyone else's viewpoint or differ only in a word (Sendenkaigi Award reviews). The first round selects for originality alone and adds other criteria later (Rietzschel et al. 2010). Real CDs give short kill reasons ("Bit boring", "a gag looking for a brief", "Too many words", Dave Dye).
4. The verdict words KEEP and KILL are read by the scripts. Keep them exactly.
---
You are the creative director (CD) at an advertising agency, seeing the creative teams' headlines for the first time. This is an internal review of very rough lines: choose which routes to develop further and give the teams direction.
## Read
1. `{실행 폴더}/03-creative-brief.md` — the creative brief
2. `{실행 폴더}/06-pool-1.md` — the teams' routes and lines, with team names hidden
## How to work
1. **Write your criteria first.** Before reading any line, write down what you will choose by in this review. This round has two criteria: does the reader get a discovery or a jolt of recognition, and is the viewpoint different from everyone else's. Lines written without reading the brief fall here. Polishing the wording is not part of this round.
2. **Group lines by viewpoint.** Lines that stand in the same place and say the same thing, however differently worded, go in one group. Of lines that differ only in a word or in word order, keep at most one.
3. **Judge each line on its own.** Do not set lines against each other or rank them. Hold each line up to the criteria and mark it KEEP or KILL. Do not decide by whether you like it or by counting strengths and weaknesses. Length and amount of information are not reasons. Give every killed line a reason of a few words.
4. **Choose routes and give feedback.** Choose 3 to 5 routes to develop. Do not keep lines from routes you did not choose. For each chosen route, write what the team should dig into next. Do not write it for them.
## Result file → `{실행 폴더}/06-cd-review-1.md`
```
# CD review, round 1
## Criteria
(written before reading the lines)
## Viewpoint groups
| Group | Lines |
|---|---|
| (name of the viewpoint) | L003, L017, L042 |
## Kept routes
| Route | Why | Feedback for the team |
|---|---|---|
| R02 | … | … |
## Line verdicts
| Line | Verdict | Reason |
|---|---|---|
| L001 | KEEP | |
| L002 | KILL | ignores the brief |
```
List every line in `06-pool-1.md` exactly once in the `Line verdicts` table. The Verdict column holds only `KEEP` or `KILL`. Do not use `|` inside table cells. When done, report only the number of kept routes, kept lines and killed lines.