Lead
Hi, I’m Pomarano.
This is Part 2 of my AI Agent Study Notes.
Series index Building Your Own AI Agents · Part 1 covered core concepts.
This part uses a sample project — Japanese post copy for X (@syouwanotaisyou) — to design two agents: copywriting and proofreading.
We fix why two agents instead of one, each role, order of handoff, and the evaluation checklist at design time. Build and comparison come in Part 3.
- Japanese version: 文案と校正の2体（第2回）*
Overview of this part
flowchart LR P1["Part 1<br/>Concepts"] D1["Design<br/>copy role"] D2["Design<br/>proofread role"] D3["6-item<br/>checklist"] P3["Part 3<br/>Comparison"] P1 --> D1 --> D2 --> D3 --> P3 classDef meta fill:#eceff1,stroke:#607d8b,stroke-width:2px,color:#1a1a1a classDef concept fill:#e8f4fc,stroke:#3d7ea6,stroke-width:2px,color:#1a1a1a classDef agent fill:#e8f5e9,stroke:#2e7d32,stroke-width:2px,color:#1a1a1a class P1 meta class D1 concept class D2 agent class D3 agent class P3 meta
Why this sample?
| Reason | Explanation |
|---|---|
| Clear rules | 140 chars, two-sentence shape, banned phrases — easy pass/fail |
| Realistic semi-auto | Human posts — same as Semi-Automating X Drafts |
| Existing copy spec | x-shuuchaku-agent-spec.md as source of truth |
| Measurable proofreading | Repeatable failures like character overruns |
Theme: Introspection Tech — short posts pairing Buddhist teaching with a small daily practice.
What goes wrong with only one agent?
With copywriting agent only, these failures show up often:
| Failure | Example |
|---|---|
| Over character limit | Spec says 140 chars; output is 200+ |
| Broken structure | Long paragraphs for teaching + practice — hard to paste on X |
| Banned phrasing | Unsourced “the Buddha said…” style |
| Checklist theater | Human checklist exists but you fix everything by hand every time |
Early drafts hit 248 and 217 characters (scored in Part 3).
That is why we split writing and fixing.
Two-agent design — copy and proofread
3-1. Copywriting agent
| Does | Does not |
|---|---|
| Topic research per weekday rotation | Post to X |
| Write post body | Guarantee 140 chars (proofread’s job) |
Create social/x-drafts/YYYY-MM-DD.md | Rule-check past files |
| Source notes | Set status: posted |
Spec: x-shuuchaku-agent-spec.md
Prompt: automation/x-daily/prompt.md
3-2. Proofreading agent
| Does | Does not |
|---|---|
| Run 6-item checklist | Write copy from scratch |
| Fix violations (max one pass) | Re-research topics |
Append ## 校正結果 / proofread section | X API or auto-post |
Re-count char_count | Edit posted files |
Spec: x-proofread-agent-spec.md
Prompt: automation/x-proofread/prompt.md
3-3. Order of handoff
Copywriting agent → Proofreading agent → Human (review · post)
Serial, not parallel. Proofreading takes the copy agent’s file as input — order is explicit.
Splitting spec and prompt
| File type | Copy | Proofread |
|---|---|---|
| spec | Theme, format, research rules | Checklist, fix priority |
| prompt | “Create one post today per spec” | “Inspect and fix md for date X” |
Change rules only in spec; keep prompts thin — same pattern as Part 1.
The harness article idea of verification as its own layer maps to proofreading as agent two.
Proofreading checklist (Part 3 evaluation)
| # | Item | Pass condition |
|---|---|---|
| 1 | Length | ≤ 140 Japanese characters (newlines, tags count) |
| 2 | Structure | One teaching sentence + one practice sentence |
| 3 | Banned phrasing | No unsourced “the Buddha said…” etc. |
| 4 | Hashtags | 0–1 |
| 5 | URL | No URL in body |
| 6 | Duplication | Not the same concept as any of last 30 drafts |
Part 3 uses first-pass pass rate on these six items as a proxy for “accuracy.”
End-to-end flow
Green = agent · Orange = human
flowchart TB
A1["① Copy agent<br/>research · write"]
A2["② Save draft md"]
A3["③ Proofread agent<br/>6-item check"]
A4["④ Fix · proofread notes"]
H1["⑤ Human review"]
H2{⑥ Edit?}
H3["⑦ Post to X"]
A1 --> A2
A2 --> A3
A3 --> A4
A4 --> H1
H1 --> H2
H2 -->|yes| H3
H2 -->|no| H3
classDef agent fill:#e8f5e9,stroke:#2e7d32,stroke-width:2px,color:#1a1a1a
classDef human fill:#fff3e0,stroke:#e65100,stroke-width:2px,color:#1a1a1a
class A1 agent
class A2 agent
class A3 agent
class A4 agent
class H1 human
class H2 human
class H3 human
With GitHub Actions each morning, a step runs before ① (operations article). Agent roles stay the same.
Summary
- Sample: Japanese X copy — clear rules, easy to measure proofreading
- Copy creates; proofread fixes — separate responsibilities
- Evaluate with a 6-item checklist
- Part 3 compares without proofreading (A) vs with (B)
This adds two-agent roles to the “rules layer” from Part 1.
コメント