SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsmixseek-prompt-builder

mixseek-prompt-builder

Agents & models

View repository

Updated 10-15 21:13

About this skill

MixSeek Prompt Builder Configuration Generation

Overview

Generate the MixSeek-Core prompt builder configuration file (prompt_builder.toml). Define three templates and control prompt delivery to each component.

Template Roles

Template Target Purpose
team_user_prompt Team (Leader Agent) Per-round instructions, presentation of past history
evaluator_user_prompt Evaluator Instructions for evaluating submitted content
judgment_user_prompt Judgment Client Instructions for determining continuation/termination

Prerequisites

  • The workspace must be initialized (seemixseek-workspace-init)
  • The environment variable MIXSEEK_WORKSPACE must be set (recommended)
  • The orchestrator configuration must exist (when using multi-round functionality)

Generated Files

  • Path: $MIXSEEK_WORKSPACE/configs/prompt_builder.toml

Usage

Step 1: Hearing Requirements

Confirm the following with the user:

  1. Purpose: What kind of task it will be used for (research, analysis, code generation, etc.)
  2. Round Strategy: Whether to improve through multiple rounds or execute once
  3. Custom Judgment Criteria: Whether there are special criteria for termination conditions

Step 2: Template Selection

Select a template according to the purpose:

Template Purpose Features
default General competition format Leaderboard reference, basic round progression
deep-research Deep Research Per-round search keyword improvement, stepwise deepening
single-round Single execution Simple structure, minimal use of round variables

Step 3: Customize

Customize the following as needed:

  • Jinja2 conditional branching: {% if round_number == 1 %} Per-round instructions using etc.
  • Additional variable usage: {{ ranking_table }} Leaderboard display with
  • Adjustment of judgment criteria: Judgment conditions in judgment_user_prompt

Step 4: Saving the File

Save the generated configuration to the following path:

$MIXSEEK_WORKSPACE/configs/prompt_builder.toml

Step 5: Validating the Configuration File

After generation, check the syntax and placeholders:

# TOML構文の検証
uv run python -c "import tomllib; tomllib.load(open('configs/prompt_builder.toml', 'rb'))"

# Jinja2構文の検証（オプション）
uv run python -c "
from jinja2 import Environment
env = Environment()
import tomllib
config = tomllib.load(open('configs/prompt_builder.toml', 'rb'))
for key in ['team_user_prompt', 'evaluator_user_prompt', 'judgment_user_prompt']:
    env.parse(config['prompt_builder'][key])
print('OK')
"

Template Types

default

  • General team competition format
  • With leaderboard reference
  • Basic round progression instructions
  • Purpose: General-purpose tasks

deep-research

  • Deep Research specialized
  • Round 1: Initial investigation, broad search
  • Round 2: Keyword improvement, deepening
  • Round 3: Final integration, verification
  • Purpose: Investigation/research tasks

single-round

  • For single execution
  • Simple prompt structure
  • Minimal use of round variables
  • Purpose: One-off tasks

Available Jinja2 variables

Team / Judgment common variables

Variable name Type Description
user_prompt string Original user prompt
round_number int Current round number (starting from 1)
submission_history string Past Submission history (formatted)
ranking_table string Leaderboard (formatted)
team_position_message string Team ranking message
current_datetime string Current date and time (ISO 8601 format)

Evaluator-only variables

Variable name Type Description
user_prompt string Original user prompt
submission string Agent submission content to be evaluated
current_datetime string Current date and time (ISO 8601 format)

For details, refer to references/JINJA2-VARIABLES.md .

Example

Deep Research configuration example

[prompt_builder]
team_user_prompt = """
現在日時: {{ current_datetime }}

# タスク
{{ user_prompt }}

# ラウンド {{ round_number }}
{% if round_number == 1 %}
## 初期調査
- 広範な検索キーワードで情報収集
- 主要な情報源を特定
{% elif round_number == 2 %}
## 検索キーワード改善
- 前回の結果を分析し、より具体的なキーワードを使用
- 情報の空白を埋める
{% else %}
## 最終統合
- 情報を検証し、矛盾を解消
- 構造化された回答を作成
{% endif %}

# 過去の提出履歴
{{ submission_history }}
"""

evaluator_user_prompt = """
# 評価タスク
以下の提出内容を評価してください。

## ユーザタスク
{{ user_prompt }}

## 提出内容
{{ submission }}
"""

judgment_user_prompt = """
# 判定タスク
調査の継続が必要かどうかを判定してください。

## 判定基準
- 継続: 情報不足、重要な観点が欠落
- 終了: 十分な回答が可能

## ユーザタスク
{{ user_prompt }}

## 提出履歴
{{ submission_history }}

## リーダーボード
{{ ranking_table }}
"""

Custom judgment criteria example

judgment_user_prompt = """
# カスタム判定
以下の基準で継続・終了を判定してください。

## 終了条件（すべて満たす場合）
1. 3つ以上の情報源から確認が取れている
2. 数値データが含まれている
3. 最新情報（1年以内）が含まれている

## 継続条件（いずれかに該当）
1. 情報源が2つ以下
2. 推測や不確かな表現が多い
3. 重要な観点が欠落している

## ユーザタスク
{{ user_prompt }}

## 提出履歴
{{ submission_history }}
"""

Troubleshooting

For Jinja2 syntax errors, variable expansion issues, and multiline TOML strings, refer to the "Error avoidance" section in references/JINJA2-VARIABLES.md .

References

  • TOML schema details: references/TOML-SCHEMA.md
  • Jinja2 variable details: references/JINJA2-VARIABLES.md
  • orchestrator configuration: skills/mixseek-orchestrator-config/SKILL.md
  • evaluator configuration: skills/mixseek-evaluator-config/SKILL.md