Principle:Spcl Graph of thoughts Keyword Counting Prompt Design

Knowledge Sources	Graph of Thoughts Graph of Thoughts: Solving Elaborate Problems with Large Language Models
Domains	Prompt_Engineering, Keyword_Counting
Related Implementations	Implementation:Spcl_Graph_of_thoughts_KeywordCountingPrompter
Last Updated	2026-02-14

Overview

Domain-specific prompt engineering pattern for decompose-count-aggregate-validate reasoning over keyword frequency counting problems.

Description

The Keyword Counting Prompt Design principle defines a comprehensive family of prompt templates for counting the frequency of country names in text passages. This is the most elaborate example in the GoT framework, supporting seven distinct reasoning approaches: IO, CoT, ToT, ToT2, GoT4, GoT8, and GoTx. Each approach uses a different combination of the prompt templates to achieve its goals.

Prompt Strategy by Reasoning Approach

Input-Output (IO): A single prompt (count_prompt) asks the LLM to count all country name frequencies in the full passage and output a JSON dictionary. Few-shot examples demonstrate the expected format.

Chain-of-Thought (CoT): The prompt (count_prompt_cot) instructs the LLM to decompose the passage into paragraphs, count frequencies per paragraph, and combine them -- all within one response. The <Approach> section provides explicit steps.

Tree-of-Thought (ToT/ToT2): Uses the CoT prompt for the initial generation, then applies tot_improve_prompt to iteratively refine incorrect frequency dictionaries. The improvement prompt presents both the original text and an incorrect dictionary, asking the LLM to compare and fix frequencies.

GoT4 (4-paragraph split): Three-phase approach:

1. Split: got_split_prompt asks the LLM to split the text into 4 equal-length paragraphs as JSON.
2. Count: count_prompt_cot is applied to each paragraph independently.
3. Aggregate: got_aggregate_prompt merges two frequency dictionaries by summing counts.
4. Validate & Improve: got_improve_aggregate_prompt corrects incorrect aggregations by showing both input dictionaries and the erroneous combined result.

GoT8 (8-paragraph split): Same pattern as GoT4 but uses got_split_prompt2 to produce 8 paragraphs, resulting in a deeper aggregation tree (3 levels of pairwise merges).

GoTx (per-sentence split): The most fine-grained approach. Uses got_split_prompt3 to split text into individual sentences, then applies count_prompt_sentence (a simpler prompt optimized for single sentences with a word-by-word iteration approach) to each. Improvement uses sentence_improve_prompt tailored for short text fragments.

Key Design Patterns

• JSON output format: All counting prompts require output as a JSON dictionary (Template:"country": frequency, ...). Double braces {{ }} are used in templates to escape Python format strings.
• Graduated complexity: Few-shot examples increase from simple (1-2 countries) to complex (12 countries, repeated occurrences) to calibrate LLM behavior.
• Improvement prompts: Both tot_improve_prompt and sentence_improve_prompt include a "Reason" field in examples that demonstrates the explicit error analysis the LLM should perform.
• Aggregation prompt: got_aggregate_prompt provides step-by-step instructions for dictionary merging, making the addition logic explicit.
• Validation-aware aggregation: got_improve_aggregate_prompt shows the LLM both input dictionaries and the incorrect merge, asking it to verify each country's sum and fix errors.

Prompt Template Inventory

Template	Lines	Purpose	Target Method
count_prompt	165-219	Full-text frequency counting	IO
count_prompt_cot	221-349	CoT-style paragraph decomposition	CoT, ToT (initial), GoT4/GoT8 (phase 1)
count_prompt_sentence	351-408	Sentence-level word-by-word counting	GoTx (phase 1)
tot_improve_prompt	410-503	Fix incorrect frequency dictionary	ToT, GoT4/GoT8 (phase 2)
sentence_improve_prompt	505-574	Fix incorrect dictionary for single sentence	GoTx (phase 2)
got_split_prompt	576-599	Split text into 4 paragraphs (JSON)	GoT4 (phase 0)
got_split_prompt2	601-633	Split text into 8 paragraphs (JSON)	GoT8 (phase 0)
got_split_prompt3	635-659	Split text into individual sentences (JSON)	GoTx (phase 0)
got_aggregate_prompt	661-678	Merge two frequency dictionaries	GoT4/GoT8/GoTx (aggregation)
got_improve_aggregate_prompt	680-729	Fix incorrect dictionary merge	GoT4/GoT8/GoTx (validation)

Related Pages

• Implementation:Spcl_Graph_of_thoughts_KeywordCountingPrompter -- Concrete Python class implementing this principle
• Principle:Spcl_Graph_of_thoughts_Keyword_Counting_Response_Parsing -- Companion parsing principle
• Workflow:Spcl_Graph_of_thoughts_GoT_Keyword_Counting_Pipeline -- End-to-end workflow using these prompts