Fleshing Out Body Paragraphs and Technical Findings
## Moving from Outline to Evidence
When transitioning from a high-level outline to drafting the actual body paragraphs of a report, writers often encounter a steep drop in AI performance. Instead of the sharp, analytical prose expected in a professional report, large language models (LLMs) like ChatGPT or Claude often default to generic summaries, repetitive transitions, or “fluff.” Worse, if overwhelmed with data, they may hallucinate metrics or lose the logical thread entirely.
The solution is not to write the sections manually, but to change how you prompt. Drafting technical findings requires moving away from single-shot generation and adopting a structured, iterative approach. By acting as an “Information Architect,” you can guide the AI to produce deep, focused, and analytically rigorous body paragraphs.
Strategic Chunking (Context Engineering)
A common mistake analysts make is the “blank page dump”, pasting a massive, unstructured dataset, a dozen interview transcripts, or an entire 50-page reference document into a single prompt and asking the AI to “write the analysis section.” This overloads the model’s context window, leading to lost details and shallow analysis.
Instead, practice Context Engineering through strategic chunking. Feed the AI small, highly relevant “chunks” of data tailored specifically to the single subsection you are drafting.
- Poor approach: Pasting the entire Q3 financial spreadsheet and asking for the “Revenue Analysis” section.
- Chunked approach: Pasting only the Q3 margin variance table and the transcript of the CFO’s comments on supply chain costs, then asking the AI to draft only the subsection on “Margin Compression Drivers.”
By restricting the AI’s universe of information to the exact data needed for one specific finding, you drastically reduce hallucinations and force the model to engage deeply with the details.
The PAIR Framework for Iterative Refinement
To move beyond generic first-draft outputs, professional writers utilize the PAIR (Prompt, Adjust, Iterate, Refine) framework. This multi-stage workflow ensures the AI’s output aligns with your analytical standards.
1. Prompt: Chain-of-Thought for Analytical Rigor
When initiating the first draft of a technical finding, do not just ask the AI to write. Force it to map out its reasoning process first using a technique called Chain-of-Thought (CoT) prompting. Instructing the AI to “analyze the raw data out loud” before drafting prose significantly reduces logical leaps.
Example CoT Prompt:
> “Here is the raw latency data for the Q4 server migration. Before drafting the body paragraphs, think step-by-step. First, list the top three latency spikes and their corresponding timestamps. Second, identify the root causes noted in the data. Once you have mapped this out, draft a two-paragraph technical finding summarizing the migration performance.”
2. Adjust: Constraint-Based Templates
When you receive the first draft, it may sound too conversational or “AI-like.” To adjust the tone, use constraint-based templates.
Crucially, avoid relying on negative constraints (e.g., “Do not use passive voice” or “Don’t write generic fluff”). LLMs often struggle to process negative instructions and may inadvertently focus on the very thing you told them to avoid. Instead, use positive, outcome-driven instructions.
- Instead of: “Don’t be vague and don’t use bullet points.”
- Use: “Write in clear, flowing prose using complete paragraphs. Lead each paragraph with a specific quantitative metric from the data provided, and use active verbs to describe the system’s performance.”
3. Iterate: Few-Shot Prompting for Structural Precision
To ensure the AI’s revised draft matches the exact style of your broader report, use Few-Shot Prompting. This involves providing the LLM with 2 to 3 self-contained, high-quality examples of previously written body paragraphs.
Example Iteration Prompt:
> “The data is correct, but the tone is too academic. I want this to read like a professional consultant’s report. Here are two examples of how I write technical findings: [Insert Example 1] [Insert Example 2]. Rewrite your previous draft to mimic this exact structural pattern, level of detail, and authoritative tone.”
4. Refine: The Human Information Architect
The final step of the PAIR framework happens outside the prompt window. You must act as the editor and validator. Never assume the AI has perfectly synthesized complex data, even with chunking and CoT prompting. Verify every technical claim, double-check that no metrics were extrapolated beyond the provided chunks, and ensure the logical flow connects seamlessly to the preceding and subsequent sections of your report.
Common Pitfalls to Avoid
- Hallucination via Extrapolation: If an AI lacks sufficient data in the chunk you provided, it may invent plausible-sounding metrics to “complete” the narrative. Always include a strict grounding constraint in your initial prompt: “Rely strictly on the provided data. If the data does not explain the cause of a metric change, state explicitly that the root cause is undocumented.”
- Losing the Narrative Thread: Drafting section-by-section can result in a disjointed report where paragraphs don’t logically flow into one another. To prevent this, always include a brief transition context in your chunked prompts: “This section follows our analysis of [Previous Topic]. Draft this new section focusing on [Current Topic], ensuring a smooth transition from the previous finding.”
By combining strategic chunking, Chain-of-Thought reasoning, positive constraints, and few-shot examples, you transform the AI from a generic text generator into a precise analytical drafting tool.