Prompt Engineering for AI News Editorials: 6 Iterations from Template to Data-Driven

Background

In late October 2026, we began building an automated AI daily news editorial pipeline. The goal was clear: let AI automatically search, filter, draft, review, and publish 3-5 high-information-density AI news stories every day.

Reality quickly delivered a lesson — the quality of AI-generated content was far below expectations.

First Attempt: Template-Based Output

The initial WRITER prompt looked like this:

Please generate an AI daily news article based on the provided search results.
Requirements:
1. Include facts and general advice
2. Distinguish between verified and speculative content
3. Output 3-5 news items

Problems:

• AI output had highly template-like structure: “Editor judgment/suggestions” appeared after every news item
• Lacked specific data: no revenue figures, user counts, version numbers
• Generic advice dominated: “recommend paying attention to”, “worth watching” filled 40% of the content

Second Iteration: Deleting Template Instructions

We removed template instructions like “Separate reported facts…” and replaced them with:

Weave analysis directly into each section:
1. Each news item must include specific data (revenue, user count, date, version)
2. Cover multiple dimensions: AI products, market trends, policy/regulation
3. Each section contains facts/impact/industry background comparison

Result: Content depth improved, but new problems emerged — AI started fabricating dates and inferring unverified cause-and-effect relationships.

Third Iteration: Date Verification

The model began producing content like:

“On Thursday 11/5, a company released a new model…”

The problem: “Thursday 11/5” was inferred by the model from context, not from raw data. We added strict date verification rules:

No relative date inference:
- If the original data doesn't contain an explicit date, do not infer
- Only use absolute dates that appear in the original data (e.g., "2026-11-05")
- All dates must be traceable to specific sources

Effect: Date hallucinations dropped by 80%, but the review stage (REVIEWER) started frequently rejecting drafts.

Fourth Iteration: Independent Review Mechanism

We introduced an independent REVIEWER model (using gpt-6-luna) with these review rules:

Independent review rules:
1. Every claim must be supported by the original source
2. Reject unverified inferences (e.g., "Democrats control both houses")
3. Reject generic purchasing advice (lacking specific data)
4. Accept single well-supported event reports

Key finding: The REVIEWER rejection rate hit 60%. Primary reasons:

1. Market event date reference ambiguity
2. Political inferences not strictly verified by source text
3. Generic advice lacking specific data support

Fifth Iteration: Selector Layer Degradation

We realized the problem wasn’t in the WRITER — it was in the SELECTOR layer being too conservative. The original prompt required:

“Provide 3-5 news items from different industry angles”

But in reality, some days might only have 1-2 high-quality news items. We changed it to:

Minimum selection: 1 item
Diversity checks moved to the review stage
If evidence doesn't support multiple industry angles, accept single-event reports

Effect: Pipeline throughput increased by 40%, review rejection rate dropped to 15%.

Sixth Iteration: Fuzzy Quote Verification

This was the most technical iteration. Quote verification is the core mechanism for ensuring content authenticity, but we encountered a problem:

AI-generated quotes are truncated, while the original text is complete.

For example, the original text says:

“The Wall Street Journal reported that Nvidia fell 9.3% to $181.60…”

The AI-generated quote is:

“Nvidia fell 9.3%…”

Standard string matching failed. We implemented the _fuzzy_match() function:

def _fuzzy_match(quoted, original):
    # 1. Significant word 60% overlap threshold
    # 2. 6-character prefix matching
    # 3. Truncated word tolerance
    # 4. Quote verification logic tightening
    ...

Effect: Quote verification pass rate increased from 30% to 90%.

Core Lessons

1. Prompts are iterated, not designed

Each iteration solved the previous iteration’s problem, but there’s no perfect prompt. The key is:

• Keep failure records: Every rejection has a reason field
• Minimal changes: Change only one variable per iteration
• Data-driven: Use pass rate, rejection rate, CTR metrics to measure progress

2. Independent review is necessary

WRITER and REVIEWER use different models (Astra + Luna) with different judgment criteria. Cross-validation catches 60% of low-quality outputs.

3. Don’t over-constrain AI

The original prompts had too many requirements (3-5 items, multiple industries, data support), causing AI to either fabricate or skip. After degrading to a 1-item minimum, quality actually improved.

4. Technical verification is more effective than language constraints

Telling the prompt “don’t fabricate” doesn’t work, but technical verification like _fuzzy_match() can 100% intercept fabricated quotes.

Future Directions

1. Automated date verification: Currently relies on manual rules, will integrate calendar verification
2. Dynamic review standards: Adjust rejection thresholds based on historical pass rates
3. Bilingual consistency checks: Ensure Chinese and English versions convey the same facts and limitations

AI-assisted drafting with a separate model review. Data from actual production pipeline run records.