How a Research Report Gets Made: My Tongling Workflow

Start: I got chapters 4-6 of the report

After the team divided the research report, I was responsible for chapters 4-6 (problem diagnosis, case study, implementation path). What I had was a pile of scattered material — interview records, data manuals, strategy docs, plus plenty of duplicates and conflicts.

My instinct: don't stitch by feel, follow a process.

Step 1: from a sea of documents to dedup

15 docx files, 26 files total; one enterprise might have 4 versions. Reading all would drive anyone mad. First I clarified duplicates and chronology:

• Liuguo Chemical, 4 files → keep the July 3 "comprehensive version".
• Tianjinghu Community, 4 files → by date: comprehensive > record > manual > reference card.

Dedup isn't deleting files; it's deciding which is authoritative — same as choosing the original voucher in accounting: with multiple documents, the latest and fullest wins.

Step 2: parallel research, not sequential reading

7 material groups + 5 directions — too slow one by one. I ran 16 parallel agents:

• 7 agents read material groups, extracting pain points (keeping quotes), interview points, data facts, policy platforms, existing proposals.
• 5 agents researched 4 comparison cities (Xuzhou, Ordos, Panzhihua, Datong) plus Tongling's public data.

What took a day or two by hand took hours to draft three chapters. At scale, "reading" stops being the bottleneck — "judgment" is.

Step 3: verifying data — accounting-level rigor

Many figures conflicted: coverage 95.2% or 97.1%? Output 17.8B or 178B? Population 1.3M or 1.7M?

I turned every conflict into a 28-item checklist, priority-sorted (hard data / medium / typos / to-fill / material gaps), each with "conflict + source + conclusion space".

For an accounting student this is natural — it's just reconciliation. Build the ledger (checklist), reconcile each item, mark "unverified" without faking completion.

Step 4: AI data-filling

Remaining 【to fill】 placeholders became 22 prompts for AI scraping, sorted into A (findable), B (reference only), C (not scrapable — needs interviews).

I also tested my own scraper: for government/news pages, readability extraction works best. AI data-filling isn't "ask it for an answer" — it's "send it for evidence."

The key moment: catching a factual error

The draft claimed central transfer payments "drop to zero by 2028" — wrong. Official policy is "9-year support window, tapering to 40%, never zero". One fix, two chapters corrected.

In research, the most expensive thing isn't writing — it's not writing it wrong.

Completion review: honestly facing the gap

Final assessment: ~85% (deliverable draft, not final). More importantly, I named the real weakness:

Chapter 4 is "diagnosis based on government interviews", but some subjects only have interview outlines, not interview transcripts — a risk of "treating the questions as the conclusions".

The most honest and valuable line: knowing where you're short beats pretending you're done. I listed A/B/C todos and sent them to the team.

One reflection

Research rigor = accounting's reconciliation mindset + AI's scale.

• Reconciliation: authoritative sources, item-by-item verification, mark gaps without faking.
• Scale: 16 parallel readers, 22 fill prompts, AI finds evidence rather than fabricating answers.