Summary of "ChatGPT untuk Penulisan Artikel Ilmiah: Demo #1"

Main idea

The video demonstrates how to use ChatGPT (GPT-3.5 and GPT-4) to assist in writing scientific articles—particularly producing abstracts and converting abstract sentences into full introduction paragraphs—while emphasizing the need for human verification, paraphrasing, and synthesis. The presenter shows a step-by-step prompt design and workflow for turning literature points into structured text, how to include test results and citations, and common prompt constraints to get useful outputs. Outputs from different model versions are compared and pitfalls (e.g., hallucinations, inappropriate examples) are highlighted, with guidance on how to correct them.

Key concepts and lessons

Use structured prompts:
- Define a clear role (e.g., “supervisor/mentor”).
- Supply an outline and explicit constraints (word count, sentence style, forbidden content).
Repeated outline structure:
- General background → specific background → knowledge gap → results → implication.
One-sentence-per-point rule:
- Each sentence in the abstract should represent one paragraph in the introduction; expand each sentence into a paragraph supported by literature.
Provide experimental results up front if they must appear in the abstract (example metrics: precision 0.896, success rate 0.755).
Include examples and limitations (e.g., low frame rate, low processing power, low-cost cameras) but verify they fit the research context—manual edits are often needed (for example, replace “smartphone camera” with “CCTV” if more appropriate).
Maintain traceability with a lookup table that maps sentences/points to verified papers so generated sentences can be linked to specific references and validated.
Citation options:
- Ask ChatGPT to place citation placeholders (numbers or short citations), or
- Feed specific papers/IDs so the model can insert references—then manually verify.
Keep generation focused:
- Request active sentences, avoid long compound sentences, and set a single-paragraph abstract with a specific word limit (e.g., 200 words) if needed.
Be cautious with plugins and other AI tools: they change the generation paradigm and may increase the risk of hallucination; do not over-rely on them.
Always paraphrase, synthesize, and validate model outputs against the literature and data—treat ChatGPT as an assistant, not the final author.

“One sentence per point”: make each sentence in the abstract represent one paragraph in the introduction; then expand each into a paragraph using literature as supporting evidence.

Detailed workflow / methodology (step-by-step)

Literature study
- Collect and read relevant papers.
- Mark and index important papers; create a lookup table mapping paper references to the points or sentences you will use.
Create outline and points
- Decide the outline for the abstract/intro (general background, specific background, knowledge gap, results, implications).
- For each outline item, write one concise sentence to be expanded later.
- Note whether each sentence requires citations; attach citation IDs from your lookup table.
Prepare the prompt for ChatGPT
- Role: specify a role (e.g., “You are my supervisor/mentor, expert in writing scientific articles”).
- Context: state the research area (example: intelligent transportation systems, vehicle tracking with low-cost cameras).
- Constraints: required word count (e.g., 200 words), single-paragraph abstract, sentence style (active voice, no compound sentences), prohibited content (e.g., don’t include test values in the implication sentence), etc.
- Supply data: include test metrics (precision 0.896, success 0.755) and example limitations (low frame rate, low processing power).
- Citation instructions: tell the model whether to use citation placeholders or specific citation formats, or to include citation numbers from your lookup table.
Generate the abstract (two modes)
- Mode A (all-at-once): ask the model to generate the full abstract per the outline in one go.
- Mode B (stepwise): ask the model to produce each outline component separately (offers more detail and is easier to verify/edit).
Expand abstract sentences into introduction paragraphs
- For each abstract sentence, instruct the model to expand it into a paragraph, using the lookup-table citations to add supporting literature.
- Request explicit citation placeholders so you can later replace them with real references.
Post-processing & validation
- Paraphrase and synthesize outputs into your own writing voice.
- Verify factual claims and citations against the original papers; remove or correct hallucinated or inappropriate examples (e.g., replace “smartphone camera” with CCTV if applicable).
- Adjust wording to journal/language preferences (English/Indonesian).
- Optionally compare outputs from GPT-3.5 vs GPT-4 and choose or combine results.
Extras (prompted steps shown)
- Ask the model for keyword suggestions (e.g., five keywords, maximum two words each).
- Ask for title recommendations based on the abstract.
- Control how results are presented (for example, instruct the model not to mention numerical test results in the implication sentence).

Practical tips emphasized

Don’t let plugins or automatic tools replace your synthesis step—human validation is essential.
Use role-play (ask ChatGPT to be a supervisor) to elicit more formal/scientific phrasing.
Ensure examples or supporting devices requested are relevant to the research context.
Prefer stepwise generation for clarity and verifiability.
Use citation placeholders and a lookup table to keep generated text traceable to verified sources.

Specifics demonstrated in the video

Research context: intelligent transportation systems (ITS), vehicle tracking using low-cost cameras.
Method demonstrated: an approach called dynamic particle / “Dynamic SW particle optimization” to improve tracking performance.
Reported experimental metrics inserted into prompts: precision ≈ 0.896 and success rate ≈ 0.755.
Prompt constraints used in the demo: 200-word single-paragraph abstract; one sentence per general background point; active sentences only; limit keywords to two words; provide five keyword suggestions.

Speakers / sources mentioned

Presenter / demonstrator: an unnamed computer science student (references “Mas Alif” in context).
ChatGPT models: GPT-3.5 and GPT-4.
Role-play persona used in prompts: “supervisor” / mentor.
Research papers / references: used via a lookup table (collective sources rather than a single named paper).
Plugins / other AI software: mentioned as possible but cautioned against overuse.
“Prof.”: referenced as the source for the one-sentence rule.
Background music: non-speech audio present in the video.