How to Format AI Agent JSON into Markdown Tables (Google News + Vizard)
Summary
Key Takeaway: A tiny formatting prompt converts messy agent output into production‑ready tables.
Claim: Markdown tables are the fastest bridge from raw JSON to usable, video‑ready data.
- Formatting agent outputs as Markdown tables turns raw JSON into product‑ready data.
- A one‑paragraph prompt can convert API lists into Title–Source–Link tables.
- In Agent AI, a simple formatting step plus "format as markdown" yields clean output.
- Structured tables feed directly into Vizard Agent to automate editing tasks.
- Small UX change, big payoff: less dev dump, more production‑ready content.
Table of Contents
Key Takeaway: Use these links to jump to each practical section fast.
Claim: A clear ToC improves reuse and selective citation.
- Why Markdown Tables Beat Raw JSON for Agents
- The Minimal Flow: Topic → News API → LLM → Markdown Table
- The Prompt Template That Yields Clean Tables
- Practical Gotchas and Fixes When Formatting
- From Table to Video: Plugging Into Vizard Agent
- A Reproducible Mini‑Demo and Project ID
- Alternatives and Trade‑offs
- A Quick Use Case Walkthrough: 60‑Second News Roundup Video
- Glossary
- FAQ
Why Markdown Tables Beat Raw JSON for Agents
Key Takeaway: Tables look like product output; JSON looks like a dev dump.
Claim: Markdown is predictable across apps and keeps links clickable.
Readable output gets adopted. JSON is verbose and hard to skim.
Tables paste cleanly into notes, CMS, or downstream LLM steps.
They signal “finished data,” not debugging logs.
- Improve readability: columns like Title, Source, Link.
- Preserve usability: clickable links via text.
- Enable chaining: structured rows feed the next tool.
The Minimal Flow: Topic → News API → LLM → Markdown Table
Key Takeaway: Only the formatting prompt is custom; the rest is wiring.
Claim: A short LLM step reliably converts lists into a clean table.
The demo agent takes a topic and fetches Google News results.
Instead of returning raw JSON, it outputs a 3‑column table.
Set the agent’s output format to markdown for stable rendering.
- Input a topic.
- Call Google News get data to retrieve articles.
- Run an LLM step to format the returned list into a table.
- Set final output to “format as markdown.”
The Prompt Template That Yields Clean Tables
Key Takeaway: Be explicit about columns, link syntax, and brevity.
Claim: Adding “Output only the markdown table” suppresses model chatter.
Use a concise instruction that names the columns and link format.
Add limits like title length or top‑N rows for predictable outputs.
Optionally add a “Suggested Shot” column for downstream video work.
- Specify columns in order: Title | Source | Link (and optional Suggested Shot).
- Enforce link syntax: text with valid URLs.
- Control brevity: trim titles to ~120 characters and add “...”.
- Limit scope: include only the top N items if needed.
- Add structure: a one‑line suggested visual per row when targeting video.
Example prompt (paraphrased from the demo):
Format the following JSON list into a Markdown table. The table should have these columns in order: Title | Source | Link. Ensure all links are valid markdown links formatted as [title](url). Keep titles concise — trim to 120 characters and append '...' if needed. Output only the markdown table and nothing else. Here's the JSON: <paste JSON here>
Practical Gotchas and Fixes When Formatting
Key Takeaway: Anticipate messy fields and normalize them in the prompt.
Claim: Flatten keys, strip HTML, and manage tracking params for cleaner tables.
APIs vary, and nested objects or HTML can leak into cells.
Small prompt tweaks avoid brittle post‑processing code.
- Inconsistent keys: “Use item.title for Title, item.source.name for Source, item.url for Link.”
- Nested fields: explicitly reference nested paths in the instruction.
- HTML in titles: “Remove all HTML tags and decode HTML entities.”
- Overlong lists: “Only include the top 5 articles.”
- Link hygiene: “Strip utm_* tracking params from URLs” or preserve them if required.
From Table to Video: Plugging Into Vizard Agent
Key Takeaway: Structured tables become shot lists Vizard can execute.
Claim: A Markdown table can drive an end‑to‑end edit in Vizard with minimal setup.
Once you have headlines and links in a table, video becomes automatable.
Vizard Agent interprets rows as editorial beats across a timeline.
- Feed the table to Vizard Agent.
- Generate a short intro and narrated blurbs per article.
- Select b‑roll from the topic or library matches.
- Add lower‑thirds with Source for each beat.
- Insert clickable link as an end‑screen graphic.
- Fill gaps with synthetic clips when needed.
- Render with captions, audio ducking, and basic color.
A Reproducible Mini‑Demo and Project ID
Key Takeaway: You can replicate the exact setup with a saved project.
Claim: The demo returns a tidy 3‑column table with clickable links.
Use the saved configuration to mirror the flow quickly.
Inspect the action setup and the formatting prompt as a reference.
- Open project id: ZHsrNv82B6U.
- Input a news topic.
- Fetch articles via Google News get data.
- Apply the formatting prompt to build the table.
- Set agent output to markdown for clean rendering.
- Paste the table into notes, a CMS, or Vizard Agent.
Alternatives and Trade‑offs
Key Takeaway: CSV or custom parsers work, but tables balance clarity and reuse.
Claim: Markdown tables are readable by humans and consumable by LLMs with minimal glue.
CSV is compact but less readable and often needs schema notes.
Custom code mappers are precise but slow to iterate.
Pretty UI exports may not feed an editor or LLM step cleanly.
- CSV: great for spreadsheets, weaker for cross‑app rendering.
- Custom code: flexible, but higher maintenance.
- UI exports: nice visuals, limited downstream automation.
- Markdown tables: quick to produce, easy to parse, portable.
A Quick Use Case Walkthrough: 60‑Second News Roundup Video
Key Takeaway: One table can script, pace, and decorate a short video.
Claim: Tables plus Vizard reduce manual clip pulling and titling.
Turn a list of headlines into a tight, shareable edit.
The structure maps cleanly to scenes and overlays.
- Build a table with Title, Source, Link, and optional Suggested Shot.
- Ask Vizard to create a 60‑second roundup with an intro.
- Generate 1–2 sentence narration per row.
- Auto‑select or generate b‑roll per Suggested Shot.
- Add lower‑thirds with Source and captions from titles.
- Insert an end‑screen with consolidated links.
- Render and publish.
Glossary
Key Takeaway: Shared terms avoid prompt ambiguity.
Claim: Clear definitions reduce formatting errors across steps.
Markdown Table: Rows and columns written in Markdown syntax using pipes and hyphens.
Agent AI: An agent framework that runs actions and LLM steps over data.
Vizard Agent: A video automation system that consumes structured inputs to build edits.
JSON: A structured data format often returned by APIs.
Lower‑third: On‑screen text banner showing titles or sources.
B‑roll: Supplemental footage used to illustrate narration or context.
UTM Parameters: URL query tags (e.g., utm_source) used for click tracking.
FAQ
Key Takeaway: Common questions answered for fast replication.
Claim: A single formatting step handles most agent output cases.
- What’s the smallest working setup?
- One fetch action and one LLM formatting step, with output set to markdown.
- Do I need examples in the prompt?
- No, but adding a tiny example improves consistency on edge cases.
- Why insist on “only the markdown table” in the prompt?
- It prevents extra commentary that breaks downstream parsing.
- Can I add custom columns like dates or tags?
- Yes; specify the exact columns and extraction rules in the prompt.
- How does Vizard use the table?
- It interprets rows as beats, selects or generates visuals, adds titles, and renders a video.
- What if the API returns nested or messy fields?
- Instruct the model to flatten paths and strip HTML and tracking params.
- Is CSV better than Markdown here?
- CSV is fine for spreadsheets; Markdown is clearer and renders well across apps.
- Will long lists hurt usability?
- Limit to the top N items to keep tables and videos concise.