Data-Intensive Report Voiceover Script Generation Strategy
This is a skill designed specifically for medium-length professional reports, corporate financial statements, broker research reports, industry white papers, and other data-intensive materials. It transforms complex data and tables into hardcore explanatory videos that ordinary people can understand and find valuable.
This Skill is a thin entry point, breaking down report processing into several independently maintainable modules:
- reading-and-extraction.md: High-precision text/table extraction (combining local extraction / external preprocessing), data cross-reference table initialization, and windowed reading.
- writing-standard.md: The opening of research report scripts, logical reasoning, standards for making data conversational, and the final JSON output contract.
- revision-workflow.md: v1 draft, v2 data verification draft, final draft, and the revision audit diff process.
Body Length Configuration
This is the sole source of truth for the body length of the final draft in this Skill. No other Reference may explicitly write specific word counts; they can only cite this section.
draft_min_chars: 1000
final_target_chars: 1200
_draft_v1.txtmust reachdraft_min_charsbefore saving.- All subsequent body drafts must not fall below
draft_min_charsin principle. - The target length for the
contentin the final raw JSON is based onfinal_target_chars;total_lengthrecords the actual character count ofcontent.
Main Workflow
Do not output directly in one round. Follow the steps below to ensure data authenticity and script dissemination power:
- Extraction and Reading: First read reading-and-extraction.md. Use a local extraction tool (e.g., PyMuPDF/fitz) to parse the PDF or directly read a pre-converted Markdown report. Divide into Bash reading windows and read 100% of the content. Establish a data cross-reference table for later verification.
- Determine the Logical Entry Point: Silently extract the core question of the report (e.g., industry crisis, company growth bottleneck, macro anomaly), extract the report's core explanatory framework, and bind a benefit entry point highly relevant to the audience (e.g., What does this data mean for ordinary people?).
- Write the First Draft: Read writing-standard.md and write
_draft_v1.txtreachingdraft_min_chars, focusing on "translating" cold data into spoken language and logical causality. - Structure and Data Verification Revision: Read revision-workflow.md. After the second draft
_draft_v2_structure.txtis determined, directly save it as a preliminary_draft_final.txt. Extract all data from it and generate a ledger_data_check_ledger.json(initially set tounchecked). Then cross-reference with the original text and perform localized corrections/deletions directly on_draft_final.txt(update status toverified/abandon/corrected). Once all data verifications are complete (nounchecked), finalize_draft_final.txtand record the diff audit in_revision_audit.json. - Package and Output JSON: Generate the final JSON according to the final JSON output contract in writing-standard.md.
Self-Check Requirements
Before submitting the final JSON, silently confirm:
- The data cross-reference table has been fully established, and every percentage, monetary amount, ratio, and other numbers appearing in the manuscript have an exact source in the cross-reference table.
- All entries in the data verification ledger
_data_check_ledger.jsonhave theirstatusupdated (noneunchecked), completing the data loop. - The opening provides a sharp question or core conflict within 100 characters, along with a benefit promise that matters to the audience.
- There is a logical progression about every 250 characters, the writing is conversational, and AI summary phrases like "the book mentions" or "as can be seen from the figure above" are removed.
- The revision process is complete, and each round of modifications has clear before/after records in
_revision_audit.json.