AI-Assisted Creative Production

What it is

A systematic, engineering-inspired workflow for producing long-form content (audio essays, documentaries, narrative nonfiction) with AI assistance. The core premise is that creative work is not pure inspiration but a decomposable process of material collection, functional classification, card-based drafting, sequencing, and style migration. The method was developed by the 诗梳风 podcast team during a week-long closed development sprint in Beijing's Sanlitun district, where they attempted to replicate a human creator's decision-making patterns through structured decomposition.

Why it matters

Most AI-assisted writing defaults to either (1) prompt-and-polish, which produces flat, AI-sounding prose, or (2) full delegation, which strips the creator of judgment. The 诗梳风 workflow demonstrates that AI can be inserted at the process level rather than the output level, preserving the creator's taste while scaling the mechanical work of material handling, structural sorting, and stylistic calibration. It validates that "taste" is not mystical but a set of implicit criteria that can be surfaced, documented, and encoded into workflow rules.

Key Principles

1. Human Baseline First

Before AI optimization, the team must produce a human-written sample that serves as the quality anchor. Without a shared human baseline, the team cannot evaluate whether AI outputs are improving or merely proliferating. This principle is documented in Human Sample Baseline.

2. Wide-Entry, Strict-Exit Material Collection (宽进严出)

Material gathering is the most time-intensive phase and is itself a creative act. The team spent five-plus hours per building researching printed books, newspapers, and academic papers—not web pages. Web search is keyword-bounded and tends to return exactly what you asked for; physical archives and databases serendipitously surface adjacent contexts that cross-pollinate the narrative.

Key tactics:

• Cross-domain borrowing: A technical paper from a 1988 high-rise engineering conference in Qingdao became the richest source for a story about a diplomatic apartment complex in Beijing—because the paper contained specs, dates, and human details that no architecture-focused web search would reveal.
• Rejecting off-topic materials: Glamorous memoirs about Sanlitun's nightlife were discarded because they were personal stories, not place stories. The act of rejection is as creative as the act of inclusion.
• Quantity before quality: "You can't be selective before you've collected." Five hours of collection per unit was the observed minimum for meaningful selection.

3. Functional Classification: Fact, Story, Resonance

All collected material is tagged by its narrative function, not its topic:

Materials are split into two master databases: Fact Bank and Story Bank. Stories are further scored for emotional resonance by a dedicated agent persona ("experienced documentary director"). The highest-scoring stories become Super Cards that anchor the opening, closing, and major transitions.

4. Card-Based Writing (卡片化写作)

The entire narrative is built from cards of 100–400 words each. This is not a stylistic preference but an engineering necessity: when context windows were smaller, the only way to process 100,000 words of source material into a 5,000-word output was to decompose the task into digestible chunks. Even with larger context windows today, the card system preserves human inspectability at every step.

Card workflow:

1. Source material is chunked into cards by function (Fact vs. Story).
2. Cards are deduplicated and merged (MECE-ified).
3. Unreliable or low-quality cards are discarded.
4. Super Cards are promoted from the Story Bank.
5. Cards are grouped into Super Groups (macro-sections), then into sub-groups, with logical ordering enforced by dependency (prerequisite information must precede dependent information).

5. Scaling and Pacing Control

AI-generated longform often suffers from two pacing diseases:

• Egalitarianism: unwilling to cut stories, treating every anecdote equally.
• Flatness: failing to expand the most important stories and compress the rest.

The workflow assigns explicit criteria for scale decisions:

• Which stories deserve full expansion (high emotional stakes, structural anchor, unique angle).
• Which stories deserve compression or deletion (redundant, weakly sourced, structurally dispensable).
• When human override is required for tricky edge cases.

6. Style Migration: Reverse-Engineering Voice

Style is not a surface polish added at the end; it is a middle-layer communication technique. The team used a super-resolution analogy to capture a writer's voice:

1. Take a finished piece by a target writer (e.g., Gao Xiaosong).
2. Imagine the "dry, encyclopedic, un-stylized" version that might have preceded it.
3. Generate a before/after pair.
4. From many such pairs, extract guidelines: what does this writer consistently do at the structural and tonal level?

Key insight: you can reverse-engineer voice from the output without knowing the writer's creative process. The team generated 50 transcripts per target writer, created dry-to-stylized transformation pairs, and distilled them into reusable style modules. These modules were applied as the final pass—after all factual and structural decisions were locked—so that style could not corrupt content.

7. Modular Beginnings and Endings

The opening and closing are treated as separate modules with fixed internal architectures:

Opening structure (1,000 words):

1. Most interesting micro-story (human-scale detail, time/place/character).
2. Pull back to the big picture (why this matters).
3. Most emotionally resonant story.
4. Core factual anchor.
5. Second-most interesting story.

This inverted-pyramid opening was observed as a recurring pattern across the team's human-created work and was formalized as a reusable template. The closing was similarly modularized.

8. Decompressed Source Retrieval

After cards are sorted, scaled, and sequenced, a final decompression step is required: for every card that was summarized during sorting, the system must retrieve the original source text and re-inject it before final drafting. Without this step, the narrative loses texture and detail. This requires maintaining traceable indices throughout the pipeline.

Workflow Summary

Material Collection → Functional Tagging → Card Generation
                                               ↓
                              Dedup / Merge / Discard
                                               ↓
                              Super Group Formation
                                               ↓
                              Card Sequencing & Scaling
                                               ↓
                              Decompressed Source Retrieval
                                               ↓
                              First Draft (neutral, encyclopedic)
                                               ↓
                              Style Migration (target voice)
                                               ↓
                              Opening Module Rewrite
                                               ↓
                              Closing Module Rewrite
                                               ↓
                              Final Human Review

Open Questions

• Does the card-based system remain necessary with 1M+ context windows, or does it now serve a different function (human inspectability, version control, team coordination)?
• How generalizable is the style-migration methodology beyond Chinese audio-essay production? Can it be applied to English technical writing, fiction, or marketing copy?
• What is the minimum viable corpus size for reverse-engineering a voice? 50 transcripts per writer was the team's choice, but no systematic testing was performed.
• The workflow assumes a team with both creative and engineering roles. Can a solo creator execute this pipeline without the "three editors" dynamic that surfaced the implicit rules?

Related

• writing/human-sample-baseline — Human sample as prerequisite for AI optimization
• writing/writing-with-ai — Katie Parrott's five-stage AI writing workflow (interview → outline → draft → review → line edit)
• writing/ai-assisted-nonfiction-material-mining — AI-assisted material collection for nonfiction
• product-trends/software-factory-ladder — Industrialization of creative/technical processes
• harness-engineering/loop-engineering — Systematic loop design as engineering discipline

Pending Jean Review

• This page is candidate status from a single source (podcast transcript). Does this workflow resonate with Jean's own creative practice, or should it be merged into a broader writing methodology page?
• The "Wide-Entry, Strict-Exit" material strategy may conflict with or complement Jean's existing information-intake habits. Verification needed.
• The style-migration technique is unverified beyond the podcast team's report. Should be treated as hypothesis until replicated.