Video summary

Claude OPUS 5 read my whole book then taught us how to prompt fiction

Main summary

Key takeaways

Technology

Summary of technological/product concepts (and what the creator learned)

1) Using Claude Opus 5 for AI-assisted editorial/reporting on a fiction draft

  • The speaker used Claude Opus 5 (soon after release) in a Backlist Intensive “working session” focused on workflow rather than directly editing books.
  • They prompted it to generate editorial reports with clear visual structure. Example shown: a 42-page AI-written report (partially displayed).
  • To test it, they provided a fully AI-written book (published earlier or submitted/presented for testing) and then observed:
    • The report surfaced serious flaws, suggesting the model’s analysis can catch issues even when the text hasn’t been heavily human-edited yet.

2) Story structure diagnostics via visual analytics

The report includes multiple chart-like breakdowns to analyze narrative mechanics:

  • Cozy genre pacing guidance

    • Example guidance: if you like a certain structure, an inciting incident around chapter ~6 is “probably too late” (for the structure being discussed).
    • Emphasis on pressure vocabulary / word choice to manage an ease vs. pressure feeling.
    • Notes that tension should rise but follow a cozy escalation pattern.

  • Beat timing & balance

    • Discusses how patterns repeat in genre media (e.g., Spider-Man beats and comedic fallout moments).
    • General rule: after big moments (action/romance), you often need immediate contrast beats (comedy fallout, separation, rejection, etc.) to maintain tonal balance.

  • Character “on stage” time (stage time)

    • A graph uses a GitHub-like colored grid / pixel-array visual:
      • X-axis: chapter numbers
      • Y-axis: characters
      • Color shades: how much each character is present/important per chapter
    • The speaker found it unexpectedly brilliant and useful for planning.

  • Character co-occurrence / connection visualization

    • Another chart shows how often characters appear with one another.
    • Critique: lines were thin vs. thick, but the nuance was hard to read; the speaker suggests color-coding or more varied line thickness.
    • They question specific gaps (e.g., “Vorna and Tam show up without connections”), treating this as either:
      • a report artifact, or
      • a real draft issue worth investigating.

  • Other report sections

    • The report also covers additional items such as point of view characters, themes, settings, dialogue, etc.
    • One speaker mentions “stealing” the visual analytics for a living series environment.

3) Dialogue proportion analysis (words vs. sentences/paragraphs)

  • The report distinguishes dialogue proportion by words, not by sentences.
  • Key insight:
    • Dialogue as a share of words is often < 30%, since dialogue snippets tend to be short.
    • Dialogue as a share of paragraphs may better match human intuition about how “dialogue-heavy” a chapter feels.
  • Observed oddities:
    • Chapter 2 has no dialogue, and other chapters have hardly any.
    • They suspect AI may add navel-gazing / introspection scenes when plotting without human direction, potentially confusing novel vs. screenplay pacing expectations.
  • Takeaway for upstream planning:
    • A scene where a character is alone may naturally lack dialogue, but the AI might still insert discussion-heavy material inappropriately.

4) Claude Opus 5 code-review behavior (precision/recall & multi-step prompting)

A later segment shifts to a more technical evaluation:

  • Claim: Claude Opus 5 can review code with high precision and recall, finding real bugs at a high rate per pass.
  • Additional finding: issues are “mostly real” rather than false positives.
  • Accuracy holds even at lower effort settings, supporting:
    • a fast first pass at review time
    • a more thorough pass later
  • Prompting lesson:
    • If the review prompt says “only high severity” or “be conservative,” the model may follow that literally and report less.
    • Recommendation: ask it to report everything, then filter in a separate pass.
  • Multi-step prompting advantage:
    • The model performs better when the task is broken into explicit steps rather than “one step doing three or four things.”

5) Prompting strategy for “effort,” and why length/verbosity instructions changed

  • They discuss that the effort parameter controls how much the model thinks, not how much it says.
  • For API usage, they criticize that examples often show only short versions of instructions and warn that prompting may need updating.
  • Example concern:
    • If you request a word range (e.g., 4k–8k words), they’re unsure how reliably Opus 5 can judge that through word count.

6) “How to measure fiction-writing length” — a detailed workaround

The creator tests Opus 5 with a prompt about measuring chapter fiction length.

  • The response rejects strict word-count thinking:
    • It says it uses “structural proxy” (e.g., estimate by paragraphs, exchanges, beats, scene completion, etc.).
  • They define a “register signal”:
    • it should feel like a full chapter, not a vignette.
  • Practical units mentioned:
    • Paragraphs of narration/dialogue
    • Story beats (e.g., “350–500 words” per beat as an estimate)
    • Scenes (e.g., “2–4 beats + entry/exit,” roughly 900–1,600 words)
  • Rough mapping they mention:
    • a target like 2,000–3,000 words corresponds to ~5–7 beats (often two scenes, or one long scene with a location shift/power shift mid-way).
    • roughly 35–50 paragraphs could be a reliable range.
  • Pipeline takeaway:
    • It’s hard to ask the model to decide exact word totals during drafting because word count is a byproduct.
    • More reliable planning is:
      • beats (fully planned)
      • then paragraphs (semi-planned)
      • letting words emerge from execution.
    • Recommended fix: ensure upstream planning by specifying enough beats so length becomes a consequence, not a constraint.

7) Efficiency guidance: low/medium effort can be enough

  • The speaker argues you generally don’t need max/high effort:
    • Lower effort often yields strong quality at a fraction of tokens, and is faster.
  • For planning drafts (e.g., outlining), they suggest iterative “cascade”-style passes:
    • outline iterations
    • checks for tension and characterization
    • incremental improvement via multiple low/medium-effort stages.

Key reviews/guides/tutorial items mentioned

  1. Workflow guide: prompt Opus 5 for editorial reports, then use its visual analytics to inspect structure.
  2. Guide for narrative diagnostics:
    • check inciting incident timing (e.g., around chapter 6 guideline),
    • balance tone via contrast beats (comedy after big moments; fallout after romance),
    • use character “on-stage time” charts,
    • evaluate character co-occurrence/connectivity and flag anomalies for review.
  3. Guide for dialogue analysis:
    • interpret dialogue % by words vs by paragraphs.
  4. Developer guide for code review prompting:
    • request all issues first; filter later,
    • use multi-step prompting for higher reliability.
  5. Prompting strategy guide for fiction length:
    • don’t over-rely on exact word counts,
    • plan using beats/scenes/paragraph decisions, letting word totals emerge.

Main speakers/sources

  • Stace (primary speaker; handles fiction workflow and testing)
  • “Claude Opus 5” (the model being evaluated and referenced throughout)
  • Additional participant (briefly responds/asks questions; not clearly named in the subtitles)

Original video