You are testing a proposed cheap replacement for an expensive method,…

← back to session · subagent transcript, not indexed

2026-08-28

User:

You are testing a proposed cheap replacement for an expensive method, for the novel at redacted. Read CLAUDE.md first and obey it.

Background

A multi-agent experiment gave each character in a scene its own agent, so no character's dialogue was generated in a context containing the others'. It worked, but cost 6–10x a conventional pass. Its own findings (samples/round6-isolated/findings.md — read this in full, especially §2 and §5) conclude that most of the gain probably came not from context isolation but from the Director writing each line's want as a role-bound instruction before any line was generated — and that a single writer working from a pre-written want-list should recover most of the benefit at about a twentieth of the cost.

That confound has never been tested. You are the test.

What to do

Rewrite the same scene the other runs used: the opening scene of chapters/019-a-question-of-rosters.md — the four dead names struck off the roster at the long table on the twenty-second. Speakers: redacted, redacted, redacted, redacted, Edwin; Peri present and silent. Target 1,200–1,800 words. Same events, facts and canon; you may restage freely but may not change what is true.

The method under test, followed strictly and in this order:

  1. redacted sentence: unpublished prose Wants only. No dialogue, no phrasing, not even a fragment. Get the whole scene's spine down as wants first.
  2. Then draft the scene normally, as one writer, from the slot list.
  3. Then score it.

redacted sentence: unpublished prose

Scoring — this is the point of the exercise

Score your scene on the project's own two instruments, exactly as samples/round6-isolated/findings.md §2 does, so the numbers are comparable:

Score your own scene honestly. You wrote it, so you are biased; compensate by being harsh, and by scoring the source scene yourself with the same hand rather than trusting the other run's figures for it.

Deliverable

samples/round7-wants/ containing:

Then commit.

redacted sentence: unpublished prose Either answer is valuable; do not flatter the cheap method because it would be convenient.

Do not edit anything outside samples/round7-wants/.

Your final message: the three-way score table, your verdict, and one line on what surprised you.

I'll start by reading the required context.

30 tool calls

🔧 Bash: redacted sentence: unpublished prose

✅ 35 lines, 689 chars

🔧 Bash: cat samples/round6-isolated/findings.md

✅ 287 lines, 14928 chars

🔧 Bash: cat notes/voices.md

✅ 44 lines, 2191 chars

🔧 Bash: sed -n '150,561p' notes/voices.md

✅ 412 lines, 24371 chars

🔧 Bash: cat notes/voice-exemplars.md && sed -n '1,120p' notes/llm-tells.md

✅ 183 lines, 10969 chars

🔧 Bash: sed -n '120,261p' redacted/notes/llm-tells.md; echo ======; cat redacted/chapters/019…

✅ 144 lines, 8405 chars

🔧 Bash: cat redacted/chapters/019-a-question-of-rosters.md

✅ 199 lines, 21815 chars

🔧 Bash: ls notes/canon/characters/ && sed -n '1,340p' notes/dialogue-voices.md

✅ 348 lines, 23461 chars

🔧 Bash: sed -n '340,620p' notes/dialogue-voices.md

✅ 281 lines, 17017 chars

🔧 Bash: sed -n '620,780p' notes/dialogue-voices.md

✅ 160 lines, 8828 chars

🔧 Bash: cd redacted; for f in edwin redacted peri; do echo "=== $f"; awk '/^#+ .*[Vv]oice/{p=…

✅ 363 lines, 21996 chars

🔧 Bash: cd redacted; echo "=== EDWIN rules"; awk '/^\*\*4\. At an emotional peak/,0' notes/cano…

✅ 101 lines, 6284 chars

🔧 Bash: cd redacted; awk '/^\*\*9\. /,0' notes/canon/characters/edwin.md | head -60; echo "===…

✅ 151 lines, 9610 chars

🔧 Bash: cd redacted; grep -n "redacted\|redacted\|redacted" notes/canon/characters/minor.…

✅ 8 lines, 552 chars

🔧 Bash: cd redacted; sed -n '180,240p' notes/canon/characters/minor.md

✅ 61 lines, 3902 chars

Now chapter 018 and Peri's key rules.

🔧 Bash: cd redacted; tail -60 chapters/018-holm-deep.md; echo "=== PERI RULES"; awk '/^\*\*6\.…

✅ 150 lines, 10062 chars

🔧 Bash: cd redacted; cat samples/round4/parole.md

✅ 110 lines, 5361 chars

Now I have what I need. Step 1: the want-list, before any prose.

🔧 Bash: mkdir -p redacted/samples/round7-wants && cat > redacted/samples/round7-wants/slots.m…

✅ 1 lines, 2 chars

Step 2: the draft.

🔧 Bash: cat > redacted/samples/round7-wants/scene.md <<'EOF' # Round 7 — the scene Written in…

✅ 1 lines, 73 chars

Step 3: scoring. Both instruments, run by hand on my scene and re-run on the source.

🔧 Bash: cat > redacted/samples/round7-wants/findings.md <<'EOF' # Round 7 — wants without agent…

✅ 1 lines, 80 chars

Committed. samples/round7-wants/ has slots.md (27 wants, untidied, including the two bracketed doubts and the struck slot), scene.md (1,500 words, 30 turns), and findings.md.

Three-way table

Prop-cover

Sample Turns Place Rate
Source ch. 019, my scoring 45 22 49%
Source ch. 019, round 6's scoring 17 6 35%
Multi-agent, round 6's scoring 16 13 81%
Wants scene, my scoring 30 14 47%

Swap test

Source (mine) Multi-agent (r6) Wants
Headline 10/15 (67%) 12/14 (86%) 12/15 (80%)
On standing/fact 4/15 7/14 6/15
On voice alone 6/15 (40%) 5/14 (36%) 6/15 (40%)
r6's figure for source 1/15 (7%)

Per-speaker voice-alone is identical between source and wants scene: Edwin 3/3, redacted 2/3, redacted 1/3, redacted 0/3, redacted 0/3.

Verdict

The wants tie the published source and beat it on neither instrument, so they do not demonstrably recover a multi-agent gain — but this round could not find a gain to recover, because the same hand scores the source at 40% voice-alone where round 6 scored it at 7%. redacted sentence: unpublished prose The mechanistic finding is that wants and isolation attack different failures: a want says what a line must do, never who may do it that way, and all four defects in my scene are register bleed into Edwin's constructions (his grading shape, his copula-plus-possessive refusal, his professional modal, his balanced antithesis) — precisely what isolation prevents by construction. Recommendation: keep the wants (cheap, and they caught Edwin's absent-reader warrant sitting in redacted's slot before a word of prose), add a construction-ownership check to editing-pass, and pay one agent to re-score all three scenes blind before adopting or retiring anything. Cost was ~1.2–1.4x a normal draft pass, one context, against round 6's 6–10x.

What surprised me

redacted sentence: unpublished prose