voice-smooth

name	voice-smooth
description	Render the verbatim scene-extraction quotes into readable, in-voice prose — a derived scene_extractions_smoothed/ layer that sits between /session-summary-consistency and session_doc. Uses each character's voice file as the guardrail. Fixes garble, run-ons, and disfluencies for readability WITHOUT changing what was said; never touches the verbatim (the VTT stays the raw record). Invoke as /voice-smooth [session-dir].
tools	Read, Bash, Write, Edit, Glob, AskUserQuestion

Voice-Smooth — voice-aware readability rendering of quotes

Turn the raw, transcriber-corrected verbatim quotes into readable prose that still sounds like the character who said it, written to a derived scene_extractions_smoothed/ layer for session_doc to consume. This is a rendering step, not an error-fix and not narration.

Where this sits (and the one inviolable rule)

VTT  (raw verbatim — IMMUTABLE, forever the record + the voice-file raw material)
  → scene_extractions/           (verbatim; transcriber errors fixed by /session-summary-consistency)
  → [THIS SKILL] voice-smooth    (readable, in-voice; guarded by the voice/ specs)
      → scene_extractions_smoothed/   (DERIVED — `## Voiced moments`, what session_doc reads)
  → session_doc narration
  → /voice-critic (checks the finished narration against the voice spec)

Inviolable rule: the verbatim is never mutated. This skill only ever writes scene_extractions_smoothed/. It does not edit the VTT or scene_extractions/. Because the raw quotes live permanently in the VTT, the smoothed layer is a safe derived rendering — not a rewrite of a record.

The optimization target flips here — this is the point of the layer

Every upstream layer optimizes for precision: what was actually said, by whom, in what order. That is correct for a record, and the VTT and scene_extractions/ remain that record forever.

This layer optimizes for narration instead. It is the last stop before session_doc, it writes only a derived copy, and nothing downstream of it is a record of anything. So the question at every quote is not "is this exactly what the tape captured?" — upstream already answered that and its answer is preserved. The question is "will this read well, in this character's voice, when the narrator reproduces it literally?"

That flip licenses edits an upstream pass would refuse:

• Recovering the word a garble is hiding. "I app and slash" is not what anyone said; it is what two ASR passes made of "I hack and slash". Upstream's job is to record that the tape says app. This layer's job is to hand the narrator a line that works.
• Completing a cue the recorder cut, when the meaning is recoverable ("Who are you talking." → "Who are you talking to?", "I rehearsed it for, like, 15." → "…for like 15 minutes.").
• Reassembling one sentence spread across four broken cues.
• Splitting stage direction out of NPC speech (step 2 rule 5 — the single biggest win).
• Dropping a disfluency that belongs to the player, not the character, when the voice file says otherwise. Stéphane says "you know" constantly; Brewbarry's spec says short, declarative, does not hedge. The voice file is authoritative, so the tic is noise here even though it is faithfully on the tape.

The guardrail is approval, not abstention. Because these edits change words, the human approves each one — see step 2.5. Do not apply them on your own judgment, and do not skip them out of caution either. Bring the candidate, the evidence from every transcript, and the options; let the GM rule.

Still off-limits, approved or not: facts, names, numbers, mechanics, attribution, and meaning. Fixing how a line reads is this layer's job. Changing what it says never is.

What this is — and is NOT

• IS: readability + voice, rendered for the narrator. Collapse run-ons, drop filler-as-noise, repair unreadable fragments, punctuate, recover garbles — while preserving each character's register, tics, and vocabulary per their voice file.
• IS NOT — silent error correction. Every garble repair is a GM ruling made here, one at a time, with the tape evidence on the table (step 2.5). What it is not is a unilateral fix, and it is not a substitute for /session-summary-consistency, which still owns the glossary and the upstream record.
• IS NOT — narration. Turning quotes into flowing scene prose is session_doc's job. This produces cleaner quotes, not narration.
• IS NOT — critique. Checking finished narration for voice drift is /voice-critic's job.

Precondition — run /session-summary-consistency first

Smooth corrected quotes, not garbled ones. If the verbatim still contains obvious transcription errors, run /session-summary-consistency first — otherwise you will fluently render a mistake, and a fluent mistake is much harder to catch than a garbled one.

That pass is where proper nouns get settled, because it rules against the glossary and the entity registry and writes the ruling back to both. Never decide a name here: a name is an identity decision, it recurs across every future session, and this layer has no way to record it. Hulkrist→Alkrist is upstream's call, always.

What it does not catch is common-word garble — app/hack, confront-free/conflict-free, bloom/show. Those aren't in any glossary, they don't trip a registry check, and they survive every upstream pass. They surface here, when you read the line and it doesn't parse. Rule on them here, one at a time, with the GM (step 2.5) rather than deferring them upstream to a pass that structurally cannot see them. Record what you ruled so /session-summary-consistency can promote anything that turns out to recur.

Inputs

• session-dir — summaries/YYYYMMDD/. Default: CWD (if CWD is the campaign root, ask which session).
• scene dir — <session-dir>/scene_extractions_new/ or <session-dir>/scene_extractions/ (suffix varies). Detect it; call it <scene-dir>.
• voice files — from <campaign-root>/voice/. Filenames are not limited to <char>_voice.md; resolve them with the three-rule lookup in step 1, which mirrors what the pipeline does. Plus voice/_genre.md (overall tone). Authoritative for how each character speaks.
• player→character map — from the glossary ## Player names → characters section (only needed if any labels still carry real names — they shouldn't after /session-summary-consistency).

Workflow

1. Locate + load the guardrails

• Detect <scene-dir>; if missing, stop.

• Read voice/_genre.md for overall tone.

• Resolve voice files the way the pipeline resolves them. The rule lives in CampaignGenerator session_doc/voice.py (load_voice_files + _resolve_voice_key, CampaignGenerator#247). Do not use a narrower one: voice/*_voice.md matches zero files in Phandalin, whose specs are brewbarry_new_pipeline.md, soma_new_pipeline.md, valphine_new_pipeline.md and vukradin_new_pipeline.md — so all four PCs of the campaign smoothing is standard on would get no spec, and this skill's own "voice files are authoritative" rule could not be honoured.

  Build the key set. Glob voice/*.md. Skip every file whose name begins with _ — _genre.md and friends are shared campaign material, not a per-character spec. For each remaining file the key is the lowercased stem with a trailing _voice removed: Brewbarry_voice.md → brewbarry, vukradin_new_pipeline.md → vukradin_new_pipeline.

  Resolve each speaker against that key set, stopping at the first hit:

  	Rule	Example
  a	exact full lowercased name	Unla Key → unla key
  b	first name only	Unla Key → unla
  c	the unique key beginning with the first name followed by _ or -	Vukradin → vukradin_new_pipeline

  Refuse on ambiguity. If rule (c) matches two or more keys, that speaker has no resolvable spec: do not guess which file the pipeline would use. Report it as ambiguous, list the candidates, and treat the spec as missing.

  Campaigns whose files happen to be named daz.md, grygum.md (out-of-the-abyss) hit rules (a)/(b), which is how the narrower rule went unnoticed.

1.5. Survey the whole scene set before smoothing any of it

Three defects are invisible one-scene-at-a-time and change what you render, not just how. Find all three before writing a single file.

Duplication across scenes — check every pair

Extractors routinely emit the same quotes into two or three neighbouring scenes. Smooth them all and session_doc narrates the same beat two or three times.

cd <scene-dir>
for f in 0*.md; do grep '^> "' "$f" | sed 's/^> //' | sort -u > /tmp/q_$f.txt; done
for a in /tmp/q_0*.txt; do for b in /tmp/q_0*.txt; do [ "$a" \< "$b" ] || continue
  n=$(comm -12 "$a" "$b" | wc -l); [ "$n" -gt 3 ] && echo "$n  $(basename $a)  <->  $(basename $b)"; done; done

One or two shared lines is noise — "Yes.", "Yeah.", "Exactly.". Real duplication is loud: Phandalin ch48 scored 67 of scene 01's 98 quote lines repeated verbatim in scene 02, with a shared tail across three scenes.

De-duplication is a SCOPE decision — what belongs where — and scope belongs to the human. Never draw the boundaries yourself. Present the overlap with counts, propose cut points, and get them approved.

Then rescue before you cut. Diff the losing copy against the keeper and read every line unique to it. In ch48 exactly one line — Brewbarry's "I hack and slash." — lived only in the half scene 01 was losing, while its setup line lived only in scene 02. Cutting blind would have destroyed the joke.

comm -23 /tmp/q_<losing>.txt /tmp/q_<keeper>.txt   # lines about to be lost — read them all

Leave a scene boundary note in each affected smoothed file saying what moved where and why, and strike through (don't delete) the now-relocated bullets in the copied ## Scene summary so the gm-assist original stays auditable.

Extractor splices

A quote fused mid-string with the next speaker's header — > "…a typical medieval bank**GM** — *introducing the banker*. Structurally broken markdown that no consistency pass catches.

grep -rn '[a-z0-9…"]\*\*\[\?\(GM\|<PC names>\)' <scene-dir>/0*.md

Repair in the smoothed layer (the split is unambiguous), and report that scene_extractions/ still carries it. Phandalin ch48 had four, in scenes 01, 02, 04 and 08 — one of which had survived a full /staged-consistency run.

Player knowledge wearing a character's name

The highest-consequence defect this layer can catch, and it is not a transcription error at all. A VTT labels by speaker, so when a player says something out of character — recalling another campaign, reasoning from GM-side knowledge, metagaming aloud — it lands in the transcript under their character's name. The narrator then reproduces it as an in-fiction thought, and the character now knows something they cannot possibly know.

Phandalin ch48 had two:

• Soma's "Were they Cambions?" and Valphine's "they're Cambions" — the players met the cambion/House Margaster connection in a different adventure. No character in the campaign knows it.
• Vukradin's "I have a feeling I know who the gnome is" — the player knows who KP is. Vukradin does not.

You cannot detect these from the tape, because the tape looks identical either way. Detect them by reading for knowledge the character has no route to — a fact nobody on-page told them, a conclusion that arrives too fast, a proper noun from outside the campaign. When one is even plausible, ask the GM; they are the only one who knows what was established at the table.

When the GM confirms one:

1. Keep the line (it was said) but relabel it OOC and annotate it — state plainly that it is the player's knowledge, name what the character does not know, and write "Do not narrate X as suspecting/asserting this."
2. Record it durably. Annotations live in a derived file that the next scene_extract run deletes. The boundary must outlive it — a hand-authored dossier at docs/<Subject>.md with a what the party actually knows table, plus a pointer from the campaign's CLAUDE.md so it loads every session. Phandalin keeps docs/KP.md and docs/Margaster.md this way.

2. Smooth each quote (per scene, from the source's ## Verbatim moments)

For every quote block:

1. Identify the speaker and load their voice spec.

2. Produce a smoothed version that is:

   • Readable — complete sentences, filler-as-noise removed, false starts collapsed, garbled fragments repaired only where meaning is unambiguous (else keep the fragment or mark [unclear]).
   • In-voice — keeps the character's register, characteristic vocabulary, sentence rhythm, and signature tics. Thorin's clipped bluntness and Zalthir's hedging verbosity are voice, not error — preserve them. Do not homogenize everyone into the same neutral prose.
   • Faithful — same meaning, same content, same names, numbers, mechanics, and attribution. Add nothing; drop no substance. If you can't smooth a line without changing what it means, leave it closer to verbatim.

3. OOC / table chatter (jokes, rules talk, real-world tangents): smooth lightly for readability only — do not force it into in-character voice, and keep any OOC marker. When in doubt, leave OOC lines near-verbatim.

   Exception — pure tooling/logistics noise gets cut, not smoothed. Not all OOC content is table talk. A distinct sub-class carries zero narratable content and doesn't belong in ## Voiced moments even lightly smoothed: VTT/roll20 tool-fighting ("Where's the init thing?... run script... screw it"), map-token placement ("put this here — Bloodstained"), and off-mic real-world asides to someone in the room ("get out of there... sorry"). None of it is banter about the game — it's the players operating the software, or talking to someone who isn't at the table. Phandalin ch3 (2025-05-28 session): the GM ruled the init-tracker fight in scene 01 should be deleted from the smoothed layer, not merely smoothed — and that standard was then applied to every other instance of the same class found while smoothing the rest of the session (the map-token line, Soma's off-mic aside), not just the one line the GM actually reviewed. Generalizing a ruling beyond its original scope is itself a judgment call — flag it explicitly for confirmation rather than assuming the GM's intent covers every future instance silently.

   Cut it; don't fake-render it. Deleting is different from marking [inaudible] (that tag is for lost meaningful content) — this is content that was captured cleanly but isn't part of the scene. Leave an italic note at the top of the affected sub-section recording what was cut and why, so the omission stays auditable rather than silent: Record each cut in the manifest's (or equivalent) so a later reader can see the ruling was made once and applied consistently, not re-litigated per line.

2.5. Rule the garbles with the GM

Any repair that changes a word rather than punctuation, filler, or sentence boundaries is a GM ruling. Collect them per scene and show a running checklist so the GM can see how many are left.

Default to one at a time. Do not apply a garble because the answer seems obvious.

But strict 1x1 does not survive volume, and pretending otherwise is its own failure. Phandalin ch48 produced 74 rulings across 8 scenes — scene 05 alone had 18. Seventy-four separate questions exhausts the GM long before scene 08, and an exhausted GM rubber-stamps. Once a scene exceeds roughly six candidates, offer the split explicitly and let the GM choose it:

• Batch the ones a second transcript settles outright — where you are not reconstructing anything, it simply heard the word Zoom dropped. List every item in the batch in full, with its before → after, in the question itself. A batch the GM cannot read is a silent decision wearing a checkbox.
• 1x1 everything else: transcripts disagreeing, nothing corroborating, or a word you are inferring from context.

A third tier appears once the run is long: candidates that are mechanical applications of a ruling the GM already made this session (a proper noun settled two scenes ago, recurring in a new one). Those can join the batch — but say so, and name the earlier ruling you are applying.

Never batch: anything that establishes canon, any name not already settled, and anything where your recommendation is a guess. Those stay 1x1 no matter how long the queue is.

Reporting at volume — the interactive review artifact

Batch-vs-1x1 is about how many decisions to bundle into one question; it doesn't solve the problem of 74 rulings across 8 scenes exhausting the GM's attention regardless of bundling. For a full-session run, publish an interactive Approve / Reject / Discuss HTML artifact instead of (or as well as) walking the queue in chat. Demonstrated end-to-end on Phandalin ch3 (2025-05-28): a full 10-scene pass reviewed in two rounds, each round-tripped by download.

Build it with the Artifact tool (artifact-design skill for treatment — this is a utilitarian tool, not editorial; artifact-capabilities skill for the downloads contract):

• Declare capabilities: {"downloads": true}.
• One card per candidate, grouped by scene (or by decision class for calibration pairs — step 4). Each card carries exactly what a 1x1 chat question would: every transcript's reading named, corroboration found, the voice-file/framing rationale, the full-sentence before→after preview, and a keep-verbatim option.
• Each card gets a segmented Approve / Reject / Discuss control. Discuss reveals a <textarea> for the GM's own words — never force a bare reject when they want to explain or counter-propose.
• A sticky tally dock (counts per verdict + how many remain undecided) so a long queue stays legible instead of becoming a wall to skim.
• A "Download decisions" button: claude.use("downloads"), then downloads.save({filename, data}); if the namespace resolves null, reveal a fallback <textarea> pre-filled with the same content so the GM can copy-paste manually instead.
• Export a single Markdown decision record, one section per candidate, verdict + any discuss-text — structured so you can parse it back mechanically.

The round trip: the GM downloads the record, then either pastes it into chat or names the saved path (on WSL2, a Windows-side save lands at /mnt/c/Users/<user>/Downloads/..., directly readable). Read it back and apply per verdict: Approve → apply the garble fix or calibration ruling exactly as a chat "yes" would; Reject → skip, don't re-litigate; Discuss → resolve using the GM's own text in conversation before acting — never auto-apply a Discuss item on your own reading of their note.

This does not relax the confirmation requirement — every garble ruling still needs an explicit GM verdict, per-item. It changes the delivery mechanism at volume, not the checkpoint itself. Below the batch threshold (a handful of candidates in one scene), a chat question is still simpler; reach for the artifact once the queue is long enough that serial chat would exhaust the reviewer.

For each candidate, bring:

• Every transcript's reading, named. They disagree, and the disagreement is the evidence. Zoom heard app and slash; the re-transcription heard act and slash. Two passes garbling one phrase two different ways is the signature of a real utterance neither caught — and it is far stronger evidence than either pass alone. When both agree, say so; that is a reason to leave it.
• Corroboration from elsewhere in the tape. The GM said "all about this hack and the slash" twenty minutes earlier. Grep for the phrase before asking.
• What the voice file or the scene's own framing implies. The GM called it "the soft bathrobe"; a pitch about softness fits, a pitch about roughness does not.
• Options as full-sentence previews, so the GM reads the line as it will land, not a diff of one token.
• A "keep it verbatim" option, every time. Sometimes the garble is the joke ("Oral B. Vance"), and sometimes the GM remembers what was actually said.

Know which transcript to trust for what — it differs per campaign and is worth recording in the manifest. Phandalin: the Zoom export is better on proper nouns, the re-transcription is better on sentence completion (it recovered "15 minutes" and "Who are you talking to?" where Zoom cut both).

When the transcripts disagree and nothing corroborates either, say so plainly and let the GM decide. Do not pick the more fluent reading because it is more fluent.

Grep the anchor, not the garble

The obvious search fails on every single candidate, and it fails silently — you get zero hits and conclude the other transcript doesn't cover the scene. It does. You are grepping for a word that, by definition, the other pass did not produce.

Search instead for a clean phrase adjacent to it — a distinctive line one or two cues away that both passes almost certainly got right — then read the region around the hit:

grep -n -B6 -A6 -i "<clean anchor phrase>" <other-transcript>.vtt | grep -vE '\-\->|^[0-9]+[-:]$'

Every hard recovery in Phandalin ch48 came this way: anchoring on "civic notice" produced "Okay, what's on the notice?" for Zoom's meaningless "In order to sake."; "disappear anyone" produced "I love having Brewbarry" for "I love having blueberries."; "recently unalived" produced "Not that long." and "He pauses." for "Not bad, Lol." and "We can pauses."

Before concluding a transcript lacks coverage, prove it with an anchor, not with a failed search for the garbled token.

A garbled number is still a number

numbers and mechanics are on the never-change list, which makes it easy to skim past a dice value that ASR turned into a word. Phandalin ch48: Vukradin's persuasion roll reads "Event persuasion. Does that help?" in Zoom — the re-transcription has "27 persuasion." It reads as prose and is in fact the roll that decides whether an NPC hands over a plot-critical name.

When a line mentions a check, a roll, or a quantity and the value is absent or reads oddly, treat it as a high-priority candidate and go to the tape. Leaving it garbled is not the safe option: the narrator will either invent a number or drop the mechanic.

Write new wrong-forms back to the glossary

A proper-noun garble ruled here is one the upstream pass will meet again next session. When the GM settles one, offer to add it to notes/vtt_transcription_corrections.md as part of the same ruling — Phandalin ch48 added Utgartian, Rieber, Vubert, Rueberg, CORN, Corin, Don Juan.

Checking the glossary also audits it. That pass found the Don-Jon Raskin row pointing at an unhyphenated canonical form, contradicting entity_registry.yaml — a live bug that had been quietly producing inconsistent output. The registry is the authority; the glossary is a lookup table that can drift from it. When they disagree, say so and let the GM pick which one is wrong.

3. Write the derived layer

Write <session-dir>/scene_extractions_smoothed/NN_slug.md, mirroring the verbatim file's structure (frontmatter, ## Scene summary, a moments section, and the same speaker labels) so it is a drop-in for session_doc — with one deliberate exception, below. Mark it as derived:

• frontmatter source: voice-smoothed and from: ../scene_extractions/NN_slug.md

• rename the moments heading to ## Voiced moments. Do NOT copy ## Verbatim moments across. The heading is a claim, and CampaignGenerator's session_doc/io.py binds it to one: ## Verbatim moments says these are the tape's words, which this skill has just stopped being true. ## Voiced moments says tidied for reading; not exact — which is what a smoothed file is (CampaignGenerator#250 R5).

  This is not cosmetic. The heading is what drops the file out of the contract axis: rules R1 and R3 police exactness inside a span marked verbatim, and firing them on prose that openly declares it was edited produces refusals for edits this layer exists to make. (It does not change the verdict counts — unverified still means untraceable to any transcript line, which is a fabrication or a splice either way, and remains a defect here too.)

  Phandalin's scene_extractions_smoothed/ currently carries ## Verbatim moments on every file, which is the state CampaignGenerator#304 detects and warns about.

• copy the ## Scene summary across, but scrub player real names as you copy (e.g. Kostadis → GM, via the glossary ## Player names → characters map) — these summaries routinely leak the GM's real name, and copying it forward propagates the leak. Flag (don't rewrite) any transcription garble you notice in the summary (its origin is gm-assist / the scene-extract step). Leave the summary's wording and facts otherwise unchanged.

• replace each verbatim quote's text with its smoothed rendering under the same speaker label

Do not modify anything under <scene-dir>/.

4. Human review — REQUIRED before it feeds session_doc

Smoothing changes words, so the human is the checkpoint (LLM drafts → human reviews → then it feeds session_doc).

Calibrate on one scene first. On a first run for a session (or a new campaign), smooth a single representative scene, present its pairs, and get the voice fidelity and the grammar-fix aggressiveness approved before rendering the rest. It catches over/under-smoothing early and keeps the review tractable. (Calibration question that came up this run: how aggressively to repair grammar the player actually spoke — clear-meaning fixes like "we got a nail Bookwyrm" → "we've got to nail Bookwyrm" are fair game; ambiguous ones stay near-verbatim; suspected transcription errors get flagged upstream, never smoothed away.)

Present the calibration scene as verbatim → smoothed pairs, grouped by decision class — filler removal, grammar repair, stage-direction split, truncation handling — not as an undifferentiated list. The GM is approving policies here, not lines; one representative pair per class, plus every pair where you are least confident, is what makes that possible. Flag explicitly any rendering that risks changing meaning, flattens the character's voice, or repairs an ambiguous fragment.

Ask: "Approve these, edit specific ones, or want a different smoothing pass on any character?"

Once calibration is locked, do not dump pairs for the remaining scenes. A full session is many hundreds of quote lines and a wall of pairs gets skimmed, which is worse than no review. The real review of the rest happens through the garble rulings in step 2.5 — those are the changes that can actually go wrong. For each remaining scene, report what you rendered, the count of stage-direction splits and truncations handled, and then go to its rulings.

If the calibration pairs themselves run long (many representative pairs, several decision classes), the same interactive review artifact described under step 2.5 applies here too — one card per pair, grouped by decision class, Approve/Reject/Discuss, download-and-read-back. Use plain chat for a single calibration scene's handful of pairs; reach for the artifact only if that set grows past what a reviewer can hold in view at once.

Apply all edits to the scene_extractions_smoothed/ files only.

5. Verify the layer

Cheap, and each check has caught a real defect:

cd <session-dir>/scene_extractions_smoothed
grep -l "## Verbatim moments" 0*.md | wc -l          # MUST be 0 — see step 3
grep -l "## Voiced moments"   0*.md | wc -l          # MUST equal the scene count
grep -l "^source: voice-smoothed" 0*.md | wc -l      # MUST equal the scene count
grep -c 'truncated' 0*.md | grep -v ':0'             # leftover markers (summary copies are OK)
grep -c '[a-z0-9…"]\*\*\[\?\(GM\|<PCs>\)' 0*.md | grep -v ':0'   # splices you re-introduced

Also confirm the player-name policy actually landed the way the GM ruled it — scan the summary sections and the quote lines separately, since the ruling usually differs between them.

6. Write the manifest — REQUIRED

Rulings made in conversation and applied as scattered annotations are unreconstructable a week later. Write <session-dir>/voice_smooth.sources.yaml recording:

• scene count, quote-line count, and the garble-ruling count broken down by scene
• the resolved voice specs, and any NPC characterization source used
• which transcripts were consulted and what each is good at — this is per-campaign knowledge that makes the next run faster
• the calibration decisions the GM approved (filler aggressiveness, grammar repair, stage-direction split, truncation handling)
• every scope ruling — de-duplication boundaries, player-name policy, knowledge boundaries — with what moved where
• garbles kept deliberately, and why, so a later reader doesn't re-flag them
• glossary rows written back, and any registry/glossary conflict found
• upstream defects found but not fixed, and carry_forward for anything left open

Validate it parses (python3 -c "import yaml; yaml.safe_load(open(...))"). Watch the indentation trap: keys that follow a list at the same indent level get swallowed into the sequence.

7. Hand-off

Note that session_doc should now read from scene_extractions_smoothed/ (the verbatim scene_extractions/ and the VTT remain the record).

Say explicitly that re-running the extractor would discard this pass, and name what is now true only in the smoothed layer — de-duplication, splice repairs, knowledge-boundary annotations. That is the difference between the GM knowing the smoothed layer is authoritative for narration and losing a day's rulings to a routine regeneration.

Conventions

• Verbatim is immutable. Never write to the VTT or scene_extractions/. This skill's only output is scene_extractions_smoothed/.
• Never carry a claim you cannot keep. The output heads its moments section ## Voiced moments, never ## Verbatim moments — see step 3.
• Voice files are authoritative — read them first, preserve the voice, never homogenize. When a player corrects a characterization, the voice file wins; update it there, not here.
• Preserve, don't rewrite. Readability + voice only. Names, numbers, mechanics, attribution, and meaning are off-limits.
• Don't over-smooth. Deliberate style is voice; only transcription noise and genuine unreadability get cleaned. A character who rambles on purpose should still ramble.
• Deciding a name is upstream; spelling a settled one is here. Establishing who someone is is an identity decision belonging to /session-summary-consistency and the registry — never make one in this layer. But applying a form the registry has already settled is spelling, not identity: Utgartian → Uthgardtian, Colin → Cullen, a house like Astra → House Margaster. Check the registry first, say which authority you are applying, and offer the glossary write-back.
• Common-word garble is ruled here. app/hack, confront-free/conflict-free, bloom/show are in no glossary and trip no registry check, so they survive every upstream pass and surface only when a human reads the line. Rule them here with the GM (step 2.5) and record what was ruled.
• Scope belongs to the human. What belongs in which scene, and who knows what, are precision decisions — never settle a de-duplication boundary or a player-vs-character knowledge question yourself (step 1.5).
• Knowledge boundaries need a home outside this layer. An annotation in a derived file dies at the next scene_extract. Push it to a hand-authored docs/ dossier with a CLAUDE.md pointer.
• Human reviews before session_doc. This is a first-draft render, not a final artifact.

Why this design

Smoothing is rendering — exactly what LLMs are good at ("taking verified structure and making it feel alive"). It is safe here because: the raw record is preserved (VTT + verbatim extractions untouched), it writes only a derived copy, each line is guarded by an authoritative voice file, and a human reviews the render before it feeds the next stage. That is the house pattern — LLM extracts → human imposes structure → LLM renders inside it — applied to the quote layer.