Prose Coach

As a document got larger, this script got slower, way slower. Fixed that! Thank goodness, really.

It is better about not grabbing up two separate lines at once.

Refactor pass to cut out some of the buildup over time.

In the config of the script, you can now say how many paragraphs around the text to look. The default is 1, the current behavior, examine the paragraph you're in.
2: current + previous paragraph
3: previous + current + next
4: two previous + current + next
Continues alternating backward, then forward
Maximum: 9 paragraphs
More context, maybe less mistakes, but more tokens/time being used. Weigh the prose and cons.
Pun intended.

Now sends lorebooks along with the context as needed, so maybe less hallucinations.

The coaching lines can now be turned off.
The buttons at the top now wrap automatically.

Oops, quiet crash fixed.

I patched Prose_Coach.naiscript so every api.v1.generate(...) request now goes through a shared generateWithBudgetWait(...) helper.
That helper:
counts uncached input tokens with api.v1.script.countUncachedInputTokens(...)
checks api.v1.script.getAllowedInput() and api.v1.script.getAllowedOutput()
waits with waitForAllowedInput() / waitForAllowedOutput() until both budgets are sufficient
only then calls api.v1.generate(...)
I routed all current generation paths through it:
SDT coaching
SDT cluster coaching
tag coaching
adverb coaching
passive coaching
general coach sub-calls
general revision coaching

Dialogue-boundary handling was reworked across SDT, tags, adverbs, and passive detection so quoted speech is excluded much more aggressively.
The detector pipeline was refactored toward a dialogue-first model:
sentence parsing now preserves offsets and quote state
shared sentence analysis uses source positions and source-line context instead of trying to rediscover sentence placement from text matches
quote-less dialogue continuation fragments are now filtered by parser state plus line-level quote parity
The old repeated per-detector rescanning was consolidated into a shared analysis pass, which improved refresh speed.
Manual Refresh was restructured:
it now returns quickly and runs heavy work on a timer tick
it shows in-progress feedback immediately
the general coach step is deferred so detector results land sooner
Refresh UI feedback was added:
button state changes while running
immediate Refreshing analysis... toast
completion and failure toasts remain
In short: 1.6.52 is the dialogue-leak reduction and refresh-responsiveness build.

Prodded it to be better still about dialogues, and a bit smarter about coaching.

Finish hunting down where dialogue leaks into detection.

Fix fix. Stop detecting speech, dang it.

It's better at tracking dialogue and thus less false positives.

More soft targets for Adverbs and SDT tags. They won't immediately trigger a squeal, but it's more ways for it to scan for and notice when they're bunching up.

Looks for more SDT tags.

• Added a shared stripBBCode() helper.
• splitIntoSentences() now sentence-splits on BBCode-stripped text, so tags like [i]...[/i] stop confusing boundary detection.
• getNarration() now strips BBCode before it decides what counts as narration vs dialogue.

The meaningful changes are:

• detectTagFlags() now recognizes narration sentences adjacent to dialogue, so patterns like "Leave." He sighed. no longer get skipped.
• coachGeneral() now actually honors general_prompt config before falling back to the hardcoded three-call behavior.
• General-coach generation now uses state.config.temperature instead of a hardcoded 1.0.
• The top-level behavior comment now matches reality: the script can edit the document when replace features are used.

• Sentence splitting no longer flattens paragraphs together.
• Show-don’t-tell clusters cannot group flags across \n.
• Cluster rewrite prompts cannot pull context past a linebreak.
• General “specific line to strengthen” rewrite now receives only one selected line of context, not the whole analysis window.

Now also looks for passive voice. It tries to be more confident before bothering you here since it really isn't sure. Will not be done automatically ever.

Now looks for adverbs (-ly words, quickly, strongly, etc.) and advises how to strengthen the verb/words instead of using them. Allows replace, auto replace, retry and all the other goodies already present.

Bug fix, the auto buttons weren't actually usable... oops.

Okay, so, you like the changes. What if you just want them without the clicking? Turn on auto replace! Enjoy the problems just erasing themselves quietly. Will pause and wait if quota runs out, but will get it done eventually.

Now it clusters problems if close together, and can replace them all too! In theory. Enjoy!

Now adds scores/priorities to what it finds, low, medium, high, so you know what you're facing. Should have less false-positives.

There is now a handy auto-scan button to make it stop auto-scanning without visiting any option pages.

Bugfixes and such.

Behavior changed in the replace path:

If the entire replaced range was already marked ai, the full replacement is now marked ai.
If the replaced range was human/prompt or mixed, the existing diff behavior remains: carried-over words stay human/prompt, newly introduced words are marked ai.

When using the replace action, any original text the user has made that is used again in the generated text will be marked as 'edited' instead of 'ai', since you made that, not the AI. We got your back.

When highlighting text and asking for advice, you can now replace and also retry what you get for more happiness!

On a failed generation, get a retry button.

1.6.7 -- You can now RETRY anywhere you're thinking about 'is this quite right to put in?'

You can now click the advice part of the coach and get more pointed advice(with find/replace button if you like it!)

The varied coach actually works now, oops.

How about a bigger general coach? That's what your getting. How the flow is, where weaknesses are seen, and a general look at the overall story.

It's better at detecting dialogue and the many sentences that live within and thus should not be scanned for filler.

detectTagFlags was scanning the entire raw sentence for tag verbs, including everything inside the quotes. So "Miss Octavia said I needed more practice" — the said inside the reported speech — matched EXPLANATORY_TAG_VERBS and triggered a false flag.

The fix applies the same getNarration() stripping that SDT detection already used: extract only the non-dialogue portions of the sentence, then scan those for tag verbs. A real attribution tag like he said, shrugging will still get caught because said and shrugged appear in the narration outside the quotes. But said buried inside quoted speech is now invisible to the detector.

When you replace, it will update the changed text as AI generated, since it will be.

It now offers a direct replace button if you want to go ahead and replace with the found answer.

Selection extraction — was using sectionId >= fromId && sectionId <= toId which only works if IDs happen to be sequential. Now uses findIndex to locate the from/to sections by identity in the document-ordered array, then iterates by array position. Works regardless of how NovelAI assigns IDs.

Prompt framing — context menu selections now get the original multi-paragraph prompt ("identify 1-2 specific sentences...") with 500 token budget, while auto-detected single-sentence flags keep the targeted single-sentence prompt with 300 tokens. Previously both used the single-sentence prompt, which is why multi-paragraph selections were getting confused responses.

Oops, bugfix.

Three categories now in the SDT panel, each with its own icon and prompt:

🔭 filter verb — perception routed through awareness (felt, heard, realized) — prompt pushes toward direct sensory experience
📢 telling word — evaluative adjectives (terrifying, miserable) — prompt pushes toward specific evocative detail
⚡ reaction tell — named reactions (yelped, flinched, relief, dread) — prompt pushes toward body language and physical sensation without naming the reaction

When making a suggestion, it won't suggest the dang word that got it flagged in the first place.

Initial release: Prompt used for the lightweight General Coach observation.