The direct answer
Filter words are small verbs and hedges that report an experience instead of delivering it. Saw, heard, felt, wondered, seemed, just, really, and very sit between the reader and the scene, so the reader watches the character perceive rather than perceiving alongside them. Paste 100 to 300 words of your draft into the finder above and it counts those words across five classes, reports density per 100 words, quotes back your five worst sentences with class labels, and hands you a rewrite direction for each class.
What filter words do to a scene
A filter word inserts the narrator as a middleman. The sentence says somebody saw, heard, felt, or thought something, and the reader must process that report before reaching the thing itself. One such verb in a paragraph costs almost nothing. Ten of them in a page builds a pane of glass across the whole scene, and the reader feels the story happening at arm's length without knowing why.
The mechanism is attention economics. Every reporting verb asks the reader to spend attention on the act of perceiving rather than on the perceived detail. The gulls, the cold air, the slipping rope — these are the images that create a scene. Felt the cold air spends two words telling the reader that a sensation happened, then still needs the cold air to do the real work. Drop the reporting verb and the same words land harder, because nothing stands in front of them.
Distance is not always wrong. A narrator who does not understand what they are seeing needs reporting verbs to show that confusion honestly. A child viewpoint, a shocked witness, or a character lying to themselves may genuinely process the world through seeming and almost-knowing. The goal of a filter-word pass is never zero. The goal is intention: every remaining filter earns its place, and every removed one hands its words back to the image.
The five classes this tool tracks
The finder groups filter words into five classes so you can see which habit dominates your draft. Each class weakens prose in a different way, and each needs a different fix.
Seeing covers verbs of visual perception: saw, seen, look, looked, watch, watched, notice, noticed, observe, glance, stare, glimpse, and their relatives. Seeing filters are the most common in the drafts writers paste here, because pointing the viewpoint character's eyes at things feels like good viewpoint discipline. It is, once. The fifth looked of the page adds nothing the reader did not already assume.
Hearing covers verbs of auditory perception: hear, heard, listen, listened, listening, sound, sounded. Hearing filters cluster in outdoor scenes, parties, and anywhere the writer wants atmosphere. The pattern to watch is the double report, where the character both listens and then hears, spending two verbs on one noise.
Feeling covers verbs of physical and emotional sensation: feel, felt, touch, touched. Feeling filters dominate love scenes, grief scenes, and injury scenes — exactly the passages where direct sensation matters most. Felt his chest tighten names a real symptom, which is why feeling filters survive revision more often than others. Keep the ones attached to concrete physical detail and cut the ones attached to abstractions.
Thinking covers verbs of cognition: think, thought, wonder, wondered, realize, realized, know, knew, remember, remembered, decide, decided, consider, believe, understand. Thinking filters are the novelist's special trap, because novels trade in interiority and these verbs feel like interiority. Often they are its opposite: a label on a thought the scene never dramatizes. She wondered whether he would come back is a placeholder for the pacing, the door-checking, the unsent message — the behaviors that would show the wondering without naming it.
Hedging and intensifiers cover the words that soften or inflate without informing: seem, seemed, appear, just, really, very, quite, rather, somewhat, suddenly, almost, nearly, actually, basically, perhaps, maybe, started, began. This class splits into two habits. Hedges like seemed, almost, and perhaps blur the sentence so the writer cannot be wrong. Intensifiers like really, very, and suddenly shout so the sentence feels strong without being specific. Both dissolve under the same fix: delete the word and check whether the sentence survives. It almost always does.
How the scoring works, with worked numbers
The tool runs a pure word-count analysis on exactly the text you paste. Nothing leaves your browser, no model rewrites anything, and the same passage always produces the same result. Understanding the arithmetic helps you read the verdict correctly.
First it lowercases your sample and splits it into word tokens, so Saw and saw count the same and punctuation never creates false matches. Then it checks every token against five fixed word lists, one per class, and tallies the hits. A 160-word sample containing 18 filter words therefore has 18 total hits distributed across whichever classes matched — perhaps 6 seeing, 2 hearing, 4 feeling, 3 thinking, and 3 hedging.
Density is total hits divided by total words, multiplied by 100. Those 18 hits in 160 words give 11.25, reported as 11.3 per 100 words. Reporting per 100 words lets you compare a 120-word flash passage against a 280-word scene fairly, which raw counts cannot do. A 300-word scene with 12 hits scores 4.0 per 100, milder than a 100-word paragraph with 8 hits at 8.0 per 100, even though the longer scene contains more filter words in absolute terms.
Next the tool splits your sample into sentences at period, question, and exclamation boundaries, scores each sentence by its own hit count, and quotes back the five worst with their class labels attached. This sentence ranking is where the practical value lives. A scene can score a moderate overall density while hiding two terrible sentences that carry half the hits, and those two sentences are usually fixable in minutes.
The verdict bands are deliberately plain. Under 3 hits per 100 words reads direct, and the tool says so. From 3 up to 6 per 100, the passage earns a warning and a pointer at the flagged sentences. At 6 or more per 100, the verdict calls for a rewrite pass starting with the top class. Samples under 20 words get an information notice instead of a verdict, because density math on a single sentence misleads more than it helps.
Try the worked example yourself. Paste the default sample and you will see roughly 160 words, mid-twenties hits, and a density in the mid-teens per 100 — a failing score produced by stacked suddenly, really, just, seemed, and felt reports in nearly every sentence. Then paste a cleaned version of the same scene with the reporting verbs removed and watch the same story score near zero. The story did not change. Only the distance changed.
Reading your density result
Treat the density number as a triage signal, not a grade. Its job is to tell you where to spend the next twenty minutes.
A pass result means the passage reports perception sparingly. Read the class table anyway before moving on. A passing overall score can still hide a cluster — four thinking hits in one paragraph of an otherwise clean scene — and that paragraph is tomorrow's revision target even though today needs nothing.
A warning result means the scene works but drags in places. Start with the quoted sentences in order, because the tool already sorted them by severity. Fix sentence one completely before touching sentence two. Writers who dabble across all five flagged sentences tend to halve every problem and solve none, while writers who finish the worst sentence often discover the second-worst reads better once its neighbor stops leaking.
A fail result means filtering shapes the reader's whole experience of the passage. Do not line-edit yet. Read the class table, identify the top class, and apply that single class direction across the full sample first. One-class passes beat scattered editing because each class has one fix, and repeating one fix down the page builds speed. After the top class is handled, re-paste and let the tool re-rank what remains.
An information result means the sample was too short. The fix is mechanical: paste a full scene of 100 to 300 words and run the scan again. Judging a single paragraph wastes a cycle, since one filter word in twelve words scores above 8 per 100 and proves nothing about your habits.
Rewrite directions for each class
The tool gives directions rather than rewritten sentences on purpose. A generated rewrite would teach you nothing and would flatten your voice toward average. A direction teaches a move you can reuse for the rest of the book. Here is each move explained.
For seeing, cut the watching verb and state what was there. She saw the lights across the water becomes The lights across the water, followed by whatever the lights do next. When the watcher matters — a lookout straining, a witness who should have been watching elsewhere — keep one seeing verb and make it carry that meaning. He watched the door means vigilance only if the scene has established why the door matters; otherwise the door simply opens.
For hearing, cut the listening verb and render the noise itself. She heard the gulls crying becomes Gulls cried over the dock. Sounds earn space through texture: the pitch, the rhythm, the interruption they cause. A character who listens hard is a character doing something, so keep the verb when the effort is the point — He held his breath and listened carries tension that He heard a floorboard does not.
For feeling, replace the felt-report with the physical detail. She felt the cold air on her face becomes Cold air burned her cheeks, or whatever the cold actually did. Abstract feelings need this move most: felt afraid becomes the throat closing, the hands checking the lock twice, the sentence the character cannot finish. Keep felt where the sensation itself is strange or medically precise and the strangeness is the point.
For thinking, dramatize the thought as action, speech, or choice. She wondered whether the boat would come back becomes She checked the horizon every hour until dark, or a line of dialogue asking the harbormaster. Realized deserves special suspicion: a realization is a scene event, and the sentence around it should show the trigger — what the character saw or heard that changed their mind — rather than announcing the change. Keep thinking verbs where the thought is genuinely hidden from everyone in the scene, including the reader, and the hiding drives the plot.
For hedging and intensifiers, delete the word and test the sentence. It seemed empty becomes It was empty, or better, the specific emptiness: No boats, no gulls, the fuel dock chained. Really and very always delete cleanly; if the sentence then feels weak, the weakness was always there and needs a stronger noun or verb, not a louder adverb. Suddenly deserves a note of its own: it promises surprise while usually marking the exact moment the writer surprised themselves. Either stage the surprise properly or let the event startle on its own. Almost and nearly state that something did not happen; say what happened instead. Started to and began to add a preparatory phase no reader needs; characters rarely begin to stand when they can stand.
Hypothetical example 1 (illustrative): a dock at dawn
Consider an illustrative hypothetical passage. A writer drafts this: Mara suddenly saw the harbor lights flicker and just stood there listening. She really felt the cold biting her fingers and heard a gull crying somewhere above. She wondered if Tomas would bring the boat back before the tide turned, though it seemed very unlikely.
The finder would flag nearly every sentence. Suddenly, saw, just, listening, really, felt, heard, wondered, seemed, very — ten hits in about fifty words, a density near 20 per 100, with seeing, hearing, feeling, thinking, and hedging all represented. The scene underneath is fine: cold, a flickering light, a missing boat, a turning tide. Every image the reader needs is present. Each one just arrives wrapped in a report.
Apply the directions one class at a time. Seeing first: the harbor lights flicker, full stop, and Mara stands in the cold. Hearing next: the gull cries above, unattended. Feeling next: the cold bites her fingers — the verb bites already does the feeling's work. Thinking next: give the wondering a behavior, such as Mara counting the empty moorings. Hedging last: delete suddenly, just, really, and very, then check each sentence standing alone. A plausible result: The harbor lights flickered. Mara stood in the cold and counted the empty moorings. A gull cried overhead. The cold bit through her gloves. Tomas was late, and the tide had nearly turned. Nearly survives there because the tide's exact state is information, not blur — which shows the real standard. Keep the filter that carries facts; cut the filters that carry only the act of noticing.
Hypothetical example 2 (illustrative): a kitchen-table argument
Take a second hypothetical illustration, this time interior. A writer drafts this: Jon felt his jaw tighten and knew the evening was about to go badly. He remembered the promise he had made in spring and realized he had already broken it. Lisa looked at him across the table and seemed to be deciding whether to speak. He just sat there, listening to the clock.
The classes here are feeling, thinking, seeing, and hedging, and the density again lands deep in failing territory. Yet this passage is closer to working than the dock scene, because two of its filters guard real material. Felt his jaw tighten names a concrete physical event. Seemed to be deciding captures genuine uncertainty between two people — Jon cannot know Lisa's mind, so a flat statement would be false.
A careful pass therefore keeps the jaw and keeps the seeming, while converting the rest. Knew the evening would go badly becomes evidence: the set-down glass, the folded arms, the sentence Jon rehearses and discards. Remembered the spring promise becomes a flash of its content — the words spoken, the kitchen they were standing in — rather than the announcement of remembering. Realized he had broken it becomes the moment of proof, the date on the letter or the name on the roster. Just sat there listening becomes what the clock actually sounds like in that silence, or the thing Jon does with his hands. Notice what changed: nothing in the story, only the ratio of labels to evidence. The revised passage runs slightly longer, and that length is earned, because behavior and detail replaced announcements.
Genre-specific guidance
Filter tolerance shifts by genre, because each genre asks perception to do different work. Use the genre selector on the tool, then read this section for what to protect.
Literary fiction can carry the most interiority, and thinking filters often do thematic work there. A remembered promise or a considered refusal may be the story's true event. Spend the filter budget on two or three load-bearing thoughts per scene, and pay for them by stripping seeing and hedging around them. Literary readers forgive a wondering narrator; they punish a narrator who seems, almost, and really hedges every perception.
Thrillers and suspense run on pace, and thinking verbs are pace-killers even when the plot races. A protagonist who wonders, realizes, and decides on every page slows time to the speed of reflection. Convert thoughts to procedure: checks, calls, movements, tests. Keep seeing verbs lean too — one visual per beat, the knife, the exit, the phone screen — and delete intensifiers ruthlessly, since a thriller that says suddenly every paragraph trains the reader to stop flinching.
Mystery and crime resemble thrillers but with one exception: observed clues earn their seeing verbs. The detective's noticing is the plot mechanism, so one noticed detail per clue is honest craftsmanship. The rule is one verb per clue, placed exactly where the clue lands, with repeats cut. Hedging needs discipline in deductions: seemed and perhaps belong in dialogue where the detective hedges aloud, not in narration where the writer hedges the reveal.
Romance needs felt emotion on the page, and readers arrive expecting interior sensation. Protect feeling hits at turning points — first touch, rupture, reconciliation — and make them physical rather than abstract. Around those peaks, cut seeing and hedging hard, because a love scene filtered through looked, noticed, seemed, and just reads cautious, which is the opposite of its intent. Thinking verbs in romance often mark withheld communication; convert half of them into things said badly or left unsaid at the wrong moment.
Fantasy and science fiction spend words on worlds, and seeing verbs do legitimate delivery work when the setting is new. The first description of a strange harbor earns its looked and watched. The twentieth does not. Hedging is dangerous here because invented worlds need authority; seemed and almost make magic read as the writer's uncertainty rather than the world's rules.
Edge cases and failure modes
Every deterministic checker has blind spots. Knowing them keeps you from following the tool off a cliff.
Dialogue inflates counts innocently. People hedge aloud, report feelings, and narrate their thoughts in speech — that is characterization, not filtering. When your sample includes heavy dialogue, expect hedging and thinking hits from the spoken lines, fix the narration first, and only then decide whether a character's verbal tics need trimming. Never smooth all characters into the same direct speech; a nervous witness who says just and really every sentence is doing voice work.
First-person narration scores higher by design. I saw, I heard, I felt, and I knew are the grammar of a person telling their own story. Compare first-person scenes against each other rather than against an absolute band, and re-paste after revisions to confirm the trend. The failure mode is sanding away the narrator's personality in pursuit of a number — a first-person voice with zero filters often reads robotic, not clean.
Common verbs create false positives. See appears in see you soon, know appears in idioms, think appears in fixed phrases. The word lists cannot read context, only tokens. When the tool flags a sentence whose matched word is clearly innocent, ignore that hit and move on. The honest response to a false positive is dismissal, not contortion of a fine sentence.
Short samples mislead in both directions. A single clean paragraph can score zero while the chapter around it filters heavily, and a single dense paragraph can fail while the chapter reads fine. The tool asks for 100 to 300 words precisely to smooth these swings. Score whole scenes, not lines, and never revise a chapter on the strength of one paragraph's number.
Overcorrection is the most expensive failure. Writers fresh from their first filter lesson sometimes delete every reporting verb and produce passages where perceptions arrive unattributed and the viewpoint floats free. If three unfiltered observations in a row could belong to any character present, restore one seeing or thinking verb to anchor the viewpoint. Clarity outranks purity.
Questions writers ask after their first scan
This section goes deeper on the practical questions that surface once writers start revising. It approaches from the revision desk rather than repeating the tool's own answers.
How much of a scene should change in one pass? Roughly a third of the flagged sentences, fully, beats touching every sentence lightly. Fully means the filter is gone and the replacement detail is on the page, not planned. Writers who fix two of five quoted sentences completely usually find the third reads acceptably once its neighbors carry more weight.
How do filters interact with psychic distance? They are its main lever. Close psychic distance narrates from inside the character — cold bit her fingers — while distant narration reports from outside — she felt the cold. Varying that distance deliberately is advanced craft: open a scene distant, slide close for the turn, pull back for the transition. The finder cannot see distance, only tokens, so a scene that moves through distances may score moderately while being perfectly controlled. Read the flagged sentences with distance in mind before cutting.
Can a filter word ever be the best word? Yes, in three narrow cases. Comic voice uses just and really as timing instruments. Unreliable narrators use seemed to smuggle lies past the reader fairly. Idiom and period voice sometimes require the supposedly weak word — a 1940s detective who never says seemed stops sounding like himself. In each case the test is the same: the word performs a job no replacement performs. If it does, keep it and spend the savings elsewhere.
Doing the revision where your draft lives
Directions only help once applied to real pages. Do the actual rewriting inside Final Edit at /final-edit, where your full manuscript sits ready for line-level work. Open each flagged sentence there, apply one class direction at a time, and keep your voice intact by choosing your own replacement details. When a scene is done, paste the revised passage back into this finder and confirm the density dropped before moving to the next scene.
A simple cleanup workflow
Work in this order and the job stays small. Scan one scene and note the top class. Fix the five quoted sentences using that class direction first. Re-paste and confirm the density moved. Then read the scene aloud from the top: any remaining filter you stumble over is a candidate, and any you glide past has earned its place.