Most AI writing tools begin with the same object: an empty chat box. That is useful when the job is a conversation. It is a poor default when the job is to change three lines, check continuity across an act, or tighten a scene without losing the reveal.
The important difference is not whether AI appears beside a screenplay. It is whether the AI can work with the screenplay as it is written—and return the result in a form a writer can actually judge.
Chat-style AI
Inkwell skill pass
Every suggestion is tied to a line in your draft — it lands exactly where it belongs.
The window problemLink to this section
Open a general chat assistant in another tab and you become the context window.
You paste the scene. The model does not know what came before it, so you paste the act. It suggests a line for a character whose name you changed in the second draft, because the version you pasted was the version you happened to have selected. You revise the page. Now the conversation is about a draft that no longer exists, and the next answer is confidently wrong in a way that reads as confident.
None of that is a failure of the model. It is a failure of the arrangement. A chat window in a different application has exactly the screenplay you last pasted into it, frozen at the moment you pasted it, with no way to tell that the scene has moved on.
A chat sidebar that lives in the draft is a different arrangement. It reads the current scene, the character's earlier dialogue, and what the screenplay has already established—because it is reading the screenplay rather than a transcript of one. When you accept a change, the next question is asked against the changed page. You never assemble the context, and you never re-paste it.
That is why "AI in your editor" is not a layout decision. It is the difference between an assistant that knows what the draft says and one that knows what you last showed it.
Chat is good at questionsLink to this section
Chat gives you room to think out loud. Ask why the midpoint feels early, whether a character is avoiding the obvious choice, or what the final image is promising. A useful answer may change how you see the scene even when it changes none of the words.
That is why Inkwell includes Ask. The conversation is grounded in the draft, but it remains a conversation. You do not have to pretend every thought is an editing command.
The trouble begins when chat is asked to stand in for every other writing action. "Make this colder" comes back as a paragraph or a replacement scene. You compare it against the page, find the useful fragment, paste it back, repair the formatting, and then check what else changed. The writer is left doing the copy-paste and cleanup.
One scene, five different asksLink to this section
A concrete case. INT. DINER — LAST LIGHT. Maya arrives with an envelope; the reveal is what is inside it. The scene runs long and the entrance gives away too much.
Depending on what you already know about the problem, five different things are the right move:
- You are mid-line and the phrase is nearly there. Ghost text offers the rest of it. One key takes it, one key dismisses it.
- You cannot name what is wrong. Ask: "why does the entrance deflate the reveal?" You get a read, not a rewrite, and you decide what to do with it.
- You know exactly which sentence is the problem. Edit selection on that sentence: "make her arrival wordless." Alternatives appear against the original words.
- You suspect the problem is not local—Maya has been over-explaining for twenty pages. A dialogue pass over the act returns proposed changes on the lines it touches.
- You know the outcome you want and not the path. Agent Mode: "tighten the diner scene without losing the reveal." It reads, changes the entrance, reads the result against the ending, and decides whether another move helps.
Same scene, same writer, five reaches. Nothing about the screenplay changes between them; what changes is how much of the judgment you are keeping.
What you review is the real costLink to this section
Every reach buys you something and charges you something. The charge is review.
- Ghost text returns the rest of the phrase. You review the words already in front of you, as you type.
- Ask returns an answer in the conversation. There is nothing in the draft to review, because it changed nothing.
- Edit selection returns up to three alternatives. You review the selected passage, in place.
- A skill returns findings, proposed changes, or both. You review each suggestion, or the evidence behind each finding.
- Agent Mode returns a worked result on a separate draft line. You review the finished run before any of it comes back.
Read that list downward and the choice gets easier. The cheapest thing to review is a suggestion you can dismiss with a keystroke. The most expensive is a multi-step run you have to read as a whole. Reaching higher than the job needs does not get you a better draft; it gets you a longer review for the same three lines.
What a skill addsLink to this section
A skill is a specialist with a defined job: dialogue, continuity, coverage, pacing, character voice, period language, or another repeatable pass. The definition matters because "check continuity" should not quietly turn into "rewrite the scene in a voice I prefer."
Some skills are editorial. They propose changes on the lines they touch. Others are analytical. They return findings without rewriting the page. Both are more useful when the result points back to the scene, character, or line that earned it.
For an edit-producing pass, you can accept or reject individual suggestions, take them all, or keep the original. For a report, you can read the evidence and decide what—if anything—you want to change. The skill does the focused read; it does not make the final craft decision.
What Inline Edit addsLink to this section
Inline Edit is for the moment when you already know where the problem is. Select the sentence, describe the change, and compare alternatives where the sentence lives. There is no reason to launch a full dialogue pass because one parenthetical is too blunt, and no reason to paste three options out of a sidebar just to see them against the surrounding exchange.
This is a small distinction with a large effect: the selected passage stays at the center of the work.
What Agent Mode addsLink to this section
Sometimes the target is clear but the next move is not. "Tighten the diner scene without losing the reveal" may require reading the entrance, changing a line, checking the new beat against the ending, and deciding whether another pass is useful.
Agent Mode works through that sequence. It can make an exact change, revise more broadly, call a specialist skill, or pause for your judgment. After each move it reads the result again. The goal stays fixed while the next move responds to the scene.
That is the distinction between a skill and an agent: a skill performs a known craft job; Agent Mode chooses the next move as the scene develops and can bring in a specialist when one helps.
When to reach lowerLink to this section
The honest cases against the larger tools are worth stating, because a tool that never says "not here" is not giving you information.
Do not give Agent Mode a goal you have not decided on. "Make Act 2 better" is not a goal; it is a mood. The agent will do something coherent with it, and you will spend longer working out whether you wanted that than you would have spent naming the problem first.
Do not run a craft pass to find a problem you could read for yourself. A continuity check across 110 pages is the right tool when you genuinely do not know whether the prop survived the rewrite. It is the wrong tool for a scene you finished ten minutes ago.
And do not reach for any of it in the middle of a first draft that is going well. The best AI decision available is often not to make one.
Keep the review where the decision happensLink to this section
Writers do not only need generated text. They need to know what moved, why it moved, and whether the new version belongs.
Ghost text can disappear with a key. Inline alternatives sit against the selected words. Skill suggestions arrive as reviewable changes, while analytical passes return findings. Agent Mode works on a separate draft line and leaves save points after landed edits. The original draft remains intact until the writer chooses what to bring back.
That is the product bet: conversation when the job is conversation; an inline edit when the words are already in front of you; a specialist when the craft pass is known; and an agent when the goal needs more than one move.
If you want the longer case for the top of that range, Ask or Agent: when to hand over the pen covers where the boundary between a question and a goal actually falls.