The AI Writing Coach That Diagnoses First, Then Teaches

CipherWrite's AI Writing Coach is built on a specific claim about what information cannot supply, and the whole product falls out of that claim. Three things are missing when you learn craft by reading about craft:

1. Diagnosis. Writers reliably misidentify their own problem. The thing you call writer's block is usually not a block.
2. Sequence. Craft has prerequisites. Line-editing a chapter whose structure is broken is wasted work, and revision done out of order costs months.
3. Feedback. Skill comes from attempting something, being told precisely where the attempt failed, and trying again. Reading about a technique produces recognition, not ability.

Key takeaways

• It disagrees with you first. The intake ends in a diagnosis of the root cause under the symptom you reported — not a restatement of what you typed.
• The curriculum is fixed data, not model invention. 32 modules, 13 disciplines, ~25 hours, distilled from 13 long-form guides, each with real attributions and a graded exercise.
• Module IDs reach the model as a closed enum, so a roadmap cannot point at a lesson that does not exist.
• Grading is criterion-by-criterion with quoted evidence — and returns an empty "what works" list rather than manufacturing praise.
• It refuses to draft for you. It demonstrates by rewriting one element in your voice, never the whole submission.

The six screens, and why the fourth one matters

The intake is short: your experience level, what you are writing, what stage you are at, your biggest blocker in your own words, how you work, and how many hours a week you actually have.

Five of those six are structured. The fourth is a free-text box, and it is doing the heavy lifting — because the gap between what a writer reports and what is actually wrong is the entire premise of the tool.

The last screen matters more than it looks, too. A roadmap that assumes ten hours a week when you have three is not an ambitious plan; it is a plan you will abandon, which is the exact failure mode a lot of these writers are already stuck in.

The symptom-to-cause table

This is the part of the page worth reading even if you never open an account. It is the mapping the diagnosis works from:

What you say	What it usually is	Where to start
"I have writer's block."	A structural hole — your scenes stopped changing anything	Sagging middle, midpoint reversal
"My characters feel flat."	No conflict between what they want and what they need	Want vs need
"I start books and never finish them."	The same structural failure recurring at the same point, plus motivation-dependent drafting	Why books die, escalation audit
"My dialogue is boring."	Characters are exchanging information instead of pursuing opposed wants	Dialogue as verbal action
"Readers say it starts slow."	The inciting incident is too late, or the opening poses no dramatic question	Three-act spine
"My writing feels distant."	Filter words — saw, felt, noticed, realised — putting a camera between the reader and the experience	Deep POV, filter audit
"I finished a draft and don't know what to fix."	No reverse outline, so the structure is invisible and undiagnosable	Editorial hierarchy, reverse outline
"It doesn't feel like it means anything."	You have a subject, not a thematic argument — nothing for the climax to prove	Controlling idea

Read the left column and the right column together and the point lands. Every one of those symptoms sends a writer looking for the wrong remedy. "Writer's block" sends people to productivity advice — timers, word-count streaks, morning pages — when the actual problem is that they wrote themselves into a middle where nothing any character does changes the situation. No amount of discipline fixes a scene that has nothing at stake. The block is the symptom; the flat midpoint is the disease.

"I start books and never finish them" is the same story on a longer horizon. It is almost never a character defect, which is how most writers experience it. It is one repeating structural failure that shows up at the same point in every project, plus a drafting method that only works on days you feel like writing.

The catalogue: 32 modules the model did not write

Here is the load-bearing design decision, and the reason this is not a chatbot in a nicer wrapper.

The craft content lives in a fixed catalogue — 32 modules across 13 disciplines, roughly 25 hours of instruction, distilled from the site's 13 long-form guides. Each module carries its concepts, its real attributions, the classic mistakes writers make in that area, and an exercise whose success criteria are the grading rubric. Not a separate rubric written later. The same criteria you are told about up front are the ones your submission is graded against.

Discipline	Modules	Examples
Writing process	2	Why books die; find your working method
Foundations	1	Premise and reader promise
Story structure	6	Three-act spine; 15-beat sheet; sagging middle; climax that lands
Character	4	Want vs need; wound, lie, fatal flaw; arc types; the antagonist who could win the argument
POV and voice	3	Choosing POV; deep POV filter-word audit; multi-POV architecture
Dialogue	2	Dialogue as verbal action; character voice
Subtext	1	Said, unsaid, unsayable
Pacing	2	Open loops; cognitive load
World-building	2	World core; delivering world without info-dump
Theme	2	Controlling idea; dramatize, don't preach
Romance craft	2	Romance beat sheet; trope stacking and the black moment
Revision	4	Editorial hierarchy; reverse outline; line edit (RUE, filters); read-aloud pass
Publishing	1	Submission package

The attributions are real and named: McKee, Campbell, Truby, Weiland, Snyder, Egri, Gardner, Booth, Grice, Leonard, Sanderson, Hayes, Browne & King — plus the cognitive research the pacing and structure modules lean on: Kahneman on cognitive load, Zeigarnik on open loops, Ekman on the physiology of emotion. Every lesson names its sources and links out to the full guide it was distilled from.

The mechanism that makes this hold at runtime: the module IDs are handed to the model as a closed enum. The model's job is to pick and sequence, not to author. It can choose the wrong lesson for you — that is a judgement call, and judgement calls can be wrong — but it cannot route you to a lesson that does not exist, and it cannot quietly substitute invented craft advice for the catalogue. If you have ever watched a general-purpose model confidently attribute a beat sheet to the wrong author, you will understand why this constraint is worth engineering for.

What a step actually contains

The diagnosis produces a roadmap of typically 6–12 steps, grouped into phases — and the grouping matters, because it is what tells you that the line-level modules come after the structural repair, not alongside it.

Each step is four things:

• A lesson — the module's concepts, taught rather than summarised.
• A worked example in your genre. This is a small thing that changes the experience a lot. A midpoint reversal illustrated with a thriller is nearly useless to someone writing a quiet literary family novel, because the mechanism is buried under genre furniture they will never use.
• An exercise on your own book. Not a generic prompt. The work you do in the exercise is work on the manuscript you came in with.
• Graded feedback on what you submit.

Grading, and the empty list

Feedback is where most AI writing tools quietly fail, because the default behaviour of a helpful model is to be encouraging. Encouragement is pleasant and teaches nothing.

Three things about how the grading works:

It grades criterion by criterion, with quoted evidence. Not a paragraph of impressions and a number. Each success criterion from the exercise is assessed separately, and the assessment quotes the line in your submission that earned the verdict. You can argue with a quote. You cannot argue with "this could be stronger."

The "what works" list can come back empty. This is the detail I would point at if I had to pick one. Most feedback systems are built so the praise section always has something in it, which means the praise section carries no information — a compliment that appears unconditionally is noise. Here, if a submission genuinely does not demonstrate the criteria, that list is empty and the feedback says what is missing instead. Praise means something precisely because it is not guaranteed.

It demonstrates by rewriting one element — never the whole thing. One paragraph. One exchange of dialogue. In your voice, applying the technique you just learned, so you can see the difference between your version and the corrected one at close range. Rewriting your whole submission would produce better prose on the page and no learning at all, which is the wrong trade for a tool whose only purpose is to make you better at this.

Three starting points

The roadmap differs sharply depending on where you enter, which is what the stage screen is for:

Never written a book. The first fortnight goes on a premise you can actually defend, one structural framework, and a drafting system with a bad-day floor — a minimum that still counts on days when writing is unpleasant. The floor matters more than the ceiling; motivation-dependent drafting is one of the two named causes of serial abandonment.

Started and stalled. Diagnosis of the stall point first, which is very often the midpoint. Then an escalation audit of the scenes you already have — do they escalate, or do they merely continue? — and one specific rewrite that lets you draft forward again this week, rather than after a restructure that takes a month.

Finished a draft and don't know what to fix. Editorial hierarchy before anything else, then a reverse outline of the manuscript you actually wrote (as opposed to the one you meant to write), then scene-level repair, then line-level passes — in that order. The order is the entire value. Line-editing prose that is about to be cut is the most common way to spend three months and improve nothing.

Versus a general chatbot

The honest comparison, since it is the first question anyone sensible asks:

	General chatbot	The coach
What you get	An answer to whatever you asked	An answer to what is actually wrong, which may not be what you asked
Order	Whatever order you happen to ask in	Sequenced by prerequisite, phase by phase
Source of craft	Generated on the spot, attributions unverified	Fixed catalogue, named sources, closed enum of module IDs
Exercises	If you ask for one	Built into every step, with published success criteria
Grading	Impressions, usually generous	Criterion by criterion, quoted evidence, praise not guaranteed
Memory	Starts cold, or approximates	Persistent book record — later lessons build on earlier decisions

A chatbot is genuinely useful when you already know what you need. The whole problem is that most stuck writers do not.

What it costs, and what "free" means here

A free account runs one plan, and the free tier is unusually complete for what it is: the full diagnosis, the whole roadmap visible rather than three steps and a paywall, step one taught in full — lesson, worked example, exercise — and your first submission graded. That is enough to judge the quality of both the diagnosis and the feedback before paying for anything, which is the only honest test.

Pro unlocks the rest of the roadmap, unlimited submissions and re-submissions, deeper lesson content, and up to 25 simultaneous plans — writers typically run one against the current manuscript and another against a specific weakness they have decided to fix. Pricing is regional and moves with the billing cycle, so it lives on the pricing page rather than in a blog post that would go stale.

What to know before you start

Sign-in is required. A plan stores your book record, the diagnosis and your progress, because every later lesson builds on decisions made in earlier ones. There is no meaningful way to do that anonymously.

Your submissions go to a model. Say this plainly: coaching works by sending your text to an AI for analysis, and submissions are saved to your account so the grading and the roadmap have something to build on. This is the coaching tool. It is not the encrypted vault side of CipherWrite, and the two should not be conflated — if you want the architecture of the private editor, that is a separate feature with a separate design, covered in our CipherWrite review. You can delete a plan whenever you want, and if you would rather not paste specifics, premise-level descriptions are enough to produce a usable roadmap.

It will not write your book. If what you want is drafted prose, this is the wrong tool and will feel obstinate.

Non-fiction gets partial coverage. Structure, theme, scene architecture, pacing, voice and revision transfer to memoir and narrative non-fiction directly. Genre modules appear only when they apply.

Who this is for

Writers who have read the craft books and still cannot finish. Writers on their third abandoned draft who have concluded the problem is discipline — it usually isn't, and repeated abandonment is more fixable than it feels, because it traces to one structural failure recurring at the same point. Writers holding a finished draft they are afraid to open, who need the repair sequence more than they need another technique. And writers who have discovered that unconditional encouragement from a chatbot is worth roughly what it costs.

Not for people who want prose generated. Not for people whose problem is purely organisational — see our roundup of writing apps for authors for that. And not for anyone unwilling to be told their self-diagnosis was wrong, because that is the first thing it does.

Run the diagnosis at cipherwrite.com/ai-writing-coach — the full root-cause read, the complete roadmap and your first graded exercise are free, which is the fastest way to test whether any of the above holds up on your own pages.