| name | writing-reader-feedback |
| description | Simulate a specified reader persona reading an article and report raw reading-experience feedback, not writing advice. Use when the user asks how a specific audience would experience a draft, whether readers can understand it, or wants reader simulation, reader-perspective feedback, or reader role-play. Near miss: for writing improvement, draft polishing, structural optimization, or a revision plan, use writing-blog. |
Writing Reader Feedback
Goal
Simulate a real reader reading an article section by section, with the background and knowledge state specified by the user, and record what happens in that reader's mind. The deliverable is a reading-experience report, not writing advice.
Use the user's primary language for the report unless they ask for another language.
Success Criteria And Stop Condition
Success means every reading unit is covered in article order, each recording the reader's expectation, confusion, emotion, and motivation to continue at that point, ending with an overall impression.
Stop after the full article has been read and the overall impression has been reported. Do not turn the output into writing advice, a revision plan, fact checking, or polishing unless the user separately asks for that.
Evidence And Reading Boundaries
- If the user provides article text directly, use that text.
- If the user provides a file path, read only the target file.
- If the user provides a webpage URL, use
url-reader or WebFetch to get the body content. Retry only when the body is missing, truncated, or obviously not the target article.
- Do not search external sources to fill in missing background for the reader. Missing reader background should appear as confusion, misunderstanding, or lack of interest, not be silently repaired.
- If the article cannot be read, state the blocker and ask the user for the text or a usable source.
Workflow
1. Establish The Reader Persona
Use reader-definition sources in this priority order:
- Reader background specified by the user in the conversation.
- The article frontmatter
audience field.
- A conservative inference from the article context.
Ask one narrow question only when missing reader information would materially change the feedback. Otherwise state the assumption and continue.
Write the Reader Model block (see Output Shape) before reading any body text. Its Knows and Does not know lists are the baseline for the whole simulation.
2. Split The Article
First scan only the heading structure. Do not read the full body yet. Find the smallest heading level present; if the article has both h2 and h3 headings, split by h3.
Splitting rules:
- Split into reading units by the smallest heading level.
- Treat the opening text before the first heading as its own unit.
- If a section is longer than 15 lines, split it further by paragraph.
Record the section list, then begin reading section by section.
3. Read Section By Section
Read one section at a time. Read the current line range, immediately write that section's reader feedback using the dimensions in Output Shape, then move to the next section.
Do not use later content to repair earlier experience. If you read the whole article at once, you already know what comes later and can no longer honestly report "at this point I expected the next section to explain X." If the full article is already in context, still simulate first-pass reading: for each section, only use what the reader could know, guess, misunderstand, or expect at that point.
4. Output Shape
Start with the reader model:
## Reader Model
- Background: <field, technical level>
- Topic familiarity: <completely new | has heard of it | has practical experience>
- Reading scenario: <opened from a feed | searched in a technical community | sent by a friend | other>
- Knows:
- Does not know:
For each section, use this shape:
## <section title or "Opening"> (L<start>-L<end>)
<sentence-level reader reactions>
Record these for every section, writing "none" where an axis is genuinely quiet:
- Expectation: what I expect the next section to cover.
- Expectation shift: what the section actually did and how it differed.
- Confusion: words or concepts I do not understand.
- Lack of interest: why I do not currently want to understand this part.
- Emotion: curiosity, agreement, doubt, impatience, irritation, excitement, or wanting to close the page.
Record these when they occur:
- Missing context: the text assumes I know X, but I do not.
- Missing subject: who is acting here - a person, agent, system, or something else?
- Cognitive load: the information is dense enough that I need to pause.
- Look-back impulse: what I want to reread.
- Active guessing: what I have to guess because the article did not explain it.
- Visual expectation: where I expect a diagram, example, or code.
Within each section, record sentence-level reactions for sentences that create a real cognitive event. Smooth sentences can be noted briefly or skipped. Focus on friction.
5. Overall Impression
After finishing the article, add an overall impression:
- Main value: what I remember after reading.
- Main confusion: what remains unresolved.
- Expectation management: where the article helped me form the right expectation, or whether I kept revising my expectation throughout.
- Information gaps: what I still need in order to fully understand the article.
Discipline
Stay faithful to the reader persona. It is better to be too confused than too understanding. You, as the AI, may know a lot, but the simulated reader may not. If the article does not explain a concept and the reader background does not include it, the reader does not understand it. Do not complete the author's logic silently. If the reader has to guess, that guess is friction.
Use the reader's plain voice, not an analyst's voice. The feedback should be direct, colloquial, and sometimes blunt:
- "I do not understand this."
- "I do not care about this right now."
- "Wasn't the previous section about X? Why are we suddenly on Y?"
- "So what?"
- "This is getting annoying."
A reader can be impatient, irritated, or ready to close the page. If the whole report sounds polite and analytical, it is no longer a reader simulation; it is a review.
"I do not care" matters more than "I do not understand." Sometimes the problem is not comprehension but motivation: the setup is weak, confusion has built up, or the article moves into details before the high-level shape is clear. Record that directly.
Expectation tracking is the core capability. Divergence between what the reader expected and what the section delivered reveals structural problems more clearly than wording comments.