Reader Panel

Reader Panel

Overview

Run structured persona reads of a chapter range and write each one as a feedback file in the shape feedback-triage already reads, so the existing triage flow takes over unchanged. The panel is a first read before human readers see the draft: it finds the problems a reader would trip on, so the human round spends its attention on what only people can tell you.

Every panel read is simulated. It is written by the agent, marked source: simulated, and never presented as feedback from a real reader.

Prerequisites

A story project (story.md in the root) with the chapters in range drafted. The feedback-triage skill must be available for the synthesis.

When to Use

• Before the first beta round, to catch the obvious problems cheaply
• After a revision, to check a fix landed before sending it to people
• When the user has no readers yet and wants a structured first read
• NOT as a substitute for human readers. A simulated round cannot close a book as ready for submission or publication
• NOT for a sensitivity or authenticity read. The sensitivity persona only flags passages for a paid human reader (use editorial-review)

Personas

Each persona has a reference file listing what it reads for, what it never comments on, and how it rates severity. Load only the personas in the panel.

Workflow

1. Scope the panel

1. Ask the user for the chapter range (default: every drafted chapter) and which personas to run (default: all five). Take genre, form, POV, and tense from the story context output in step 2 (its Story essentials section), not from story.md, whose Synopsis may describe the ending. The genre reader needs the genre. Take language only from the Language contract in that same story context output: it is no spoiler, and every persona reads the book as a reader of that language would.

2. Pick the round number: the next N that is free under feedback/ and not yet the name of a tag or snapshot panel-round-{N} (git tag --list 'panel-round-*' and story snapshot --list --path . list them). A panel gets its own round. Never add simulated reads to a round of human readers, so the human synthesis stays independent.

3. Save the text the panel will cite under the name panel-round-{N}, so feedback-triage can map its labels to a later draft with story compare once the chapters change:

   • Git project: work from the book's folder, the one that holds story.md (cd there first), because -- . below means the current folder. Check that .gitignore lists dist/ (add the line if it is missing), so earlier review copies stay out of the commit. Then run git status --untracked-files=all -- . and show the user what it lists: the copy is built from the working tree, but a tag points at the last commit, so uncommitted changes would make the two differ. Look through it for private files (a .env, keys or credentials, scanned documents): unless the user says to commit one, add it to .gitignore first. .gitignore does not untrack a file git already tracks, so also run git ls-files -- . and look for the same kinds of file: if one is listed, the commit below would include its changes, so stop and ask the user before committing. A tag holds only tracked files, so run git ls-files --others --ignored --exclude-standard -- . as well: if it lists a markdown file outside dist/ and .snapshots/, or the cover or stylesheet story.md names, the build reads a file the tag would miss, so take a snapshot as below instead. Ask before committing and before tagging. With approval, commit the project folder only (skip the commit when the tree is already clean) and tag that commit:

     git add -A -- .
     git commit -m "Simulated reader panel round {N}" -- .
     git tag panel-round-{N}

     If the user declines the commit, never tag the last commit over an uncommitted tree: take a snapshot as below, or stop. If git tag says the tag already exists, an earlier attempt saved this round: never move it; ask the user, or pick the next free N. Never push, and never move or delete a tag, without the user's approval.

   • Project without git, the user declined the commit, or an ignored file the build reads: take a snapshot with story snapshot panel-round-{N} --path ..

4. Build the review copy the panel cites from that text, so its labels match the ones human readers will use:

   story build . --format html --stamp panel-round-{N}

   The build holds the whole book, so use it for labels only: a persona reads the chapter files in range, never the review copy. Cite each note with its paragraph label (ch03-p12) and the paragraph's first few words, exactly as the feedback template asks of human readers.

2. Gather context without spoilers

A persona reads the book as a reader would: only the chapters in range, in order. For background (who the characters are, what the story has promised so far), run story context on the last chapter in range:

story context chapter-{NN} --path . --budget 4000

It leaves out later chapters and the synopsis. Its What <name> knows that the reader has not seen (do not reveal) subsection holds facts from chapters the reader has not reached: delete that whole subsection, heading included, before any persona sees the output. Do not read story.md (the context already gives its language), plot/, arc files, promise payoffs, or chapters after the range: a persona that knows the ending cannot tell whether the setup works. When the genre reader needs the genre's promises, take them from the chapters and the genre, not from the outline.

3. Run each read independently

1. Run one persona at a time. If you can start subagents, give each persona its own subagent with only its reference file, the chapters in range, the story context output, and the book's language, so no persona reads another's notes. The line editor's subagent also gets style-sheet.md (when the project has one) and the story prose . findings for the chapters in range. Run story prose . first, and keep only its per-chapter findings for chapters in range. Otherwise finish and save one persona's file before starting the next, and do not revise earlier files after reading later ones.
2. Follow the persona's reference file. Read the whole range before writing any note.
3. Every problem cites evidence: the paragraph label, the quoted first words, and the words that cause the problem. A note with no location in the text is not a note; drop it.
4. Report what the persona finds, not a quota. A persona that finds nothing writes that in Overall Impression and leaves Problems empty. Never pad a read to look thorough, and never invent a problem another persona or a craft checklist would expect.
5. Read in the book's language. Judge prose, dialogue, and punctuation by that language's conventions and the style sheet, not English ones: guillemets or dialogue dashes are not errors in a French or Spanish book. Quote the text exactly as written, and write the notes in the language the user works in with you unless they ask otherwise.
6. Mark uncertainty. When a note depends on something the persona cannot see (a later payoff, the author's intent), say so in the note instead of asserting it.