Feedback Triage

Overview

Process alpha/beta reader feedback as a structured, reconcilable workflow: collect per-reader feedback files, hold all revision until the round is complete, synthesize convergent/divergent/single-reader findings into a decision record with a readiness verdict, and hand a concrete revision plan to the revision-continuity skill.

Prerequisites

A story project with at least one drafted chapter (or a complete draft) that readers have read. Verify story.md exists in the project root.

When to Use

• Starting a feedback round (recruiting readers, sending chapters out)
• Giving readers a review copy, including the GitHub Pages review-copy workflow and the reader-note issue form (step 1 of the workflow; other skills point here)
• Recording feedback as it arrives
• Synthesizing a completed round into decisions
• NOT for revising the manuscript (use revision-continuity with the synthesis's revision plan)
• NOT for the agent's own critique of the draft (use revision-continuity audits; reader feedback is external input)
• NOT for producing simulated persona reads (use reader-panel; this skill synthesises the round it writes)

Workflow

1. Set up the round

1. Decide the round scope: which chapters readers get (chapters-read range) and how many readers (2–4 per round is typical; one reader is a data point, not a round).

2. Create the round folder: feedback/round-{N}/.

3. Give readers a review copy they can open without a terminal. Build it from text you can get back, so the round can be rebuilt and old labels mapped later (step 2 of Collect). Save that text first:

   • 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/ (story init writes one that does; 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. 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 "Feedback round {N}" -- .
     git tag feedback-round-{N}
     

     If the user declines the commit, never tag the last commit over an uncommitted tree: take a snapshot as below, or stop.

   • Project without git, or the user declined the commit: take a snapshot with story snapshot feedback-round-{N} --path ..

   Then build the stamped HTML copy from that text:

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

   Never push, and never move or delete a tag, without the user's approval: a push to main can publish the manuscript through the review-copy workflow below. The single-file HTML copy in dist/ has a table of contents, the build stamp at the top, and a paragraph label beside every paragraph (ch03-p12 is chapter 3, paragraph 12). A label is the chapter and the paragraph's position in that build, not a permanent id: any earlier edit in the chapter renumbers it, and story move changes its chapter part. Ask readers to cite the label, the build stamp, and the paragraph's first few words with each note.

   GitHub review copy. This is the setup the other skills point to. For a project on GitHub, offer the templates from the Story Skills repository (https://github.com/danjdewhurst/story-skills, templates/github/). The review-copy.yml workflow publishes the HTML copy to GitHub Pages on each push to main, stamped with the date and short commit, with a Note link beside every label (--note-url) that opens the issue form prefilled with the label, build, and first words. Like every build, it leaves out a matter page whose permission is pending or unclear (warning permission-pending-left-out), so an uncleared epigraph never reaches Pages; never add --include-pending to the workflow. The ISSUE_TEMPLATE/manuscript-note.yml issue form asks readers for the label, build, first few words, note type (typo or wording, confusing, continuity, pacing, character, sensitivity or authenticity, loved this, other), how much it affected their reading, and the note. Before asking to copy them, warn that a public Pages site makes the manuscript public unless the repository and Pages are private, and confirm the visibility the user wants. Then copy them into the story repository's .github/workflows/ and .github/ISSUE_TEMPLATE/ only with the user's approval, and create a manuscript-note label first; GitHub only applies existing labels.

4. For each expected reader, create a stub file from references/feedback-template.md at feedback/round-{N}/{reader-kebab}.md with frontmatter filled in and the body sections empty. The stub list is the round's checklist.

2. Collect feedback (the discipline)

1. As each reader's notes arrive, record them in their file using the template. Quote or closely paraphrase; do not editorialize yet. Keep each note's paragraph anchor (ch03-p12) in its Where line; convert chapter or page references from other formats to anchors when the location is unambiguous. For notes filed through the issue form, fetch them with gh issue list --label manuscript-note --state open --json number,title,body,author and map the form's "How much did it affect your reading?" answer to the template's severity: Made me want to stop reading is major (blocking when several readers stopped at the same place), Pulled me out for a moment is minor, Barely noticed is nit, and a blank answer is minor. A Typo or wording note with a blank answer is a nit unless the reader says more.

2. Map old labels to the current text. When a note's build is older than the manuscript, its label may point at a different paragraph now. Resolve every label from the round in one run against the text the round saved in step 1: the git tag (or the short commit in the note's build stamp), or the snapshot of the same name when the project has no git or the user declined the commit. A reader-panel round saved its text the same way as panel-round-{N} instead of feedback-round-{N}. git tag --list 'feedback-round-{N}' and story snapshot --list --path . show which one the round has. Readers type labels into the issue form, so before a label goes into the command, check that it is lower-case letters, digits, and hyphens ending in -p and a number (ch03-p12, front-epigraph-p1), and ask about any other. Quote each label:

   story compare . --ref feedback-round-{N} --anchor 'ch03-p12' --anchor 'ch07-p4'
   

   For a snapshot, use --snapshot feedback-round-{N} in place of --ref. Each line gives the current label: (text unchanged), or (edited, NN% similar) when the paragraph was revised (check it is the one the reader meant). not found in the current text ("…") means the paragraph was cut or rewritten past recognition: search the chapter for the reader's quoted words, or the words shown, and mark the note ambiguous if nothing matches. no such label means the label never existed in that build: check the note's build and the reader's typing. Record the current label in the Where line, keeping the reader's original label in brackets. Sensitivity and authenticity reads use the same file shape; see the editorial-review skill for commissioning them.

3. Run the canon check on each problem note: verified against the bible, contradicts canon (usually a setup problem — note the canon file), or outside canon scope. Record the result in the file.

4. Do NOT revise until all feedback for the round is in. Revising on partial feedback optimizes for the first reader and invalidates the others' reads. If a reader is late, either wait or formally close the round without them (note it in the synthesis) — never silently proceed on a partial set.

3. Synthesize

Only when every expected reader file is collected. If the round's files carry source: simulated, it is a panel round from the reader-panel skill: follow "Simulated rounds" below as well.

1. Read all reader files for the round.
2. Sort every finding into exactly one category:
   • Convergent — ≥2 readers agree independently. Strongest signal; becomes a revision item by default.
   • Divergent — readers disagree. Adjudicate: check both sides against canon and premise, record which side wins and why.
   • Single-reader — one reader only. Weigh by specificity: specific + canon-verifiable → investigate or accept; vague + taste-based → usually decline.
   • Declined-with-reason — explicitly rejected, with a recorded reason referencing canon, premise, genre contract, or craft principle.
3. Write feedback/round-{N}/synthesis.md using references/synthesis-template.md, including the frontmatter readiness verdict: ready | needs-revision | not-ready.
4. Build the numbered revision plan with concrete file targets.

4. Hand off the revision plan

1. Present the synthesis summary and readiness verdict to the user.
2. If the verdict is needs-revision or not-ready, hand the revision plan to the revision-continuity skill for execution. The synthesis is the input; revision-continuity owns the edits.
3. If the verdict is ready, the round is closed — proceed to the next round, the next drafting stage, export, or the submission skill.

Simulated rounds

The reader-panel skill writes persona reads in this skill's file shape, with source: simulated and persona in the frontmatter. A file without source, or with source: human, is a human reader's. Synthesise a panel round as usual, with these differences:

• Label it. Set source: simulated in the synthesis frontmatter and start the readiness line with "Simulated round:". Refer to the files by persona ("the line-editor persona"), never as readers or beta readers.
• Personas are not independent. Several personas run by one model agreeing is one signal, not convergence. Sort every simulated finding as single-reader or declined-with-reason, weighed by how specific and checkable it is: a quoted POV slip or a contradiction with both sides cited is worth acting on; a taste note is usually declined.
• Check before acting. Personas leave the canon check to you (they cannot read past their range), so run it here, and confirm each simulated problem in the text before it enters the revision plan. A note the text does not bear out is declined with the reason "not borne out by the text".
• What ready means. A simulated round's ready means ready for human readers, nothing more. It never closes a book for submission or publication; hand off to a human round, not to submission.
• Never mix rounds. Simulated and human reads go in separate rounds. If a round holds both, move the simulated files to their own round before synthesising. When a later human round repeats a simulated finding, the human readers' notes carry it; the earlier panel adds no weight.
• The sensitivity persona only flags passages for a human reader. Its notes become items for an editorial-review brief, never a finding that a portrayal is fine.

Conventions

• Feedback lives under feedback/round-{N}/; {N} is a plain integer (round-1, round-2).
• Reader files use kebab-case reader ids: feedback/round-1/maria-chen.md.
• Locations cite paragraph labels from story build --format html (ch03-p12) where available. Labels are paragraph positions in one build, so tag and stamp each round's build, rebuild and resend the review copy between rounds, and map an old label to the current text with story compare . --ref <round-tag> --anchor '<label>' (or --snapshot <round-name> when the round was saved as a snapshot; step 2.2) rather than reusing it after a revision.
• Every feedback file and the synthesis carry YAML frontmatter (reader, round, chapters-read, overall-verdict / readers, readiness). Simulated reads and their synthesis also carry source: simulated; simulated reads carry persona.
• Findings are quoted or closely paraphrased from readers, never invented. If a note is ambiguous, mark it ambiguous in the file rather than resolving it silently.
• Declined findings always carry a recorded reason. A synthesis with unexplained rejections is incomplete.
• Bidirectional discipline: when synthesis creates or resolves continuity/questions/ entries (reader confusion often reveals clarity gaps), update those files too.

CLI Maintenance

Use the Story CLI when it is available. If story is not installed, use the bundled fallback node ../story-maintenance/scripts/story.js with the same arguments. Use node <checkout>/bin/story.js instead only when the user names a Story Skills repository checkout or you are working in one. Write the script as an absolute path (resolve the fallback relative to this skill folder) and run it from the folder you would run story from, so . and other relative paths keep their meaning. Use Node, not Bun or a package script: Bun would load that folder's bunfig.toml (which can run code) and .env, and a package script runs from the checkout's root. If no CLI is available, perform the registry, backlink, and word-count checks manually.

After creating or updating feedback files and synthesis:

story reindex .
story wordcount . --write
story check .

Reference Files

• references/feedback-template.md - Per-reader feedback file template with frontmatter (reader, round, chapters-read, overall-verdict, and source/persona for simulated reads), paragraph anchor citations, the review-copy note for readers, and canon-check discipline
• references/synthesis-template.md - Round synthesis template: convergent/divergent/single-reader/declined-with-reason categories, readiness verdict, revision plan

Shared Conventions

Every story skill follows the shared conventions in ../story-maintenance/references/conventions.md, resolved relative to this skill folder. Read it before creating, renaming, or linking story files. If that file is missing because this skill was installed without story-maintenance, the essentials are: kebab-case ids and filenames, YAML frontmatter on every story-project file, _index.md registry tables that story reindex rebuilds (never edit them by hand), bidirectional links between entities, characters for who is on the page and mentions for who is only referred to, status: deceased plus died-in: chapter-{NN} for deaths, and no project-local generator or build scripts (run only the installed or bundled Story CLI).