skillZs
★ LIVE SKILL TAGS ★
>>> LIVE SKILLS INDEX <<<
* OPEN SOURCE *
NO LOGIN, NO TRACKING
※ REAL INSTALL DATA ※
← back to all skills
shengchen1998/author-toolkit49 installs

author-review

Use when running independent final or current-scope manuscript reviews for intent and consistency, writing standards, and a fresh target reader, or preparing separate review handoffs when subagents are unavailable.

How do I install this agent skill?

npx skills add https://github.com/shengchen1998/author-toolkit --skill author-review
view source ↗

Is this agent skill safe to install?

  • Gen Agent Trust Hubpass

    The skill provides a specialized framework for reviewing literary manuscripts across three distinct perspectives (intent, standards, and fresh reader experience). It implements a layered review protocol using a chunking strategy to manage context limits. While it relies on local sibling skills and a custom 'author-project' tool for reporting and state management, these are standard project-specific capabilities. A low-severity surface for indirect prompt injection exists due to the processing of untrusted manuscript data.

  • Socketpass

    No alerts

  • Snykfail

    Risk: CRITICAL · 1 issue

What does this agent skill do?

Author Review

Organize three independent reviews and produce findings only. Accept all, one perspective (intent, standards, or fresh), or collect, plus a project root, review job, or review handoff. Accept review_mode=final|current-scope|request; the default is final.

This skill does not revise prose, update Story Bible creative content, or apply fixes. When called by write-fiction, revise-fiction, revise-draft, revise-increment, or revise-request, the caller owns manifest and tracker writes. Standalone use leaves shared project state unchanged unless the author explicitly incorporates the resulting reports later.

Prepare

  1. Resolve the project root from the supplied project root, review job, or review handoff. Toolkit skills from one installation are siblings. Resolve <setup-skill-root> as ../setup-author-toolkit relative to this skill's installed directory. Read ../author-project/SKILL.md and ../author-project/references/read-operations.md, then run its read and status procedures separately. If it reports an incomplete toolkit installation, stop and report the missing installed files; do not deploy contracts into the project. These are the only skill procedures this skill may run. Do not run other skills from a reviewer or subagent.
  2. Read the controlled workflow terminology in <setup-skill-root>/assets/references/terminology.md, references/checklists.md including its review-mode applicability rules, and references/chunked-review.md before planning, resuming, or collecting a review. Read Report envelope and Finding identity across reruns; state-write instructions remain the caller's responsibility.
  3. Select the review mode and its output identity:
Review modeReport directory under paths.reportsReport v2 stageHandoff subdirectory under paths.handoffs
finalfinal-review/final-reviewfinal-review/
current-scopecurrent-scope-review/current-scope-reviewcurrent-scope-review/
requestcurrent-scope-review/current-scope-reviewcurrent-scope-review/

With default paths, these report destinations are reports/final-review/ and reports/current-scope-review/. Use unique UTC-stamped filenames. Creation callers retain their own operation identity and never add a creation operation to report v2 stage. A request report's operation_coverage describes the requested perspectives and checks, while its Review coverage identifies the actual selected units; neither claims full-target reading. 4. Establish one confirmed scope_id and one manuscript scope for all three perspectives. A final review covers the whole book. A current-scope review covers only the confirmed current scope and never claims whole-book final approval. A request review retains the target scope ID but reads only the declared changed and dependency-affected passages with necessary local context; it never claims full-target coverage or current-draft revision convergence. Apply the checklist's skipped-item rules for whole-book and unreviewed-passage questions that do not apply. 5. Create or resume one review job per requested perspective and its chunk -> volume/phase -> book plan from actual manuscript paths, hashes, chapter or scene boundaries, and the available context budget. Store it under .author-toolkit/reviews/ or the registered paths.sessions; record fresh reader-state at its reviewer_state_path. Use the installed review-job and reader-state artifact contracts; the handoff carries their paths rather than copying their content. 6. Use the uppercase finding ID prefixes INTENT, STANDARD, and READER. Before dispatch, request the required consecutive ID block from author-project with next-finding-id and place the exact block in each handoff or reviewer prompt. Assign disjoint subranges of that block to independently written chunks. Reuse IDs only after the caller reconciles prior observations; an isolated reviewer does not read prior reports or allocate its own IDs. Every reviewer uses judgment confidence and includes a recommendation. The caller records every returned report, including clean reports, through author-project record-findings; that transaction enters new findings as open and indexes each audit run. Status is not a finding v2 property. 7. Preserve input independence as follows:

PerspectivePermitted inputs
Intent and consistencyManuscript, Story Bible, approved draft-baseline when supplied, approved revision plan, and intent checklist
Writing standardsManuscript and writing standards checklist only, plus references/chinese-writing-standards.md for Chinese work; exclude previous self-assessments, Story Bible, draft-baseline and other baselines, plans, and prior review reports
Fresh readerActual manuscript units in sequence, target reader description, fresh-reader questions, and only its own reader-state after the first unit; exclude all other review and baseline inputs

For language=zh* or an actually Chinese manuscript, the standards reviewer must load references/chinese-writing-standards.md. If the language label and manuscript disagree, follow the actual manuscript and disclose the mismatch. Include this reference in that reviewer's permitted inputs and handoff; do not give it to the intent or fresh-reader perspectives.

Run layered reviews

Follow references/chunked-review.md for unit coverage, synthesis dependencies, reader-state, findings, and completion. Dispatch independent perspectives with fresh contexts. intent and standards chunk work may run in parallel; the fresh chain remains strictly sequential with one logical reviewer. Do not share one perspective's output with another.

Write immutable chunk reports and each applicable volume/phase and book synthesis report through author-project scaffold-report with human-readable scope, a Language coverage section, a Review coverage section, and exactly one report v2 JSON envelope. Mark inapplicable whole-book checklist items skipped with reasons in current-scope mode; in request mode, mark checks requiring unread passages skipped with outside request review and record every other gap truthfully.

For Chinese standards, Language coverage must itemize the Chinese subchecks in the loaded reference as actually performed, with full, partial, or skipped status and reasons for gaps. Claim full only when every applicable Chinese subcheck covers the requested actual manuscript scope; a sampled, translated, unread, or unexecuted check cannot be folded into full. Keep Review coverage tied to actual manuscript units and syntheses.

Use engine.name: "conversational" and record the actual execution path, version or unknown, and degradation. Every report's engine.notes contains:

  • scope_id=<actual-scope-id>;
  • review_mode=final|current-scope|request;
  • actual operation_coverage=full|partial|skipped;
  • subchecks=intent, subchecks=standards, or subchecks=fresh for one perspective;
  • a language coverage summary with the actual status, for example Language coverage: full (zh-CN) or coverage=full.

For a consolidated report containing all perspectives, use subchecks=intent,standards,fresh and the same scope ID. Preserve every finding ID and disagreement. collect validates the review job and coverage, including all units and syntheses and the fresh reader-state chain, before reporting consolidated or remaining coverage.

Run without subagents

Resolve paths.handoffs and create resumable handoffs from assets/review-handoff-template.md beneath the mode-specific <paths.handoffs>/<review-mode-directory>/<review-id>/. Use the registered custom directory when present. Give each handoff only its permitted inputs, review-job path, next unfinished unit, report destination, and resume command. Only the fresh handoff names its own reader-state path.

Stop while work is pending. Resume the named perspective in a separate session, write the completed chunk or synthesis report, update the review job and handoff to the next unfinished unit, and stop again when no further work fits. Fresh sessions continue the same logical reviewer only through that reviewer's own reader-state. Once all required evidence exists, the author can start author-review with collect and the same review mode or return to the calling strategy's review operation.

Findings and completion

Use author-review-intent, author-review-standards, and author-review-fresh-reader as audit names. Report concrete observations at verified locations. Linked cross-file findings retain real locations; no broad issue receives an invented single-file location. Produce no manuscript replacements or automatic resolutions.

A perspective is complete only when every required unit and applicable synthesis validates and covers the agreed scope, with all gaps disclosed. all is complete only when all three perspectives meet that condition for the same current scope_id and review mode. Return report paths, actual coverage, and finding counts; do not start another operation or describe current-scope or milestone evidence as whole-book final review.

Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.

<a href="https://skillzs.dev/skills/shengchen1998/author-toolkit/author-review">View author-review on skillZs</a>