Prerequisites
When the user specifies an article, check that file; otherwise run wewrite run show, prioritizing reading artifacts.draft, and read
artifacts.brief、artifacts.claims and artifacts.sources. Only fall back when an old task has no draft artifact
artifacts.article. Within the pipeline, first run wewrite run step review in_progress。
Read in full:
读取: {skill_dir}/../wewrite-write/references/article-brief.md
读取: {skill_dir}/../wewrite-write/references/editorial-quality.md
读取: {skill_dir}/references/seo-rules.md
When the user only says "check it" and provides an external file, give only a report; in the main workflow, for an in-task draft, or when the user explicitly says "optimize," the necessary revisions must be completed directly.
Review and revision
1. Align with the task
Check whether the article truly answers audience.question, whether the core judgment is consistent with thesis , whether each section advances the specified
claim, and whether the ending delivers goal.takeaway/action. Delete irrelevant paragraphs; do not use a pretty structure to cover up an answer that misses the point.
2. Verify facts and personal material
Run wewrite sources list --json, and match the specific numbers, dates, quotations, research conclusions, and time-sensitive facts in the body
item by item against claims.yaml and the original sources. When the sources do not support them, prioritize further checking; if they cannot be found, delete them, narrow them, or turn them into explicit
inferences. Do not use model memory to fill gaps.
Check all first-person events, friends and colleagues, interviews, conversations, times and places, and sensory details. They must come from user material explicitly recorded for this task; otherwise they are blocking issues and must be deleted directly or changed into non-firsthand discussion.
3. Five editorial judgments
Score each of "accuracy, viewpoint, usefulness, voice fit, readability" from 1-5, and list blocking issues and at most 5 major issues:
pass: average score at least 4, no individual item below 3, and no blocking issues.revise: can be corrected within the existing material. After making the revisions directly, re-execute 1-3; do not merely write suggestions.needs_input: limited to cases where the user explicitly requests a personal story, the material is insufficient, and the framework cannot be safely changed.
At most two rounds. If problems remain in the second round, delete unreliable content, narrow the promises, and generate a reliable version that can pass; do not label an article that has not
passed as "deliverable." After it passes, write the final body into artifacts.article。
4. Title, summary, and tooltips
Generate one main title, two accurate alternative titles, a summary within 40 characters, and 3-5 tags. Keywords should appear naturally; do not force them in based on density, and do not fabricate numbers or promises for open rate. Review does not call image generation.
Run wewrite score {article_path} --json. This score only helps locate clichés, overly uniform sentence patterns, and paragraph rhythm risks;
it does not set a mechanical pass line, nor does it repeatedly rewrite to raise the score.
Save the editorial report
First write the editorial judgments as a temporary assessment.yaml：
decision: pass
pass_number: 1
dimensions:
accuracy: 4
viewpoint: 4
usefulness: 4
voice: 4
readability: 4
blockers: []
major_issues: []
notes: ""
in the task directory, then generate artifacts.review_report：
wewrite content-eval --draft {draft_path} --final {article_path} \
--assessment {assessment_path} --output {review_report_path} --json
Only the publishable=true in the report can complete the review. Update the task and mark it complete:
wewrite run update --patch '{"editorial":{"decision":"pass","pass_number":1,"publishable":true},"seo":{"title":"...","alt_titles":[],"digest":"...","tags":[],"quality_score":0},"provenance":{"verified_sources":0,"unverified_sources":0}}'
wewrite run step review completed
Use natural language for the self-check report, list at most 5 issues ordered by impact, and explain how they have been corrected. If there are no hard flaws, say directly that the body can be delivered; if needed, images or layout can be added separately afterward, without outputting a long score table.