Constrain text review intotraceable findings and actionable conclusionsdeliverables
Input and Output
| Direction | Content |
|---|---|
| Input | The subject under review (text / document), optionally accompanied byreference input(for SSOT checks); with multiple files, they serve as reference input for each other, and cross-document rule groups are enabled accordingly |
| Output | Conversational analysis answer or report form, routing seesection |
Execution Principles
- Do not modify files on your own: the output is conclusions / a report; only after the user explicitly agrees do you proceed to rewriting (seesection）
- Every finding is traceable: map to at least 1 catalog rule name
- Scope convergence: only review scenarios included in the catalog; for doubtful facts, only point out the location and let the user verify manually, without expanding into verification
- Keep outputs within the conversation: reports / conclusions are not written to files (writing to files creates a second source of truth with no maintainer); intermediate outputs of the multi-model pipeline (materialized snapshots, scanned candidates written to disk) are not subject to this clause; for the disk-writing policy seeMulti-model Orchestration
- Self-review compliance: before delivering written outputs (report-form output / full rewritten text from Step 6), machine-check them with the Step 2 scanning script (materialize in-conversation ones into temporary files first) and clear to zero any hits attributed to yourself; hits arising from quoting the reviewed original text are exempt from adjudication, and for the remaining exemption policy see section; the appearance of penalized markers in your own writing equals a loss of credibility for the conclusions
Review Stance
The five stance points have a single source insection, and the default path and multi-model reviewers take that section as authoritative
Workflow
Step 1: Collect input
Parse the input, determine language + genre + length; determine whether there isreference input. Directory form by default includes only .md prose files;
other files enter the reviewed subject set only when the user names them. Length is counted by python3 tools/count_words.py <被审对象> (pasted text is first materialized into a temporary file; run from the skill root); when > 2000 characters, confirm the segmentation granularity with the user
(by chapter / by subsection; with multiple files, by file is allowed), and present findings in batches according to that granularity
Step 2: Load the catalog and fill in the gaps
Required reading ref/catalog.md. For candidates that can be judged mechanically, run the scan first:python3 tools/scan_fingerprints.py <被审对象>(run from the skill root; hits are handled according to the corresponding rule, exemptions according to section). Then use
catalog rules to fill in findings; when there is no reference input, state clearly in the conclusion that "this time only the document itself is reviewed". When the user explicitly requests multi-model /
cross-review, the finding production in this step as a whole switches toMulti-model Orchestration's subagent per-vote process, and the orchestrator no longer produces findings itself
Step 3: Form routing
By default go to conversational analysis answer(seesection); when the user explicitly asks for a report, enter report form(seesection); the form can be switched midway
Step 4: Conversational analysis answer (default)
- Understanding restatement: first use 2-4 sentences to explain clearly what this passageis saying(the core claim) andhow the logic proceeds(the chain of argument), and where uncertain explicitly say "here I understand it as X; if that is wrong, please correct me"
- Conclusion first: give in one sentence "is there a problem" (if not, explain the reason and wrap up)
- List key points: in document position order (file:line / section order), findings carry their own handling, without adding severity grading labels (Blocker / P0 type); if findings ≤ 3, give all directly (position + rule name + problem + suggestion); if > 3, give a full brief table (one line per item)
- Closing inquiry: when there are many findings, ask "go through them one by one or produce a report"; going through one by one presents them in document position order,
format "1/3 position
规则名: problem + suggestion. Accept / revise / skip / follow up?", and for each item wait for the user's stance and handle it according to the meaning of the words; after follow-up expansion, proceed to the next item - After going through all items one by one, output a summary (accept / revise / skip counts + adoption list), and ask whether to switch to report form or enter rewriting (seesection）
Step 5: Report form
Output according to the section order of ref/report-template.md (the first section of the template is the understanding restatement); at the end ask the user for the next step (refinement / one-by-one review / rewriting)
Step 6: Rewriting handoff (after confirmation)
For each accepted finding, execute the fix method of the corresponding catalog rule; for the silent path, execution constraints, output and self-check see rewriting handoff