A workflow: lay out the evidence on both the supporting and opposing sides of a viewpoint at the same time, so you can see clearly where this viewpoint holds up and where it is most fragile. This is not about making the judgment for you; it is about making the situation clear.
First read references/sci-cli.md—CLI detection steps, usage, and iron rules.
Then read references/output-system.md—HTML shell specification and design tokens.
Steps (fixed order)
Step 0 — Confirm CLI availability
Run sci --version. If not installed, give the installation command and stop; if there is an authentication error, prompt sci init and stop.
Step 1 — Lock down the claim
Compress the viewpoint into a precise, falsifiable sentence, and dig out the implicit scope (population, conditions, measurement method). Number the claim C1; when there are multiple claims, split them into C1、C2、C3, and by default handle the most central one first.
Tell the user that this run will execute about 10–13 searches, then start directly without waiting for confirmation.
Step 2 — Supporting-side search
For the claim, execute sci search，--mode low，--limit 15, leaning toward angles that directly support it. Filter according to a four-level standard (direct support / partial support / indirectly related / excluded), and mark the strongest one.
Step 3 — Opposing-side search
Construct five types of opposing queries and search them one by one (use ultra_lowfor broad angles, and rerun fruitful directions with low ):
- Direct contradiction
- Boundary / complication
- Alternative explanation
- Methodological criticism
- Replication failure / null result
Rate the threat level of each opposing source (high / medium / low), with a one-line response direction attached.
Step 4 — Cross-analysis After the searches are complete, do two cross-checks:
- Angles with support but no opposition: list them (this does not mean safe; it may simply mean insufficient research)
- Angles with opposition but no support: list them (this is the most fragile place)
Step 5 — Comprehensive judgment Based on the search results (without adding memory-based inference), write a 2–4 sentence description of the evidence landscape: the overall signal strength on the supporting side, the most concentrated type of challenge on the opposing side, and the overall state of evidence balance. Do not make a final judgment about "whether your viewpoint is right or wrong".
Outputs
After the searches are complete, create in the current directory the {topic-slug}/ subfolder, and output four things:
- Terminal — directly print the comprehensive summary: preferred citation on the supporting side + highest threat on the opposing side + key points of the cross-analysis
report.md— stable YAML front matter + comprehensive judgment + supporting-side evidence ladder + opposing-side threat matrix + cross-analysis + retrieval transparencyreport.html— comprehensive report, see the HTML specification belowreferences.bib— merge all queries'.bib, and after deduplication keep only the literature cited by the report
HTML report specification
Accent color #7c3aed(Purple, distinct from the green when bear-support is run alone and the red of bear-counter.)
Above the fold: comprehensive judgment panel
.verdict Three columns side by side:
- Left: strongest supporting evidence (author year + one-sentence supporting relationship, green accent)
- Middle: 2–4 sentence description of the evidence landscape (purple accent, core position)
- Right: highest threat on the opposing side (author year + threat type + one-sentence response direction, red accent)
.metric-strip: number of effective supporting references / number of high-threat counterevidence / number of angles with support only and no opposition / number of angles with opposition only and no support
Cross-analysis panel(below the fold, standalone section)
Two columns side by side:
- Left column "Has support, no opposition found": list angle names + note "absence of opposition is not the same as absence of controversy"
- Right column "Has opposition, lacks support": list angle names + label fragility level (high / medium)
Detailed evidence area (tabs)
Two tabs: "Supporting side" and "Opposing side", with "Supporting side" shown by default.
Within the supporting side tab: evidence ladder (direct support / partial support / indirectly related), with reference cards containing DOI links and collapsible Chinese academic abstracts.
Within the opposing side tab: displayed in three groups by "high threat / medium threat / low threat", with the five challenge directions labeled as pill tags on each card; do not nest another layer of tabs. Each card contains threat level + 回应：... + DOI link + collapsible Chinese academic abstract.
Add the .card-primary class to the most important cards (the preferred citation on the supporting side and the highest threat on the opposing side), with abstracts expanded by default.
Search transparency(bottom): list all query terms, mode, number returned, number adopted, and directions with empty results.
Honest boundaries
All references come from this search. "Has support and no opposition" only means that none was found in this search, not that the angle is free of controversy. The cross-analysis is a slice of search coverage, not a complete evidence map.
Gotchas
- The query terms for the searches on both sides must be based on the same claim, do not search broad background literature on the supporting side and the core controversy on the opposing side—inconsistent standards on the two sides will make the cross-analysis meaningless
- The cross-analysis cannot be omitted: this is the core of why bear-review is more valuable than running the two skills separately, do not skip it just because the search results are complex
- The description of the evidence landscape must be restrained: 2–4 sentences, state well-grounded judgments such as "the supporting-side signal is overall fairly strong but concentrated under specific conditions", do not say "therefore this view is correct"