For independent screenwriters
Most writers have never seen it. Paste a scene, upload a draft, or load a demo and see what WHO, Samaritans, and national guidance actually say, then keep, dismiss, or question every note.
Deterministic screenplay triggers select candidate passages. Google Agent Search retrieves source-linked clauses; a Google ADK agent on Gemini explains them and must self-check its alternative through the hard filter. No citation means no flag. The same service is a keyed public API.
MARA reaches for the phone. Her friend stays beside her.
“We can ask for help. We only need to get through this minute.”
Try the complete workflow
The honest premise
Some research associates fictional portrayal with changed suicide outcomes; reanalysis disputes whether the portrayal explains the underlying trend. The product does not use uncertainty as permission to overstate.
An informed reviewer should see the disagreement, source dates, and limits, not a dramatic statistic stripped of context.
Samaritans runs a human Media Advisory service. This pre-review extends access and routes difficult cases toward expertise; it is not a substitute for it.
No clinical claim. No compliance claim. No censorship path.
The fine print, in full
Bridge et al. (2019, JAACAP) and Niederkrotenthaler et al. (2019, JAMA Psychiatry) associated one series release with increased US youth suicides. Romer's 2020 reanalysis (PLOS ONE, Annenberg Public Policy Center) found no clear evidence of contagion after controlling for the underlying rise. We cite both sides. Overstating contested evidence is how a tool like this loses an informed reader.
WHO, Preventing suicide: a resource for filmmakers and others working on stage and screen; Samaritans' guidance on depictions of suicide and self-harm in drama and film; the 2019 National Recommendations for Depicting Suicide (National Action Alliance, SAMHSA, Entertainment Industries Council); Mindframe's resource for stage and screen (Australia); and the Mindset guide's suicide chapter (Canada). Thirty-six attributed clauses, each with a source link and version. The research behind them, fourteen verified studies including the ones that disagree, is on the Evidence page and is retrieved beside every note.
Samaritans' Media Advice team runs a script review service, staffed by people, with finite capacity. This exists to reach writers that service will never have capacity to serve, and it routes users toward expert review where one is warranted.
We never display "your scene is safe." A pre-review can only report what it detected against a corpus. It is not clinical or professional judgement and cannot substitute for it.
Depicting help-seeking, coping and recovery is associated in the literature with protective rather than harmful effects. The tool explains the mechanism rather than issuing a verdict, so a writer learns instead of complying.
An output filter compares the specificity of any suggested alternative against the writer's original and rejects it if method detail increased. The agent must run that filter on its own draft before it answers, the API runs it again, and Google Cloud Model Armor screens the result. A safety tool that generates more harmful detail than it found would be worse than nothing.
Deliberately absent. A single rating invites optimisation, and optimising a script toward a safety number is exactly the wrong behaviour.
The product surface must be something a writer can simply open: Replit hosts the deployed app, keeps a signed-in writer's decisions, and re-checks the corpus nightly. The agent itself is a Google ADK agent on Gemini, deployable to Vertex AI Agent Engine, grounded by Google Agent Search, and called over HTTP.
Architecture
Pure code shows the exact phrase and rule that produced a candidate. It cannot issue a guidance flag.
Google Agent Search returns a versioned source URL. A Google ADK agent on Gemini can explain only what was retrieved, and must run the safety filter on its own alternative before answering.
Accept, dismiss, or request expert review. Dismissal leaves the reasoning visible. On Replit, signed-in writers can keep their decisions. No overall score exists.
Integrate it
No account. Anonymous calls keep working; a key raises the limits and names the caller in meta. Every response carries the disclaimer as a required field, so an integration cannot strip it by accident.
The ribbon at the top of every page shows which Google Cloud and Replit services are answering right now. Status is earned by a real round trip, not painted on.
Approved corpus
World Health Organization · Samaritans · National Action Alliance
Corpus records include publisher, title, jurisdiction, version, retrieval date, applicable trigger classes, and source URL. Multi-jurisdiction results remain separate.