简体中文 · English
A privacy-first clarity tool for conversations that leave you guilty, confused, frightened, or smaller than before.
AI-assisted analysis ·
No sign-up ·
No diagnosis ·
English / 中文
Some of the hardest conversations do not look dramatic from the outside.
They arrive as a calm “I’m only doing this for you,” a joke that leaves someone ashamed, a manager quietly rewriting what happened, or a family member turning every boundary into proof of disloyalty.
After enough conversations like these, people often stop asking “Was that fair?” and start asking “Am I the problem?”
That gap matters. Psychological aggression can involve language used to cause emotional harm or exert control. Emotional abuse may appear as isolation, intimidation, humiliation, guilt or denial rather than physical violence. Workplace bullying and harassment are also recognised psychosocial risks—not merely personality clashes.
Grow Into Yourself was built for the moment before someone has the right words: when they have a chat in front of them, a knot in their stomach, and no neutral way to read what just happened.
It does not decide who is good or bad. It helps the user slow the conversation down, examine the actual words, notice pressure, and recover enough clarity to choose what happens next.
| Read the conversation | Find the pressure | Recover your choices |
|---|---|---|
| Separates the other person’s words from the user’s reply | Flags possible guilt, humiliation, blame-shifting, threats, denial, isolation or control | Suggests soft, firm and exit-style boundary replies |
| Reviews individual sentences instead of producing a generic summary | Explains why a sentence may hurt and where evidence remains uncertain | Surfaces safety concerns only when the text supports them |
The result is designed more like careful margin notes than a chatbot verdict:
- What happened
- Possible pressure signals
- Patterns in the user’s own reply
- Sentence-by-sentence annotations
- A clearer reading of each sentence
- Three boundary reply options
- Risk level and proportionate safety guidance
Behaviour, not labels. The tool does not diagnose NPD or any personality disorder. It discusses observable language, repetition, power, boundaries and impact.
Uncertainty stays visible. One sentence cannot define an entire relationship. When the evidence is incomplete, the analysis should say so rather than inventing context.
Safety without alarmism. Urgent guidance is reserved for credible signs such as threats, stalking, forced control, danger to a child, self-harm coercion or harm to others.
The user keeps agency. The output offers interpretations and possible wording. It does not order the user to reconcile, confront, forgive or leave.
Other person’s messages ─┐
├─> /api/analyze ─> local evidence-linked analysis ─> structured JSON
My reply or draft ───────┘ │
└─> optional external AI overlay
- The user pastes the other person’s messages into the first box.
- Their previous reply or draft response can be added separately.
- The server-side route builds a local, evidence-linked result first.
- When strict external-AI configuration is present, an external model may add a validated structured overlay without replacing the local result.
- If the provider is unavailable, times out, or returns invalid output, the local result remains available and clearly labelled.
Keeping the speakers in separate fields avoids unreliable speaker guessing and removes slow mobile OCR from the critical path.
| Layer | Implementation |
|---|---|
| Interface | React, TypeScript, Vite / Vinext |
| Runtime | Cloudflare Vite integration with a server-side POST /api/analyze route |
| Core analysis | Local-first, evidence-linked rule engine |
| AI overlay | Configurable OpenRouter-compatible server request, strictly schema-validated |
| Configuration | ANALYSIS_API_KEY, ANALYSIS_API_URL, ANALYSIS_MODEL, ANALYSIS_STRICT_PRIVACY |
| Response | Structured JSON with sentence analysis, evidence links, and reply options |
| Resilience | Local result survives missing configuration, timeouts, and invalid AI output |
| Languages | English and Simplified Chinese |
| Safety | Non-diagnostic prompting, confidence calibration, and urgent-risk gating |
The API key belongs only in server-side environment variables. It must never be committed to GitHub or exposed in client-side JavaScript.
ANALYSIS_API_KEY=
ANALYSIS_API_URL=
ANALYSIS_MODEL=
ANALYSIS_STRICT_PRIVACY=trueThe external overlay is optional. Without these variables, the local analysis remains available. Keep secrets only in server-side environment variables; never commit them to GitHub or expose them in client-side JavaScript.
npm install
npm run devThe development server uses Vite / Vinext with the project’s Cloudflare-compatible local bindings. To verify a production artifact locally, run npm run build; to serve it, run npm start.
npm testThis runs the verified build and Node test suite, including evidence-grounding, calibration, bilingual behaviour, fallback, timeout, API, and regression coverage. The current final regression suite passes 71/71 tests.
Text may be sent to the configured AI model to generate this analysis. This site does not save the conversation. Remove names and avoid identity numbers, bank details, passwords and exact addresses.
The project is privacy-conscious, but it does not make the inaccurate claim that AI analysis happens entirely on the user’s device.
Grow Into Yourself is not a therapist, emergency service, legal adviser or diagnostic system. It cannot determine a person’s intentions from a short excerpt, and it should not replace qualified local support where immediate safety is involved.
The product direction is informed by established definitions and research:
- WHO — Violence against women
- CDC — About intimate partner violence
- UN Women — Signs of relationship abuse
- The National Domestic Violence Hotline — Emotional abuse
- The National Domestic Violence Hotline — Types of abuse
- WHO — Mental health at work
- ILO — Experiences of violence and harassment at work
- UNICEF — Violence against children
These sources do not endorse or validate this software. They support the underlying premise that psychological pressure, controlling behaviour, emotional abuse and workplace harassment can be consequential even when no physical injury is visible.
The most useful thing GPT-5.6 did was not agree with me too quickly.
Whenever I made a strong interpretation, it pushed back:
- Which exact words support that conclusion?
- Is there another reasonable explanation?
- Are we seeing concern, or a pattern that narrows someone’s choices?
- What new evidence should lower our confidence?
Those questions became part of the product itself.
I reviewed 73 peer-reviewed studies and translated the research into 38 decision rules and 39 minimal pairs. Strong labels require stronger evidence. Ordinary budgeting should not become “economic control.” A clear breakup should not become “silent treatment.” Concern and pressure can exist in the same conversation.
I made the core product decisions:
- no personality diagnosis;
- every important conclusion must return to the original words;
- uncertainty must stay visible;
- the local result appears first;
- external AI may improve the analysis, but it must never erase an already useful result;
- real danger should lead to real-world support, not just a softer reply.
Codex turned those decisions into working constraints:
- evidence-linked analysis;
- separate handling for family, dating, friendship and workplace conversations;
- local-first fallback behaviour;
- bilingual UI and analysis rules;
- timeout and failure states that preserve the result;
- regression tests for evidence mismatch, overclaiming and cross-case contamination;
- deployment and repository synchronisation.
The final regression suite passes 71/71 tests.
GPT-5.6 helped me make empathy testable. Codex helped me make restraint executable.
This is an early public test built to explore a narrow question:
Can AI help someone read a difficult conversation more clearly without diagnosing strangers, exaggerating danger, or taking away the user’s choices?
Feedback is welcome, especially where the analysis feels vague, overconfident, culturally awkward or unsafe.
Understand the behaviour. Keep your choices. Grow into yourself.