Draft the hard message without outsourcing your judgment.
The raw draft often contains the most sensitive names, events, and emotions. A useful assistant should help structure it without silently sending that context elsewhere.
At a glance
AI can help draft a difficult message by separating the intended outcome, observable facts, feelings, request, and boundary, then producing several tone options for the user to review. Use a local model for sensitive context when possible, remove unnecessary identifying details for online routes, and never send the first output automatically. The sender remains responsible for accuracy, timing, consequences, and whether a message is safe to send at all.
How Phos handles it
Three clear routes
Local model
Sensitive background and drafts can remain in an on-device prompt path.
Local server
The personal server receives names and details included in the request; redact if the host is shared.
Bring your own key
The provider receives the submitted context, so use placeholders and the smallest useful excerpt.
Who this is for
Built for people with real private work to do.
People writing a boundary, apology, feedback note, rejection, or repair message.
Managers and founders who need clarity without leaking personnel context.
Students handling conflict without turning a private situation into a cloud record.
Anyone who wants multiple drafts before deciding what sounds like them.
Plain comparison
The point is control, not a louder chatbot.
These pages are for people comparing real options. Phos should win when someone wants privacy, local control, no account wall, and an assistant that still feels good to use.
| Feature | Phos | Typical cloud chatbot | Raw local app |
|---|---|---|---|
| Input style | Messy notes become facts, goals, request, and boundary. | General rewrite prompt. | Depends on user's system prompt. |
| Tone choices | Direct, warm, brief, or formal alternatives. | Usually available. | Model-dependent. |
| Privacy path | Visible before generation. | Hosted by default. | Local when the model is local. |
| Send control | Draft for review; no autonomous sending. | May integrate with communication tools. | Normally copy and paste. |
Define the outcome before choosing words
Write one sentence describing what should be different after the recipient reads the message. Possible outcomes include acknowledging harm, setting a boundary, requesting a decision, documenting a concern, declining work, or reopening a conversation. If there are three unrelated outcomes, split them or choose the most important one. A message that tries to punish, persuade, explain, and solve everything at once is hard for any model to repair.
Then list only observable facts: dates, actions, commitments, and direct effects. Put interpretations in a separate section. Ask the assistant to flag claims that sound like mind-reading or absolutes. The goal is not to make the message emotionless; it is to keep emotion from disguising speculation as fact.
Use a five-part drafting frame
A durable structure is context, observation, impact, request, and boundary. Context explains why the message exists. Observation names the specific behavior. Impact describes what changed for the sender or work. Request asks for a concrete next step. Boundary states what the sender will do if the situation continues, without threatening an outcome they cannot or will not enforce.
Ask for three versions: a short direct draft, a warmer draft that preserves the request, and a formal record. Compare what each omits rather than automatically choosing the nicest. A model often softens the important sentence until it becomes ambiguous. Keep the request measurable and the boundary under the sender's own control.
Redact before using an online model
Replace real names with roles, remove addresses, account details, health information, customer identifiers, and quoted private messages that are not necessary. Summarize the relevant event in neutral terms. If exact wording matters, local mode reduces provider exposure, but the draft still lives on the device and can enter backups, exports, screenshots, or the clipboard.
For workplace disputes, legal notices, safeguarding concerns, harassment, or situations involving power and retaliation, AI drafting is not a substitute for policy, counsel, or a trusted human reviewer. Do not upload confidential records merely to improve tone. The more serious the consequence, the smaller the model's role should be.
Run a pre-send risk review
Check every factual statement, name, date, and promise. Read the message aloud and remove language you would not say in person. Ask whether the channel is appropriate: some topics need a call, a formal process, or no immediate response. Delay sending when angry, sleep-deprived, intoxicated, or afraid, unless urgency involves safety and a qualified person should be contacted.
Use the model one last time as a critic: 'What could a reasonable recipient misunderstand? Which sentence escalates without advancing the request? What obligation am I implying?' Review those suggestions skeptically. Keep a final draft in your own voice and send it yourself. Phos is designed for drafting and thought support, not autonomous outreach.
Direct answers
Frequently asked questions
Can AI send a difficult message for me?
It should help draft and critique, not autonomously send. The user must verify facts, consequences, tone, recipient, and timing.
How can I protect the other person's privacy?
Use placeholders, summarize only necessary facts, remove private identifiers, and prefer local mode when exact sensitive wording is required.
When should I avoid AI drafting?
Avoid relying on it for emergencies, legal notices, safeguarding decisions, or high-risk workplace matters without an appropriate qualified human reviewer.