Hit Send

Before you send, let AI read the message as the recipient.

中文说明 · Examples · Install · Use it · Optional H5 interface · Latest release

Hit Send is an open-source Agent Skill for short workplace messages. It checks whether a draft is clear, actionable, appropriately direct, and safe to send, then makes the smallest useful revision without changing the writer's facts, position, ownership, deadline, or voice.

It is designed for messages such as requests, reminders, status updates, disagreements, risk alerts, handoffs, short emails, and replies based on selected conversation context. It never sends the message for you.

Examples by scenario

Hit Send does not force every message into the same polite template. It keeps the writer's intent and makes the smallest change needed for the recipient to understand or respond.

Following up on progress

Draft

Why isn't this done yet? We need it tomorrow. Please get it sorted ASAP.

Suggested

We need this tomorrow. Where does it stand, and when do you expect it to be ready? Please flag it now if that timing is at risk.

What changed: The urgency stays, while blame is replaced by questions the recipient can answer directly.

Challenging a proposal

Draft

This approach doesn't make sense. We tried it before and it failed. Why are we doing it again?

Suggested

I don't think we should move forward with this approach as-is. We tried it before and it didn't work. What's different this time?

What changed: The disagreement and supporting fact remain, without turning the message into a judgment about the people proposing it.

Raising a delivery risk

Draft

The API is still unstable, so we'll probably miss Friday. The product team needs to decide today what to cut.

Suggested

The API is still unstable, which puts Friday's release at risk. The product team needs to decide today what can come out of scope.

What changed: The uncertainty, cause, decision owner, and timing are easier to scan without overstating the outcome.

Setting a boundary

Draft

I can help this time, but I can't keep taking last-minute requests. You always send them too late. Next time, send them earlier.

Suggested

I can help this time, but I can't take last-minute requests as the default. For future requests, please send them earlier so I can confirm whether I can support the timing.

What changed: The boundary and request stay clear, while the absolute accusation is removed.

The default is one recommended version. A second version appears only when it represents a genuinely different strategy, such as more direct versus more relationship-sensitive wording. Actual wording depends on the context provided; Hit Send does not invent missing facts.

Install

Ask your agent

Send this instruction to an Agent Skills-compatible agent:

Install Hit Send from https://github.com/BobbyYue/hit-send,
using the complete subdirectory skills/hit-send. Register it in the agent's normal
Skill directory, then confirm that SKILL.md and its references can be loaded.
Do not overwrite an existing installation without asking.

Clone manually

git clone --depth 1 https://github.com/BobbyYue/hit-send.git
cp -R ./hit-send/skills/hit-send "<YOUR_AGENT_SKILLS_DIR>/hit-send"

Copy the complete folder, not only SKILL.md, and reload the agent if required by the host.

Web or desktop import

Download the latest release, extract it, and import skills/hit-send through the client's Skill interface. Do not import the repository root unless that client explicitly supports repository subpaths.

Use it

Ask naturally:

Use Hit Send to review this before I send it. Preserve my intent and only change
what materially affects clarity, actionability, or tone:
<paste message>

Other useful prompts:

Keep my disagreement, but make sure it does not sound like a personal accusation.

Draft a reply from the two selected messages. Do not add any status, decisions,
or commitments that are not in the context.

Make this risk update clear. Check the impact, decision owner, and next update
time, but do not invent missing details.

What it protects

• Meaning: preserves facts, names, numbers, dates, scope, uncertainty, ownership, commitments, and the intended ask.
• Voice: keeps natural workplace language instead of turning every message into formal corporate prose.
• Actionability: makes unclear requests easier to answer without inventing deadlines or commitments.
• Relationship safety: keeps legitimate disagreement and emotion while removing sarcasm, mind-reading, and unsupported personal judgment.
• User control: returns a preview for explicit review or copying; it never posts, replies, or auto-sends.
• Privacy boundary: processes only text the user supplies or explicitly selects and does not require conversation-history access.

Optional H5 interface

The repository includes an optional browser and Feishu/Lark H5 interface under apps/feishu-h5. It supports:

• local browser use with an authenticated Codex CLI;
• a private model endpoint implementing the documented JSON contract;
• Feishu/Lark chat-box menu and message-action entry points;
• preview and copy without calling a message-send API.

Run a local preview:

cd apps/feishu-h5
npm test
HIT_SEND_PROVIDER=mock npm start

Open http://127.0.0.1:8787. See the H5 deployment guide for model and Feishu/Lark configuration.

Development

Version 0.6.0 adds a host delivery entrypoint that rechecks the actual evidence and current caller context before returning content. It also separates editing instructions from copyable messages in the existing review. Genuine recipient requests and useful context remain. Applications still need explicit integration; the response JSON, lightweight self-check and no-send rule are unchanged. No native/H5 deployment is included. The new calibration utility accounts for errors and missing runs without claiming measured human outcomes.

Version 0.5.0 adds checked local delivery: scripts/deliver.py rechecks the recommendation and alternatives against the supplied source and returns the exact response bytes. Missing, pending, failed or stale reviews block that path. Validation and fingerprinting use the same byte buffers. It remains a lightweight self-check, never an independent review; the response JSON and no-send rule are unchanged. Native/H5 applications still need explicit integration, and no application deployment is included.

Version 0.4.0 identifies the supplied recipient context and writer's stance before rewriting. Existing reviews explain what misunderstanding or reply burden an edit removes; clearer or shorter wording alone is not sufficient. Clear messages stay unchanged, with intent, urgency, uncertainty and useful emotion preserved. Contextual cases and complete examples are loaded only when needed. Review version 2 and the response JSON are unchanged. The optional H5 and native applications are not newly integrated with this check. See the upgrade notes.

Run the public-package checks:

python3 scripts/validate_repo.py
npm test --prefix apps/feishu-h5
python3 skills/hit-send/scripts/validate_response.py skills/hit-send/evals/valid-response.json --schema-only
python3 skills/hit-send/scripts/validate_behavior_cases.py skills/hit-send/evals/behavior-cases.json
python3 -B -m unittest discover -s skills/hit-send/tests -v

Repository layout:

skills/hit-send/   installable Agent Skill
apps/feishu-h5/    optional browser and Feishu/Lark interface
.github/workflows/ public CI checks

Privacy and limitations

• Do not put credentials, private conversation exports, internal document links, or real sensitive messages in issues or test cases.
• Local Codex mode requires an installed and authenticated Codex CLI. Other hosts may use only the Skill folder.
• A local H5 URL is available only on that computer. Shared use requires an HTTPS deployment and an independently secured model endpoint.
• Hit Send improves message quality; it cannot determine facts or relationship context that the writer did not provide.

See CONTRIBUTING.md and SECURITY.md. Licensed under MIT.