Back to blog
Client requirements, project decisions, schedules and useful links often live in WeChat. The hard part is not sending a transcript to an AI tool. It is preventing irrelevant private data, stale conclusions and missing context from going with it. A useful handoff has a defined question, the smallest necessary record set and an output that can be checked.
WeChatBridge is an open-source macOS tool built around WeChat’s native “select → merge forward → forward to another app” flow. Its documentation describes handoff to Codex, Claude Code, Obsidian or the clipboard from a user-selected export. The project states that it does not read WeChat databases, decrypt files or inject the WeChat process. Teams should still decide what their own policy permits.
Workflow: select and redact messages before handoff; keep the agent output in review until a person approves the Obsidian note.
Choose the job before choosing messages
Start with a job: client review, project decision log, daily group summary, topic research or personal note. Each one needs a different message range and output.
| Job | Select | Exclude | Deliverable |
|---|---|---|---|
| Client review | needs, commitments, objections, dates | other clients and unrelated contacts | needs, risks, open questions, next steps |
| Project catch-up | decisions, owners, deadlines, blockers | chatter and unverified claims | decision log and actions |
| Topic research | links, viewpoints, cases and sources | conclusions without sources | topic card and source list |
| Personal archive | reusable notes and files | private identity and payment data | tagged Markdown note |
If a new colleague could not understand the selected batch, add one line of background. Do not solve missing context by sending hundreds more messages.
A four-step workflow
- Select a continuous, relevant chat segment and include the few messages that explain any attachment or decision.
- Redact accounts, phone numbers, addresses, pricing floors, internal URLs and restricted attachments. Keep dates, roles and constraints needed to understand the decision.
- Add a scenario instruction.
You received a [client review] chat export. Do not add facts absent from the messages.
Return: confirmed needs; our outstanding commitments; risks and conflicts; next actions by owner and date.
Cite the message time for every conclusion. Mark uncertainty as [NEEDS CONFIRMATION].
- Review the draft before it becomes an Obsidian note. Use a searchable title such as “Client A | quote and delivery scope | 2026-09-28”, not “WeChat chat 9.28”.
Three prompts worth keeping
Project decisions
Turn this chat into a decision record. Extract only what was explicitly approved or rejected.
Output: decision / rationale / owner / due date / supporting message time.
Put proposals that were not decided under open questions.
Group daily digest
Remove greetings, repeated forwards and unrelated messages. Return completed, in progress, blocked and tomorrow's actions; keep each item below 40 words.
Abstract clients, prices, accounts and private data into roles.
Obsidian knowledge note
Turn this chat into an Obsidian Markdown note with a one-line summary, conclusions, source links, reusable method, items to verify and tags.
Keep attachment filenames. Do not turn attachment contents into unsupported conclusions.
Permissions, failure modes and a red line
WeChatBridge documents separate uses for its permissions: Accessibility activates a target app and pastes; Screen Recording reads the WeChat title bar to match a chat to a scenario. If a target app is missing or automatic paste fails, the archive remains in the clipboard for manual paste. Enable only the share entries you use and review system extensions periodically.
Never make an entire history, group or contact list the default sync source. Local processing is useful, but it does not replace data minimization, scenario instructions, human review and deletable archives. Chat is context, not a source of truth: validate high-impact claims against contracts, project documents, original links or the responsible person.
Sending selected chat to Codex reduces copying and catching up; saving a reviewed note to Obsidian makes it retrievable later. The human judgment between those two steps is the part worth keeping.
Keep evidence with every note
An AI summary alone is a poor record. Preserve the message range, timestamp, processing method, and human reviewer. If a claim cannot be tied to a message time, mark it unverified. If later messages reverse a decision, keep the change history instead of silently overwriting the earlier version.
---
topic: Client A delivery scope
source: WeChat project group; 2026-09-24 14:00 to 2026-09-28 18:00
status: awaiting owner review
reviewer: content owner
tags: [client-communication, delivery-scope]
---
# Client A | delivery scope | 2026-09-28
## Confirmed
- Deliverable: ... (evidence: 09-28 15:42)
## Unresolved
- Quote expiry: no deadline in selected messages.
## Next action
- Owner, action, due date.
The source field does not mean copying a whole conversation into every note. Where later audit is justified, retain a redacted export in a controlled folder and link to it. Otherwise keep only an authorized conclusion and minimal evidence pointer. Set a review or deletion date for retained exports.
Give Codex a narrow evidence bundle
Place selected, redacted Markdown in a dedicated folder rather than granting access to a full chat backup or home directory. First ask Codex which questions have evidence and which are unsupported. Let it draft notes for Obsidian in a staging folder; move them into the main vault only after human review. This confines a misread joke, outdated decision, or missing attachment to the draft stage.
A useful search question is precise: “What was the final delivery scope agreed in the last week of September? Show message times and conflicting statements.” A useful answer contains the conclusion, evidence pointer, and uncertainty. Fluent prose without a source location is not a reliable retrieval result.
Troubleshoot the share route in order
Confirm that WeChat shared only selected messages, then check the macOS share extension and receiving app's installation and focus. The project usage guide describes clipboard fallback. Preview the plain text before pasting and inspect names, attachments, and sensitive fields. If an image or file is missing, return to the original message for context; do not infer its contents from a filename.
Worked example: from "someone said it in the group" to a decision
Suppose the selected conversation contains: "The demo will be ready Friday," "Let's leave payments out for now," and "The client would like to see data export." A quick summary can misclassify all three as approved requirements. Before forwarding, identify who made each statement, which project it concerns, and whether a later message changed the Friday deadline. Select only those messages and nearby context, then redact private identifiers.
Ask for three output columns. Confirmed: the Friday target, with speaker, timestamp, and any later revision. Explicitly excluded: payments, but only if the decision owner agreed. Unresolved: what "see data export" means, who must answer, and by when. Put a message-time pointer on each row. If the export request is merely a suggestion, it belongs under unresolved rather than confirmed. The owner can now audit a decision by returning to the original group.
Set a retention and correction rule
Not every chat belongs in a permanent vault. One-time codes, payment details, private photos, unrelated third-party messages, and casual scheduling usually stay out of the archive. For a retained redacted export, record its owner, readers, review date, and deletion trigger. A project closing, a customer withdrawing permission, a team member leaving, or a formal document replacing the chat should prompt a review.
If Codex runs in a cloud environment, separately check workspace access and data-handling settings. "Processed locally" can describe one handoff step without describing later paste, sync, backups, or model calls. Begin with your own non-sensitive chat, inspect the resulting note, and only then expand to project material.
A week later, sample three notes. Can a teammate find each one in 30 seconds using topic or owner? Can they trace a key conclusion to a message time? Can they correct an obsolete conclusion without erasing its history? If any answer is no, fix the note structure before importing more messages.
Sources: WeChatBridge repository, usage guide.
