email-sorting
Skill2 starsChanged 31 days agoMITView on GitHub
Customize and use a human-reviewed email sorting system for Gmail, Google Workspace, Outlook, Microsoft 365 or another connected mailbox. Use for first-time setup, folder or label recommendations from a read-only email sample, inbox triage, and applying approved sorting batches. Not for sending, replying to or deleting email.
In this repository10 other files it gives its agents.
- maya-starter/CLAUDE.mdCLAUDE.md
- facebook-personal-browserSkill
- gam-campaign-checkupSkill
- generate-meSkill
- gpt-imageSkill
- grok-browserSkill
- last30days-plusSkill
- muse-browserSkill
- perplexity-browserSkill
- wiki-llmSkill
The file
We split this skill into the rules it gives an agent
8 rules. About 0 are common, 8 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈8
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 12
Install or customize: No approved private profile exists, the account changed, or the member asks to redesign the system.
- Distinctive: line 14
Apply approved rows: Use the exact reviewed proposal and the member's explicit selection.
- Distinctive: line 41
Read only the relevant provider guide: Gmail, Microsoft, or other providers.
- Distinctive: line 45
Check human signals and the rules the member actually approved.
- Distinctive: line 47
Read relevant thread context, including later replies and sent responses within the authorized scope.
- Distinctive: line 48
If completion is uncertain, inspect related mail or business records only when that source is separately authorized.
- Distinctive: line 58
Use the fields in review-and-apply.md.
- Distinctive: line 71
Read back after each change.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Email Sorting
Help the business owner work in their existing mailbox and talk directly to their AI assistant. Learn how they work before recommending another structure. The first run is a read-only consultation. Setup and sorting are separate approvals.
Choose the mode
- Install or customize: No approved private profile exists, the account changed, or the member asks to redesign the system. Follow SETUP.md.
- Propose a sort: Load the approved profile and provider reference. Read a bounded batch, normally 15 conversations, and show recommendations. Write nothing to the mailbox.
- Apply approved rows: Use the exact reviewed proposal and the member's explicit selection. Follow the apply procedure below.
Installing the files does not connect an email account or grant write permission. Use the member's existing authorized mail tools. Follow the host's tool and account rules. If only read access is available, complete the recommendations and give a manual setup checklist. Never claim setup happened without readback.
Private configuration
Use a private workspace outside the installed skill and outside any Git repository. Keep its path in the member's private project instructions, not in the distributed package. Read profile.json there at the start of each run. If the path is unknown, ask for it; do not search unrelated private folders.
The blank profile template is a starting point, not an approved policy. It deliberately contains no account, people, sender lists or industry rules. Store customized profiles, email samples, proposals, feedback and receipts only in the private workspace. Minimize copied email content and use restricted local file permissions where supported. Never add these files to a shared skill, repository or member download.
The starting framework
Recommend these four work stages unless the member's existing structure works better:
| Stage | Meaning |
| 10 Action | A next step belongs to this business or person |
| 20 Waiting | A concrete response or event is expected from somebody else |
| 30 Reading | Optional reading or useful FYI with no required action |
| 90 Filed | Finished work or reference material with no unresolved obligation |
Numbers leave room for future additions. Names, numbering, whether Reading is useful, and the filing hierarchy are customizable. Keep work stage, topic, owner and follow-up date separate. Topics describe the work; they do not prove it is complete. Start with a few evidenced topics, not one folder per sender. Only create reference subfolders the member approves.
- Gmail uses user labels. Inbox membership is a separate label; applying a work-stage label does not itself archive a message.
- Microsoft mail uses folders for location and categories for topics or owners. Show explicitly whether a recommendation moves mail out of Inbox.
- Existing labels, categories, flags, stars and folders belong to the member. Map and reuse them where suitable. Remove only sorting labels/categories the member explicitly designated as managed by this skill.
Read only the relevant provider guide: Gmail, Microsoft, or other providers.
Decide from evidence
- Check human signals and the rules the member actually approved. Confirm what stars, flags and categories mean; they are not universal stage indicators.
- Apply approved narrow sender rules. A marketing domain can also send invoices or security notices, so a sender alone is usually insufficient.
- Read relevant thread context, including later replies and sent responses within the authorized scope. One message often cannot establish the current work stage.
- If completion is uncertain, inspect related mail or business records only when that source is separately authorized. Never import another company's memory or assume a shared memory service exists.
- Recommend stage, topic, owner and any follow-up date with a short reason and source reference. Unknown ownership stays unassigned and is surfaced. A proposed date is not a promised deadline; distinguish stated dates from suggestions and confirm timezone.
- Escalate critical items and ambiguity. Do not clear an obligation simply because it is old, read, acknowledged, or past its deadline. A later reply can reopen completed work. Missing evidence means review, not Filed.
Examples of items for human handling include suspected fraud, account-access changes, payment-detail changes, legal demands and urgent customer complaints. Let the business customize these criteria. Critical rows stay visible and are excluded from ordinary bulk application. The human handles them separately.
Use the current agent or an available economical model for routine topic suggestions. Escalate ambiguous work-stage decisions with better evidence or a stronger available model. Do not hardcode a vendor, model version, paid API or confidence threshold from somebody else's benchmark. If an optional specialist classifier is used, obtain consent for the data destination, keep credentials private and validate on fresh mail. Confidence alone does not authorize a write.
Show a reviewable proposal
Use the fields in review-and-apply.md. Show:
Row | Short subject | Current location | Proposed stage and destination | Topic | Owner | Reason / evidence | Needs review
Include the target provider and mailbox, proposal ID/version, message versus conversation scope, and the exact effects of the changes, including whether Inbox membership changes. Save a snapshot for each message and the latest thread context. If a conversation contains different obligations, split it into message-level rows or leave it for review. Never archive a whole conversation to hide an unresolved message.
Ask the member which rows to change and what you misunderstood. Revise the profile or proposal from feedback, then show the revised version. An answer about naming folders is not approval to move messages. Approval for one batch does not cover future batches. If approval is partial, apply only the specified unchanged rows.
Apply and verify
- Confirm the approved profile version, proposal version, account and exact rows. If the proposal changed after approval, obtain approval for the changed rows again.
- Re-read the message and relevant thread immediately before writing. If content, thread membership, location, labels/categories, human signals, read state or the provider's version changed, skip the affected row and refresh its proposal. Use conditional writes when the tool supports them. If it cannot guard against concurrent changes, disclose that limitation and keep the pilot small.
- Apply only the reviewed folder/label/category changes with the authorized provider tool. Preserve unrelated categories/labels, read state and flags. Do not set follow-up flags or dates unless those exact additional changes were reviewed and approved.
- Read back after each change. Record original state, operations, returned/new IDs and verified final state in a private receipt. A partial write is a partial result, not a success.
- On a failure or mismatch, stop the remaining batch. Reconcile the actual state before resuming; do not replay uncertain writes. Explain what changed and what needs attention.
- Return counts for verified, skipped, failed and review-only rows, plus a private receipt path. Offer a reversal plan from the recorded state when requested, but recheck current state and obtain approval before reversing.
This skill never sends, drafts, replies, deletes, trashes, marks read/unread, changes access permissions, or creates forwarding. It runs on demand. Installing it does not authorize server rules, automatic filing, scheduled jobs or notifications. Those are separate future work.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

