uygnoey Slack thread delegation
Claude Tag puts Claude inside chat tools like Slack. You tag it in a thread and it completes tasks using the conversation's context: it can follow the conversation, use standing instructions you have given it, and decide when to participate.
The leverage is not that Claude writes faster. It is that the raw material — a 15-message thread, 120 scattered reports, a marketing asset and the sources that back it — is already in chat, and gathering it is the part that costs a human hours. Delegate the synthesis; keep the judgment.
This skill covers three delegation shapes drawn from how teams at Anthropic use the feature, and the verification loop that makes each one safe to trust.
Instructions
1. Decide what to delegate
Delegate work where the inputs live in chat and the cost is information gathering:
- Synthesis — a thread contains a decision, a spec, or an argument that someone now needs as a document.
- Consolidation — the same class of request or report is scattered across channels and nobody has the list.
- First-pass review — a specialist reviews every item, and most items are fine.
Keep for yourself: which claims need sign-off, what the final call is, and whether the output is right. In every workflow below, the human focuses on judgment, verification, and decision-making while Claude handles information synthesis and processing.
2. Prompt for the shape of the work
The three shapes need different prompts. See references/prompt-patterns.md for each pattern in full.
- Synthesis — point at the thread and name the artifact and its requester. A short prompt is enough because the thread is the specification.
- Consolidation — state three things: the search targets, what counts as a match, and the desired output format with an example row. Without the match definition Claude cannot dedupe; without the format example you get prose instead of a list you can act on.
- Review — set the rules once as standing instructions for a dedicated channel rather than repeating them per asset.
3. Verify before you circulate
Do not treat the first draft as the deliverable. Ask Claude to check its own factual accuracy, and have it separate claims that are verified from claims that need sign-off from the owner. Then feed it the official resources for the unverified parts and have it rewrite those sections.
In the marketing one-pager case this took four versions before the draft went to the product lead — each round narrowing what still needed a human. Budget for that loop; it is where the quality comes from.
4. Give a recurring workflow its own channel and standing instructions
For work that repeats, create a dedicated channel and write the rules once. Grant the access the work actually needs — company chat, an internal knowledge index, the public web — and no more.
Start from templates/review-channel-instructions.md and adapt the rules to your domain.
5. Make the instructions improve themselves
When you correct Claude, ask it to carry the correction forward as a standing instruction rather than as a one-off fix. In the legal review channel, "verify flagged claims in real time when possible" started as feedback on a single review and became the rule for all future ones.
Add a recurring review of the instructions themselves: a weekly routine where Claude reads the feedback it received that week and proposes instruction updates for approval. Approval stays human.
6. Expect scoped access, and notice the gaps
Claude Tag's access is deliberately scoped — only the channels and documents it has been granted. When it lacks the access a task needs, it tells you. Treat that message as a signal about the task, not only about permissions: work that keeps hitting inaccessible sources may belong elsewhere.
Examples
Full case studies with the prompts, timings, and outputs reported in the source post are in examples/tag-workflows.md. In brief:
Thread to document. A product marketer tagged Claude on a 15+ message thread with roughly "go through this Slack thread and come up with a one pager that [the requester] is asking for." Claude produced a two-page draft in two minutes covering feature description, business case, implementation details, and an appendix. Four verification rounds later it went to the product lead — 45 minutes total.
Scattered requests to one list. A product strategy team member asked Claude to find every place a feature had been requested and who had asked on a customer's behalf. Claude ran roughly 20 search variants across channels, deduplicated the results, and returned about 24 accounts with requester handles, teams, account names, and source links in about 26 minutes. The weekly version of the same task turned about 120 raw reports into 23 open and 14 resolved issues organized by product area in 50 minutes — work estimated at a full week by hand.
Review queue triage. A product counsel set up a channel where Claude reviews marketing assets before legal review: spotting unsubstantiated claims, verifying factual statements, flagging issues with specific remediation instructions, and working with requesters directly. On one newsletter it flagged three items and resolved one itself after finding support in internal documents. Turnaround went from a day to 30 minutes per asset.
Setup
Claude Tag is available on Team and Enterprise plans through Anthropic's first-party service, set
up at claude.ai/admin-settings/claude-tag. It is in public beta.