Claude Tag Closes 65% of the Claude Code Team's PRs, and the System Prompt Was Cut by 80%

Cat Wu and Thariq Shihipar of Anthropic's Claude Code team said at the AI Engineer World's Fair conference in early July 2026 that their internal tool Claude Tag already closes 65% of the team's engineering PRs, and that the Claude Code system prompt was cut by 80%.

What Claude Tag Can Do

Claude Tag is Claude embedded directly into team tools; the Slack version launched in early July 2026. Its main difference from Claude Code is that it is multiplayer and proactive by default: once you add it to a Slack channel, the whole team edits a single PR together, while the bot itself watches bug reports and opens fixes without being manually invoked.

According to Cat Wu, the internal version of Claude Tag already closes 65% of the Claude Code team's product-engineering PRs, which is more than half of all PRs. Claude Code remains the place for the most complex interactive tasks, while Claude Tag works proactively in the background.

• Claude Tag is a collaborative Slack integration that launched in early July 2026
• It closes 65% of the Claude Code team's product-engineering PRs
• It is multiplayer: the whole team works on a single PR
• Team memory is stored in one markdown file per channel
• It runs on top of auto mode, the same safety mechanism as Claude Code

Why the System Prompt Was Cut by 80%

The Claude Code system prompt was cut by 80% because the new models, Fable 5 and Opus 4.8, perform worse with a large number of examples and rigid prohibitions. According to Thariq Shihipar, the team had been "over-constraining" Claude: removing examples produced a noticeable improvement because the model turned out to be more creative than the templates it had been fed.

Anthropic now maintains a different system prompt for each model: only frontier models got the 80% token reduction, while older models keep the full prompt. Lists of "don't do X, don't do Y" also hurt: a strong prohibition impulse conflicts with the user's later instructions and confuses the model.

OpenAI offers similar advice: by its data, leaner system prompts raised scores by 10–15% in a sample of internal co-agent runs while cutting tokens by 41–66%.

How Safe Is Auto Mode?

According to the team, auto mode reduces the main risks, prompt injection and data leakage, more effectively than the average human reviewer. Almost all Anthropic employees use it internally, and the mode has been battle-tested since January 2026; it has been publicly available since March 24, 2026.

Here is how it works: on every action Claude takes, a separate Sonnet-based classifier evaluates the tool call and the conversation context, including dynamic permissions such as "push to GitHub" or "don't push." The team ran thousands of evals and hired external red teams to try to provoke harmful actions and, by its own account, fixed every problem they found.

"Almost everyone at Anthropic uses auto mode.

It's the best way to run long tasks in Claude Code while staying safe." — Cat Wu, Claude Code team, Anthropic

Auto mode is what makes Claude Tag possible: the bot reads user feedback in channels, which opens up many attack vectors, and without layered "Swiss cheese" protection, launching such a Slack bot would be dangerous.

What This Means

Anthropic is increasingly taking humans out of the review loop and relying on the models' own judgment. Shorter prompts, proactive agents and auto mode point in one direction: development teams are shifting from manually controlling every step to delegating routine work and taking on more ambitious tasks.

"The only way to make up for this sense of loss is to become more ambitious." —

Thariq Shihipar, Claude Code team, Anthropic

Frequently Asked Questions

What Is Claude Tag?

Claude Tag is Claude embedded in team tools; the Slack version launched in early July 2026. It is multiplayer and proactive: it watches bug reports in the channel, opens PRs on its own and tags the engineer who last touched that part of the code.

Why Was the Claude Code System Prompt Cut by 80%?

Because the new models, Fable 5 and Opus 4.8, perform worse with a large number of examples and rigid "don't do X" prohibitions. Anthropic removed the excess examples and restrictions; the full prompt is kept for older models.

How Safe Is Auto Mode?

According to the team, the main risks, prompt injection and data leaks, are lower than with the average human reviewer. Anthropic has used auto mode internally since January 2026 and has run thousands of evals with external red teams.