Claude Tag закрывает 65% PR команды Claude Code, а системный промпт ужали на 80%

CatWu and Thariq Shihipar from the Claude Code team at Anthropic, speaking at the AI Engineer World's Fair conference in early July 2026, revealed that their internal tool ClaudeTag already closes 65% of the team's engineering PRs, and the Claude Code system prompt has been cut by 80%.

What ClaudeTag can do

ClaudeTag is Claude built directly into team tools; it launched in Slack in early July 2026. The main difference from Claude Code is that it is multiplayer and proactive by default: once added to a Slack channel, the whole team works on a single PR together, and the bot itself watches for bug reports and opens fixes without being manually invoked.

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

*ClaudeTag is a collaborative Slack integration, launched in early July 2026

*Closes 65% of the Claude Code team's product-engineering PRs

*Multiplayer: the whole team works on a single PR

*Team memory is stored in a markdown file for each channel

*Runs on top of automode — 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 hard restrictions. 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 given.

Anthropic now maintains a different system prompt for each model: only the frontier models got the 80% token cut, while older models kept 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: according to its data, "leaner" system prompts in a sample of internal co-agent runs raised scores by 10–15% while cutting tokens by 41–66%.

How safe is automode

According to the team, automode reduces the main risks — prompt injection and data leaks — more than the average human reviewer does. Inside Anthropic, almost all employees use it, and the mode itself has been in testing since January 2026; it has been publicly available since March 24, 2026.

Here's how it works: for every action Claude takes, a separate Sonnet-based classifier evaluates the tool call and the conversation context, including dynamic permissions — for example, "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 says it fixed every issue found.

"Practically everyone at Anthropic uses automode.

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

It is precisely automode that makes ClaudeTag possible: the bot reads user feedback in channels, which opens up a huge number of attack vectors, and without multi-layered "Swiss cheese" defenses, running a Slack bot like this would be dangerous.

What this means

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

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

Thariq Shihipar, Claude Code team, Anthropic

Frequently Asked Questions

What is ClaudeTag?

ClaudeTag is Claude built into team tools; it launched in Slack in early July 2026. It is multiplayer and proactive: it watches for bug reports in a channel, opens PRs on its own, and tags the engineer who last touched that piece of 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 hard "don't do X" restrictions. Anthropic removed the unnecessary examples and restrictions; the full prompt was kept for older models.

How safe is automode?

According to the team, for the main risks — prompt injection and data leaks — it performs better than the average human reviewer. Inside Anthropic, automode has been in use since January 2026, and thousands of evals have been run with external red teams.