Back to the journal

Claude writes over 80% of Anthropic’s code. Written instructions make the difference

From Anthropic’s own reports: Claude writes over 80% of the code it merges, but the real gain comes from what teams write for it: CLAUDE.md, skills and commands. In data analytics, accuracy never passed 21% without skills and topped 95% with them. And where it still falls short.

Parsa5436
6 min read

In When AI builds itself, Anthropic says that as of May 2026 Claude wrote more than 80% of the code merged into its codebase; before Claude Code launched in February 2025, that share was in the low single digits. In an August 2025 survey of 132 of its engineers and researchers, the same company heard that most of them could fully delegate only 0–20% of their work to Claude.

The two numbers do not contradict each other: one is a share of code, the other a share of work handed over without supervision, and they are nine months apart. Side by side, though, they show where the real work stayed: in the text people write for Claude, and in reviewing what it hands back. Anthropic’s team-by-team reports show the same thing in detail.

The numbers, with their own caveats

  • Over 80% of merged code was written by Claude as of May 2026. Anthropic’s leadership had earlier said 90% or more; the report gives the more conservative number because its attribution pipeline has gaps. (When AI builds itself)
  • Eight times as much code per day per engineer in the second quarter of 2026, compared with 2024. The report itself says this almost certainly overstates the real productivity gain, because line counts do not measure quality. (Same source)
  • In the August 2025 survey, employees said they use Claude in about 60% of their work and that their productivity rose 50% (20% a year earlier). 27% of the work done with Claude would not have been done at all without it. The most common daily use was debugging (55%), then understanding code (42%). (How AI is transforming work at Anthropic)
  • Across about 200,000 internal Claude Code transcripts from February to August 2025, the maximum number of consecutive tool calls per transcript rose from 9.8 to 21.2, and human turns fell from 6.2 to 4.1. (Same source)

The common pattern: what teams write for Claude

Data analytics. Anthropic says Claude answers 95% of its business analytics queries, with about 95% accuracy in aggregate. The same post says that without skills, Claude’s accuracy on their own evals did not exceed 21%. On those evals, the gap between the two numbers is skills: business knowledge and analysis routines written down, alongside a semantic layer and validation. The post also warns that this accuracy means no obvious gaps were found, not that a wrong answer never comes back. (How Anthropic enables self-service data analytics with Claude)

RL engineering. For Claude’s repeated tool-calling mistakes, the team added a line to CLAUDE.md along these lines: “run pytest not run and don't cd unnecessarily - just use the right path”. According to the report, consistency improved significantly. (How Anthropic teams use Claude Code, PDF)

Security engineering. This team wrote half of all the custom slash commands in Anthropic’s monorepo. (Same PDF)

The Claude Code team itself. For prototyping, they turn on auto-accept mode and leave Claude in a loop of writing, running tests and iterating; then they review and finish a solution that is about 80% complete. Vim key bindings in Claude Code were built this way, and about 70% of the final implementation was Claude’s autonomous work. Their condition: start from a clean git state and commit checkpoints regularly, so that rolling back is easy if Claude goes off track. (Same PDF)

Data science and ML. They call it a slot machine: commit your state, let Claude work for half an hour, then either accept the result or start over. In their experience, starting fresh often succeeds more than fixing Claude’s mistakes. (Same PDF)

Outside engineering

  • The legal team prototyped a “phone tree” that routes each person to the right lawyer at Anthropic. In one hour, a team member built an app that suggests replies through speech-to-text and speaks them aloud, for a family member who has difficulty speaking because of a medical condition.
  • Growth marketing built a Figma plugin that generates up to 100 variations of an ad, half a second per batch; writing ad copy went from two hours to 15 minutes.
  • Product designers say Figma and Claude Code are both open 80% of the time. Removing the phrase “research preview” from the entire codebase, which would have taken a week of back-and-forth, was done in two 30-minute calls.
  • When Kubernetes pods stopped scheduling, data infrastructure gave Claude Code screenshots of the dashboard; Claude walked them menu by menu to a pod IP exhaustion warning and supplied the commands to create a new IP pool, with no networking specialist needed.

All four come from the same PDF; a summary is also on the Claude blog.

Where it does not work

  • The RL team says Claude gets a small-to-medium PR right on the first attempt only about a third of the time. Their advice: try one-shot first, and if it fails, switch to guided work. (PDF)
  • Employees worry about their skills atrophying. The 2025 report points to a paradox of supervision: overseeing Claude takes the very skill that delegating wears down. Some also say they now work more with Claude than with their colleagues. (How AI is transforming work at Anthropic)
  • The 2026 report says many staff believed Claude-written code was still lower in quality than Anthropic’s human-written code in late 2025; there is no full consensus. (When AI builds itself)
  • A retrospective analysis found that if Claude had automatically reviewed every code change, roughly a third of the bugs behind past claude.ai incidents would have been caught before production. The rest would not. (Same source)

In our own repository

This site’s repository is private, so what follows is owner-reported.

Our CLAUDE.md has a section called Traps already paid for: each line is a trap we fell into once and do not want Claude to fall into again. For example, do not pipe an Nx run into tail in a script, because the daemon inherits the pipe and the command never returns. That is the RL team’s pytest line, in another project. Another rule: motion code is never written from memory, and Claude reads the GSAP skill first. To date, 66 of this repository’s 71 commits carry Claude’s co-author line.

This post, and this translation, were created as drafts by Claude Code from an idea on our content board, through our own CMS’s MCP server. Publishing them is up to a person who reads them.

Sources

Drafted with AI, reviewed by [name]