diff --git a/CLAUDE.md b/CLAUDE.md
new file mode 100644
index 0000000000..41edef2cfd
--- /dev/null
+++ b/CLAUDE.md
@@ -0,0 +1,147 @@
+# CLAUDE.md
+
+Code in this repository follows strict conventions. Read
+`docs/coding-guidelines.md` if unfamiliar with any rule below.
+
+## What is mold
+
+mold is a fast drop-in replacement for GNU ld. It links ELF object files
+into executables and shared libraries. It is a cross-linker — it can
+target any architecture from any host. The man page source is
+`docs/mold.md`; do not edit `docs/mold.1` (auto-generated).
+
+## Build and Test
+
+```bash
+cmake -DCMAKE_BUILD_TYPE=Debug -DCMAKE_CXX_COMPILER=clang++ -B build
+cmake --build build -j$(nproc)
+cd build && ctest --output-on-failure -j$(nproc)
+```
+
+Quick smoke test — verify mold can link a trivial program:
+```bash
+echo 'int main() {}' | cc -B. -o /dev/null -xc -
+```
+
+## Code Style
+
+- **Integers**: Use `i64` for everything. Use `i32` only when allocating
+  millions of the same object. Use `ub32`/`ul32`/`il32` from
+  `lib/integers.h` for file I/O fields (endian-safe, alignment-safe).
+- **`auto`**: Use only for:
+  - Lambdas: `auto fn = [&](i64 val) { ... };`
+  - Structured bindings: `auto [it, inserted] = map.insert({key, val});`
+  - Verbose types: `auto flags = std::regex_constants::optimize | ...;`
+  Never for simple variable declarations — always write the explicit type.
+- **`std::tie` vs `auto [x, y]`**: Use `std::tie` when the variables are
+  already declared. Use `auto [x, y]` for new declarations.
+- **`constexpr`**: Use `if constexpr` for template metaprogramming (target
+  selection, arch-specific code paths). Don't add `constexpr` to functions
+  unless required.
+- **No exceptions**: The build uses `-fno-exceptions`. Use
+  `Fatal(ctx)` / `Error(ctx)` / `Warn(ctx)` — RAII classes that print
+  and exit in the destructor. Fatal calls `_exit(1)`.
+- **No deep inheritance**: Class hierarchies are at most 2 levels deep.
+  Prefer composition.
+- **Variable names**: Short — `ctx` (Context), `mf` (MappedFile),
+  `sym` (Symbol), `esym` (ElfSym), `loc` (relocation target address).
+- **Helper lambdas**: Define locally for repeated operations within a
+  function (e.g., `auto check = [&](i64 val, i64 lo, i64 hi) { ... };`
+  for relocation range checks).
+- **`append()`**: Use `append(vec, x)` from `lib/lib.h` instead of
+  manual loops to append vectors.
+- **Direct field access**: Prefer `isec.name` over `isec.name()` when
+  the field is directly accessible. Remove unnecessary local variables
+  that just alias a field.
+- **Range-based for over index loop**: Prefer `for (Symbol<E> *sym : vec)`
+  over `for (i64 i = 0; i < vec.size(); i++)` when the index isn't needed.
+- **Semantic names over raw checks**: Prefer `ctx.gnu_debuglink` over
+  `!ctx.arg.separate_debug_file.empty()`. Use meaningful boolean fields
+  instead of checking argument emptiness.
+- **Simplify**: The project values simplification — removing redundant
+  code, replacing index loops with range-for, extracting complex lambdas
+  into named functions, using semantic names. Simplification should be
+  a standalone commit, not mixed with feature work.
+
+## Architecture
+
+- **Explicit template instantiation**: Every `.cc` file ends with
+  `template class Foo<E>;` for the target type `E = MOLD_TARGET`.
+  This keeps compile times fast. Don't use header-only templates.
+- **`Context<E>` is the god object**: Almost every function takes
+  `Context<E> &ctx`. It owns all linker state.
+- **`namespace mold`**: All code lives here. No nested namespaces.
+- **Include `mold.h`**: Most `.cc` files include only `mold.h`.
+  Put `#include "mold.h"` first, then system headers.
+- **ELF types**: Use types from `src/elf.h` (e.g. `Elf64_Sym`,
+  `Elf64_Rela`). The template alias `U32<E>` selects endian-correct
+  32-bit integers automatically.
+- **Determinism**: mold's output must be bit-for-bit identical given the
+  same inputs. No host-specific defaults. No non-deterministic iteration
+  (hash tables, thread scheduling) in output-affecting code.
+- **Linker scripts**: Intentionally minimal support — just enough to
+  read `/usr/lib/x86_64-linux-gnu/libc.so`. No plans to add more.
+- **Parallelism**: Uses TBB (`tbb::parallel_for_each`,
+  `tbb::parallel_sort`). Use simple `for` loops when parallelism isn't
+  needed. Don't parallelize just because you can — measure first.
+  Avoid per-operation overhead (e.g., hash table insert per relocation).
+- **Comments**: Explain WHY, not WHAT. Use full sentences. Detailed
+  comments for complex logic (e.g., symbol resolution, .eh_frame parsing).
+  No section separator lines (`// ====...`, `// ----...`).
+
+## Writing a Test
+
+Tests are shell scripts in `test/`. Every test sources `common.inc`:
+
+```bash
+#!/bin/bash
+. $(dirname $0)/common.inc
+
+cat <<EOF | $CC -fPIC -c -o $t/a.o -xc -
+int foo = 42;
+EOF
+
+cat <<EOF | $CC -c -o $t/b.o -xc -
+extern int foo;
+int main() { return foo != 42; }
+EOF
+
+$CC -B. -o $t/exe $t/a.o $t/b.o
+$t/exe
+```
+
+- `$t` is a temp directory auto-created by `common.inc`.
+- `$CC` / `$CXX` are set by `common.inc`.
+- Use `not` to test expected failures.
+- Use `skip` for arch-specific exclusions (`[ $MACHINE = aarch64 ] && skip`).
+
+## Commits
+
+- Short imperative: "Fix overflow in relocation", "Add SHT_NOTE support".
+- No conventional commit prefixes (no `feat:`, `fix:`, `chore:`).
+- No WIP commits. No draft PRs.
+- No non-functional / stylistic changes — these are explicitly rejected.
+- One logical change per PR. The maintainer splits and refactors after merge.
+- Separate concerns: don't mix adding a feature with changing output format.
+- Commit messages explain the problem, root cause, and fix — especially
+  for bug fixes. Reference issue numbers when applicable.
+
+## What NOT to Do
+
+- Don't use `auto` for simple variable declarations — only for lambdas,
+  structured bindings, and verbose types.
+- Don't introduce `std::ranges` or `std::views`.
+- Don't add `constexpr` to functions that don't need it.
+- Don't use exceptions or `throw`.
+- Don't refactor existing code for style — only functional changes.
+- Don't use `int`, `unsigned`, `int32_t`, `uint32_t` for local integers.
+- Don't add CMake options (`--enable-foo` / `--disable-foo`).
+- Don't edit `docs/mold.1` directly — it's auto-generated from `docs/mold.md`.
+- Don't propose optimizations without measuring memory impact —
+  optimizations where memory cost outweighs speed gain are reverted.
+- Don't parallelize code unnecessarily — simple `for` loops are fine
+  when parallelism isn't needed.
+- Don't use per-operation overhead (e.g., hash table insert per
+  relocation) — solutions that add O(n) allocations are rejected.
+- Don't extract complex lambdas into standalone functions unless the
+  lambda is too complex to read inline.