Claude Clone

Goal: produce the smallest possible new code that faithfully reproduces a chosen capability from a high-signal upstream repo, by reading the upstream deeply rather than guessing at the shape.

This is a meta-skill. The deep upstream archaeology is delegated to build-vs-clone (prospect mode). The dev environment is delegated to the sibling skillbox repo. This skill's job is the stitching: scoping the clone, calling the right sub-skills, extracting the irreducible core, and landing it in the target repo.

On Trigger

Start the first progress update with:

Using claude-clone ...

Then, in one short paragraph, name:

1. The upstream repo (or candidate set) being cloned
2. The specific capability (a verb + object, not a vague area)
3. The target repo the port lands in
4. Whether the clone is INLINE (one file dropped in) or MODULE (a small subdirectory with its own boundary)

If any of those four is missing, ask once before proceeding. Do not ask about anything else.

Non-Goals

• Full forks, vendor-ins, or "let's just submodule it". Use build-vs-clone in ADOPT mode for those.
• Borrowing patterns without reading code. If the user wants conceptual inspiration only, route to build-vs-clone BORROW.
• Greenfield design. If no upstream is named or implied, route to build-vs-clone to find one first.
• Style-only translations (e.g., "rewrite this Go file in Python verbatim"). This skill recreates capabilities, not files.

Modes

CLONE mode (default)

Use when the user has named both an upstream and a target. Five steps:

1. Frame — pin the four things from "On Trigger" in writing.
2. Archaeology — delegate to build-vs-clone prospect mode against the upstream. Ask it to map only the subgraph that implements the capability: entry points, core data structures, the 2–5 functions that do the real work, and the tests that pin the behavior.
3. Reduce — apply the extraction recipe in references/extraction-recipe.md to collapse the subgraph to its irreducible form.
4. Port — write the minimal new code in the target repo. Match the target's language and idioms, not the upstream's. Inline what you can, drop what the target already provides.
5. Validate — run the port inside the skillbox dev container (see references/skillbox-devcontainer.md) so the toolchain (compilers, linters, fuzzers, hyperfine, etc.) is already there. Land at least one behavioral test that pins a non-trivial case from the upstream's own tests.

SCOUT mode

Use when the user has named a capability but no upstream ("I want fuzzy matching like fzf does it, find me the right repo first"). Run build-vs-clone ecosystem-fit mode to pick the upstream, then re-enter CLONE mode at step 2.

EXTRACT mode

Use when the user already has the upstream cloned locally and wants the minimal extraction without re-discovery. Skip step 2's prospect call; go straight to reading the named files and applying the extraction recipe.

Hard Rules

• Read before writing. Do not write a single line of port code before you have named the exact upstream files and line ranges that justify it. If you cannot cite source, you are guessing.
• Smallest viable port. Every function, struct, flag, and dependency in the port must earn its place against a concrete behavior the user asked for. Cut everything else, even if "it might be useful later."
• No transitive imports. If the upstream pulls in three helper packages to do the job, prefer to inline the 20 lines you actually need over vendoring all three. Document the inlining at the top of the ported file.
• Match the target, not the upstream. Ported code obeys the target repo's language, formatter, error-handling style, and test framework. A faithful clone reads like the target wrote it themselves.
• One test that proves it. Land at least one behavioral test ported (not copied verbatim) from the upstream's own test suite. This is the contract that says "the clone is real."

Delegation Contract

claude-clone does not duplicate work that build-vs-clone already does.

Question	Owner
Is this worth cloning vs adopting vs building?	build-vs-clone
Which upstream is the right one?	build-vs-clone (SCOUT)
What does the upstream's relevant subgraph look like?	build-vs-clone (prospect)
What is the minimum code to recreate it?	claude-clone
Where does the code physically live in the target?	claude-clone
Does it pass a behavioral test?	claude-clone (in skillbox)

If build-vs-clone returns ADOPT or BORROW for the underlying question, stop and report that result instead of forcing a clone.

Dev Container

All build/test/bench work for the port runs inside the sibling skillbox repo's main workspace container. See references/skillbox-devcontainer.md for the exact entry commands and the assumed package surface. Do not install language toolchains on the host to support a clone — that is the entire reason skillbox exists.

Output Shape

When the clone is complete, report:

1. Upstream + commit SHA read during archaeology
2. Files cited (path + line range) that the port is derived from
3. New files in the target with line counts
4. Behavioral test(s) added and what upstream test they mirror
5. Deliberately omitted upstream features and why
6. skillbox commands the user can rerun to re-validate