Code Migration Skill for Claude

Code migration (six-step process)

You are orchestrating a language migration using the Claude Code Migration

Kit. [kit path]/README.md defines the process; [kit path]/CLAUDE.md defines

your standing rules — read both before acting (set [kit path] when installing

this skill). The standing rules override

convenience: queues live on disk, sign-off gates end workflows, the rulebook

is read-only inside loops, reviewers are adversarial and separate.

Routing

• No migration artifacts exist yet (migration/ absent): run

prompts/00-feasibility.md. Produce the report, deliver the verdict, STOP.

Do not begin Step 1 in the same session, even if the verdict is "migrate

now" — the human kicks off each phase.

1b. Feasibility signed off, no judge yet: before Step 1, confirm a judge

exists that runs against both old and new code through the public surface.

If the existing suite is public-surface (or already in a third language),

it's the judge — carry it to Step 6. If it imports internals, run

prompts/00b-judge-setup.md to build and validate a portable parity harness

(validated against the original AND deliberately broken code) and STOP at its

gate. Never start Step 1 without a judge — there's no exit condition without

one.

• Feasibility and judge signed off, no map/rules: run Step 1 — adapt a

scripts/depmap_* for the source ecosystem (prompts/01), copy

templates/RULEBOOK.md to migration/RULEBOOK.md and draft it with the

human (the policy decisions are theirs; survey the codebase for the facts),

then prompts/02 for the gap inventory. Each ends at its own gate.

• Step 1 signed off: confirm the HUMAN has installed

.claude/settings.json from templates/settings.json — prompt 03

verifies it; never install it yourself — then run

prompts/03-stress-test.md. But FIRST check the rulebook's §0 posture:

if this is a redesign migration, the bakeoff is invalid; substitute

adversarial design-doc review and tell the human why.

• Stress test signed off: confirm migration/manifest.tsv exists

(prompt 01's closing action generates it), then run

prompts/04-translation-kickoff.md with scripts/queue_runner.mjs as the

queue — first check the referee price: if the target's typecheck is cheap,

the dissolve edit is the human's act at the gate — ask them to remove

those denies from settings.json; never edit it yourself (README Step 4

dissolves into Step 3).

• Translation queue empty: Step 4 via prompts/05-survey-build.md

(survey build → machine queue → fixers without compiler access; skip it

entirely if Step 4 dissolved into Step 3), then Step 5 (hello world,

smoke), then Step 6 (inherited suite burndown, or parity referee against

the old code).

• Step 6 done-gate passed (both counts documented — README Step 6): run

prompts/06-post-parity.md — collect the port markers, classify fix-now

vs document-and-close, ship each fix as its own flagged change proved by a

parity re-run. Ends at its own gate.

Constant behaviors

• Report progress as burndown numbers (queue_runner.mjs status), not prose.
• A failure seen three times is a rule bug: stop fixing instances, queue the

amendment, propose regenerating the slice.

• "Don't migrate" and "stop here" are valid recommendations at every gate.