arsdlc-craft

Ponytail decides whether the code needs to exist. Craft decides how the code that must exist is written: least lines, honest types, organized so the next change touches few files - then delete what no longer earns its place.

0. Companion resolution (silent, every session)

1. If a skill named code-craft is available (Claude Code plugin ansh-other-skills@ansh-other-skills, a .agents/skills/code-craft/SKILL.md copy, or any harness that reads SKILL.md), invoke it for the full ladder, smell→fix tables, and reference files (types, control-flow, naming, organization, refactoring catalog, tests). Then apply §3 of this file.
2. If not installed, apply §1-§3 below. Say once per session: "official code-craft not installed; using arsdlc's embedded distillation. Install with npx arsdlc addon code-craft."
3. Never install anything without asking.

1. The ladder (climb before writing)

Stop at the first rung that solves the problem:

1. Does it need to exist? No present, real requirement → don't build it (YAGNI). Existing code now unused and in scope → delete it.
2. Stdlib / the platform? Use it; add no dependency.
3. An already-installed dependency? Use it before adding a new one.
4. One line / one expression? Write that; don't wrap it in a class, layer, or config.
5. The minimum honest code. Smallest version with good names and real types.

If you cannot point to two concrete cases today, an abstraction (interface, base class, strategy, plugin hook, generic param) is premature. Wait for the third (Rule of Three).

2. Smell → fix (touched lines only)

smell	fix
dead / unused code (exports, params, branches)	delete; confirm dynamic references first (reflection, routes, DI, templates)
deep nesting	guard clauses, early returns; invert the condition
long argument list	options object - but only from 3-4 args that travel together
any / casts / non-null assertions	type the boundary once, honestly; no lying types to silence the checker
swallowed errors / floating promises	handle, wrap with context, or let it crash loudly - never catch {}
comments restating the code	delete the comment; rename if the code is unclear
copy-paste duplication	extract only on the third instance
god class / fat interface	split along the seams the change actually uses - not speculative axes

The official code-craft skill carries the full tables and six reference files; this row set covers the 90% case.

3. The arsdlc part: craft serves the spec

1. Blast radius = the task's requirement and its scenario. Fix in-scope lines; out-of-scope mess gets flagged for arsdlc-review, not fixed in passing. Unrequested cleanup is how diffs grow teeth.
2. Deletion that changes spec-relevant behaviour is a delta. Dead-code removal touching a SHALL line's path goes through arsdlc-delta, never a quiet edit.
3. A tool's candidate list is not proof. knip/vulture/tsc findings must be confirmed by hand (dynamic references) before deletion - and the deletion still needs the evidence run to pass.
4. Craft never trades away the never-lazy list (arsdlc-ponytail §4): trust boundaries, data loss, security, a11y, anything the spec says.
5. Report what you changed and what you deliberately left alone; the second half is what the reviewer reads first.