Add scope-discipline directives to the always-on working-style standard

README.md

@@ -11,7 +11,7 @@ Everything is branched so it loads only when it is relevant, keeping context lea
 
 | Loaded | What | Where it lands |
 |---|---|---|
-| **Always-on** | `working-style` — propose-before-implementing, a question isn't a work order, report honestly, flag reusable logic | `.claude/standards/`, imported by `CLAUDE.md` |
+| **Always-on** | `working-style` — propose-before-implementing, a question isn't a work order, report honestly, flag reusable logic, scope discipline & the simplest solution | `.claude/standards/`, imported by `CLAUDE.md` |
 | **When editing `*.py`** | `python` — 3.13+, type hints, PEP 695, `uv`, structured logging, definition ordering, suppression discipline, async safety | `.claude/rules/` |
 | **When editing `*.py`** (opt-in) | `scalable-architecture` — ports-and-adapters, dependency injection, boundary validation, idempotent writes; patterns for services/larger apps | `.claude/rules/` |
 | **When editing `*.md`** | `prose-and-docs` — no historical narration, verify before documenting, voice, ordering, RFC-2119 & leanness | `.claude/rules/` |

plugin/.claude-plugin/plugin.json

@@ -1,5 +1,5 @@
 {
   "name": "preflight",
   "description": "Starter kit of reusable AI-agent standards, personas, and workflows. Run /preflight:setup to scaffold editable copies into a project.",
-  "version": "0.1.0"
+  "version": "0.2.0"
 }

plugin/templates/standards/working-style.md

@@ -16,6 +16,25 @@ These apply to every task. They are deliberately short — this file is always i
 - **Flag reusable logic.** When generic utility code (transforms, math helpers, shared parsing) is
   buried as a private method in one module, raise it: propose extracting it to a shared location
   so there is one source of truth.
+- **Prefer the simplest solution.** Backward-compatibility shims, dead code paths, and versioning
+  that nothing depends on are maintenance cost, not safety — don't add them. Keep compatibility only
+  where a real consumer relies on it: a versioned or published API, an SLA, a production pipeline, or
+  a persisted data format with outside callers. When it is genuinely unclear whether something still
+  depends on the old behaviour, ask instead of defensively keeping both paths.
+- **Solve the specific ask, not its generalization.** When a task names a concrete need, implement
+  exactly that; don't build the generalized version — every case handled, exposed everywhere —
+  until a second concrete caller exists. A seam at an I/O boundary for testability or genuine
+  swappability is a present need, not speculative generality — the `scalable-architecture` rule
+  governs that call where a project follows it.
+- **Don't re-derive what the input already states.** Before writing code to compute, select, or
+  reconstruct a value, check whether the source data already states it authoritatively — if it does,
+  read and apply it rather than adding a second source of truth that can drift.
+- **Name the defect in one sentence, then make that sentence false.** Write the current wrong
+  behaviour as one sentence and let the smallest change that falsifies it bound the fix — "stop
+  discarding the table" stays small where "add X support" invites machinery.
+- **Justify each touched file against the requirement, not architectural symmetry.** Mirror the
+  shape of existing patterns, not their breadth — that one feature spans a frame, a schema, meta, and
+  inventory usage does not mean a related change must touch all four.
 - **Confirm destructive or outward-facing actions** unless already authorized. Before deleting or
   overwriting something you did not create, look at it first; if it contradicts how it was
   described, surface that instead of proceeding.