Ban plan shorthand, name forward references

docs/references/kernel-style-as-example.md

@@ -9,7 +9,7 @@
 - Phase model: Phase0 planning.md (non-trivial trigger), Phase1 coding.md always hot, Phase2 review.md + exemplars-routing mandatory before commit, Phase3 commit.md + changelog-style mandatory.
 - ID taxonomy: `R0-*` factual integrity, `CL-*` changelog, `CS-*` code structure, `CC-*` comments, plus `<!-- ID -->` matching hot ↔ cold rationale via `check-orphan-ids.py`.
 - Lint chain: `lint-changelog.py` (CL-12 50/70), `lint-code.py`, `checkpatch.pl --strict -g HEAD`, `verify-cover-letter.py`.
-- Generic distilled rules we re-used: WHY not WHAT 50-word cap, one source at definition, helper extraction by theme, should_/is_ bool naming, minimal obvious fix, one idea per paragraph.
+- Generic distilled rules we re-used: WHY not WHAT 50-word cap, one source at definition, helper extraction by theme, should_/is_ bool naming, minimal obvious fix, one idea per paragraph, forward references naming the artifact (GC-14b), working-context shorthand ban (upstream-hygiene class 6).
 
 ## License
 

skills/changelog-quality/SKILL.md

@@ -51,6 +51,10 @@ Good: "Fix by having the offline path wait for pending work for the memcg before
 
 Split helper, sync bool, trylock vs mutex belong in code comments/doc, not changelog.
 
+### Forward references name the artifact, for purpose not status
+
+<!-- GC-14b --> A patch that exists for a later patch names what the later patch introduces — the function, helper, or contract — never the later patch's plan label or status. Bad: "switched to it in the next patch", "until step B lands", "a later change needs the same three lines" (names neither the change nor the lines, so in the committed log it points at nothing). Good: "cleans up the name space for adding a scoped-query variant". Justify a prep patch on its own terms: open with the problem that exists now, not with what a follow-up will do — at this patch that behavior does not exist yet, so stating it as motivation misleads a reviewer reading in order.
+
 ### GC-15 — Say what change does NOT do / its limits if non-trivial
 
 Scope guard: "Does not change behavior for GET", "No fast-path impact".

skills/upstream-hygiene/SKILL.md

@@ -54,6 +54,7 @@ Do not enumerate private names — ban classes, with `example.invalid` placehold
 3. **Internal tooling/process:** private scripts, cron jobs, internal ticket IDs, private CI job names, local-only wrappers — unless project documents them.
 4. **Cross-project references:** "like our fix in <other-repo>" unless direct documented dependency (import, API, shared lib). Upstream fix should stand alone.
 5. **Personal dev notes:** `TODO(<real-person>)`, "we do X in our infra", "for our perf lab", internal benchmark machine names.
+6. **Working-context shorthand and process narration:** step labels from a private plan (`phase-2 item`, `step B of the rollout`), bare first-name references (`the case <first-name> flagged`), and process verbs carrying no technical content (`review found three problems`, `as decided in sync`). Rewrite by naming the artifact — the function, contract, or finding — not the session it came from.
 
 **Why class-based:** Listing private names to ban them leaks them. Teach agent the class, use `example.invalid` in examples, never paste real private suffixes into skill or commit.