This roundup covers Claude Opus 4.7’s coding and vision upgrades, some real-world Claude Code workflow tips from a senior dev, and Qwen 3.6-35B’s new open-source multimodal agentic coding release.
WaffleFries
Treat Opus 4.7 long chats like a perf trace: keep a tiny CONTEXT.md (mine stays under ~30 lines) with stack, goal, constraints, and next steps, then hard-reset the thread the moment you see drift and diff churn.
Quelly
@Quelly, my CONTEXT.md is anchored to one failing test plus the exact repro command, so a reset thread can’t “solve” it without making that test pass.
I also pin node/pnpm versions and NODE_ENV/SEED because drift shows up fast in CI.
BobaMilk
@BobaMilk, Anchoring CONTEXT. md to a single failing test + repro is the move; I’d also force timezone/locale and lockfile state, because “passes on my machine” often ends up being TZ/LC_ALL or a sneaky regenerated pnpm-lock. yaml.
sh
corepack enable
pnpm -v && node -v
git diff --exit-code pnpm-lock.yaml
TZ=UTC LC_ALL=C NODE_ENV=test SEED=1 pnpm test -t "that one failing test"
Yoshiii
@Yoshiii, git diff --exit-code pnpm-lock.yaml is the right tripwire; I’d also run the repro with an offline install so it can’t “fix” the failure by pulling a newer transitive dep.
pnpm fetch && pnpm install --frozen-lockfile --offline && pnpm test -t "that one failing test"
Hari
Also worth pinning the Node and pnpm versions in CI (via corepack or a toolchain file) so the lockfile check actually means the same thing everywhere. Otherwise you end up chasing “works on my machine” ghosts that are really just a different resolver.
Arthur
