Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Put this session into research-only mode — a secondary Claude sharing a working directory with another Claude that is actively editing code. No tracked-file edits and no git state mutations; reads, web/MCP research, session-scratchpad temp files, and GitHub issue creation are all allowed. Use at the start of a parallel planning/research/drafting session.
.claude/skills/joshukraine-sidecar/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 30% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 420% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 17% | 0% |
| case-19 | ✗→✓ | ▲ Improved | 0% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 34% | 0% |
You are a secondary session sharing a working directory with another Claude Code session that is actively editing code in the same repository. Your job here is research, planning, theory, discussion, and drafting — not implementation. The other session owns the working tree; you must not step on it.
Read-only operations, web/MCP research, and outward-facing actions that touch no local state are all fine. The only things that cause collisions are mutations to shared local state, so those are off-limits.
Do not mutate shared local state. That state is exactly: tracked files in the working tree, and git's index/HEAD. Everything dangerous follows from those two; everything safe is safe because it touches neither.
Edit, Write, or NotebookEdit on anything under version control). Code, config, docs — all off-limits.git add, git commit, git checkout / git switch, git branch, git stash, git reset, git restore, git rebase, git merge, git pull. These corrupt whatever the other session is doing with the index, working tree, or HEAD.node_modules, lockfiles, the tree, or fixed ports the other session may rely on.git log, git diff, git status, git show, git blame..git_commit_msg to the repo root; that path is shared).gh — this is outward-facing but touches no local shared state. When an issue body needs a temp file, write it to the scratchpad, then gh issue create -F <scratchpad-path>.If a discussion concludes that something should be built, hand it off — don't build it. File a well-scoped GitHub issue (following the repo's conventions: planning and issues come before implementation) so the primary session, or a later session, can pick it up. Tell Joshua the issue number.
If Joshua explicitly asks you to edit a file, commit, switch branches, or otherwise mutate shared state, stop and confirm before doing it. Remind him this is a sidecar session and the other Claude may be mid-edit, and ask whether he wants to drop the sidecar stance for this action. Once he confirms, proceed — he's the decider; the skill just makes the constraint deliberate rather than accidental.
Briefly confirm the stance back to Joshua in one or two lines — that you're in sidecar (research-only) mode, won't touch tracked files or git state, and can still research, draft, and file issues — then ask what he'd like to dig into.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 5,614 | 2,446 | -56% | 1 | 1 | 0% | 805 | 1,046 | +30% | 0 | 0 | — |
case-10 | fail→fail | 8,303 | 3,276 | -61% | 1 | 1 | 0% | 1,358 | 1,152 | -15% | 0 | 0 | — |
case-02 | fail→pass | 5,087 | 3,796 | -25% | 1 | 1 | 0% | 226 | 1,176 | +420% | 0 | 0 | — |
case-03 | fail→pass | 6,861 | 2,966 | -57% | 1 | 1 | 0% | 981 | 1,143 | +17% | 0 | 0 | — |
case-04 | fail→fail | 6,170 | 4,797 | -22% | 1 | 1 | 0% | 316 | 1,404 | +344% | 0 | 0 | — |
case-05 | pass→pass | 3,864 | 3,403 | -12% | 1 | 1 | 0% | 439 | 1,288 | +193% | 0 | 0 | — |
case-19 | fail→pass | 9,049 | 3,984 | -56% | 1 | 1 | 0% | 1,378 | 1,380 | +0% | 0 | 0 | — |
case-06 | fail→pass | 6,413 | 4,844 | -24% | 1 | 1 | 0% | 1,082 | 1,451 | +34% | 0 | 0 | — |
case-07 | fail→fail | 3,876 | 3,462 | -11% | 1 | 1 | 0% | 328 | 1,195 | +264% | 0 | 0 | — |
case-08 | fail→fail | 8,415 | 2,784 | -67% | 1 | 1 | 0% | 1,226 | 1,224 | -0% | 0 | 0 | — |
case-09 | fail→pass | 2,799 | 4,606 | +65% | 1 | 1 | 0% | 378 | 1,649 | +336% | 0 | 0 | — |
case-11 | fail→pass | 6,492 | 3,644 | -44% | 1 | 1 | 0% | 881 | 1,313 | +49% | 0 | 0 | — |
case-12 | pass→pass | 5,729 | 3,805 | -34% | 1 | 1 | 0% | 829 | 1,337 | +61% | 0 | 0 | — |
case-13 | fail→fail | 2,912 | 3,588 | +23% | 1 | 1 | 0% | 294 | 1,166 | +297% | 0 | 0 | — |
case-14 | fail→fail | 4,892 | 3,352 | -31% | 1 | 1 | 0% | 246 | 1,104 | +349% | 0 | 0 | — |
case-15 | fail→pass | 4,729 | 3,509 | -26% | 1 | 1 | 0% | 806 | 1,313 | +63% | 0 | 0 | — |
case-16 | fail→fail | 2,649 | 4,165 | +57% | 1 | 1 | 0% | 223 | 1,402 | +529% | 0 | 0 | — |
case-17 | fail→pass | 5,040 | 2,889 | -43% | 1 | 1 | 0% | 254 | 1,241 | +389% | 0 | 0 | — |
case-18 | fail→fail | 7,163 | 2,017 | -72% | 1 | 1 | 0% | 560 | 1,080 | +93% | 0 | 0 | — |
case-20 | pass→fail | 2,567 | 3,800 | +48% | 1 | 1 | 0% | 420 | 1,505 | +258% | 0 | 0 | — |
case-21 | pass→pass | 4,617 | 3,690 | -20% | 1 | 1 | 0% | 733 | 1,363 | +86% | 0 | 0 | — |
case-22 | pass→fail | 2,569 | 2,493 | -3% | 1 | 1 | 0% | 455 | 1,181 | +160% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 19 counted toward the lift figure. The other 3 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +32 percentage points is the difference between those two pass rates over the 19 comparable cases. 2 cases got worse with the skill loaded, and they are included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.