diff --git a/NEXT_SESSION_PROMPT.md b/NEXT_SESSION_PROMPT.md
new file mode 100644
index 00000000..b7f46f64
--- /dev/null
+++ b/NEXT_SESSION_PROMPT.md
@@ -0,0 +1,150 @@
+# Next session: make write-to-system installers traceable
+
+> **Paste this whole file into a fresh Claude session opened in `~/testsAndMisc`.**
+> It is self-contained. Do not go looking for the previous session's context.
+
+## The one job
+
+Extend `meta/scripts/trace_shell_split.sh` so that scripts whose job is
+**placing files** can be executed safely and diffed. Right now they are the
+only class of remaining file-length violations that cannot be verified at all,
+which blocks roughly half of what is left.
+
+This is a **tooling task, not a split.** Do not split any script this session
+until the harness change is landed and proven. Then, if there is room, use it.
+
+## Why this is blocking
+
+`meta/scripts/trace_shell_split.sh` (read it first, ~180 lines) runs a script
+with mutating _binaries_ shadowed by `PATH` stubs that record their calls and
+exit 0. That works for `sudo`, `pacman`, `systemctl`, `adb`, `mount`, and about
+25 more.
+
+It does **not** work for shell redirections. `cat > /etc/modprobe.d/x.conf` and
+`cat > "$HOME/.local/bin/foo.sh"` are bash builtins-plus-redirection; no `PATH`
+entry can intercept them. So today the choice for these scripts is: stub the
+copy tools and get a nearly empty trace, or run for real and rewrite the live
+system. Both are useless.
+
+Concrete blocked files (all real, all currently unverifiable):
+
+| File                                | Writes                                                                                 |
+| ----------------------------------- | -------------------------------------------------------------------------------------- |
+| `nvidia_troubleshoot.sh` (336)      | `cat > /etc/modprobe.d/nvidia-gsp-disable.conf`, `mkinitcpio`                          |
+| `install_usage_monitoring.sh` (290) | `$HOME/.local/bin/*.sh`, `$unit_dir/*.service`, `*.timer`                              |
+| `install_pacman_wrapper.sh` (316)   | the pacman wrapper + immutable-file handling                                           |
+| `block_compulsive_opening.sh` (705) | copies the running script into `/usr/local/bin`                                        |
+| `install_leechblock.sh` (485)       | `mkdir -p`/`rsync`/`cp -a` into `$INSTALL_ROOT`, `$VERSION_DIR`                        |
+| `setup_thorium_startup.sh` (443)    | `cat >` a launcher + a `.service`; `sudo -u` `mkdir -p` into systemd/autostart/i3 dirs |
+
+`block_compulsive_opening.sh` is the highest-stakes one: the handoff records
+that a bad split there "breaks three daily-use apps + the pacman rewrap hook".
+Treat it as the last of this group, not the first.
+
+## What is already known about the shape
+
+Run these before designing anything — the answers are load-bearing:
+
+```bash
+grep -n 'unit_dir=' linux_configuration/scripts/periodic_background/system-maintenance/bin/install_usage_monitoring.sh
+grep -rhoE '>\s*"?(/etc|/usr/local/bin|/usr/share|/var/lib|\$HOME[^"]*)' <each file above>
+```
+
+Measured on 2026-08-16: most destinations are **`$HOME`-rooted variables**
+(`unit_dir="$HOME/.config/systemd/user"`), not hardcoded absolutes. That
+matters, because redirecting `HOME` to a temp dir handles a large share with
+**no change to the scripts themselves**. The genuinely hardcoded `/etc` and
+`/usr/local/bin` writes are the smaller, harder residue — decide explicitly how
+to treat those rather than assuming one mechanism covers everything.
+
+Two cases that a bare `HOME` redirect will **not** catch, so design against
+them rather than discovering them later:
+
+- `setup_thorium_startup.sh` writes via `sudo -u "$SUDO_USER" mkdir -p ...`.
+  `sudo` is already stubbed to exit 0 without running its argument, so those
+  directories are never created and anything depending on them takes a
+  different path. Decide whether the `sudo` stub should _execute_ its argument
+  under the prefix instead of swallowing it — that is a real behaviour change
+  to the harness and affects every existing trace.
+- `install_leechblock.sh` writes through `rsync -a --delete` and `cp -a` into
+  `$INSTALL_ROOT` / `$VERSION_DIR`. `--delete` against a wrong root is
+  destructive, so confirm where those variables point _before_ the first run.
+
+## Design constraints (do not violate)
+
+1. **Never mutate the real system to verify a split.** This is Decision 6 in
+   `refactor_claude_todo_resume.md` and it is not negotiable. The whole point of
+   the harness is that it makes running these scripts safe.
+2. **The trace must still be diffable** — same format as today: exit status,
+   ordered stubbed calls with arguments, stdout, stderr. A before/after diff at
+   a detached worktree is how a split gets proved.
+3. **A silent stub is a trap, not a pass.** `--stub du` with no value makes
+   every `((size > 0))` false and the trace walks past every guarded branch;
+   two such traces match while exercising nothing. Any redirect you add must
+   not create the same illusion — if a write is redirected, the trace should
+   _show_ the write, with its destination and ideally its content hash.
+4. **Do not weaken the gate to land this.** No suppressions, no `exclude:`.
+5. `trace_shell_split.sh` is itself under the 250-line cap. If the change
+   pushes it over, split it — the harness is not exempt from the rule it serves.
+
+## Suggested approach (your call — argue if you disagree)
+
+A `--prefix <dir>` flag that runs the target with `HOME` pointed at a temp tree
+and, for the hardcoded-absolute residue, either a `bwrap`/`fakeroot` bind-mount
+of `/etc` and `/usr/local` or an explicit refusal with a clear message. Prefer
+one honest mechanism plus a named limitation over two half-working ones.
+
+Whatever you choose, the acceptance test is the same: pick
+`install_usage_monitoring.sh`, trace it, and confirm the trace shows the unit
+files being written **into the prefix**, with `~/.config/systemd/user/` and
+`~/.local/bin/` on the real machine untouched afterwards (`ls -la` them, and
+check `systemctl --user list-unit-files | grep usage-report` is unchanged).
+
+## Verify it the way this repo verifies things
+
+- Prove the harness catches a **real** break, not just that it runs. The
+  existing fixtures in the session log did this by reproducing the two
+  documented bug classes; do the equivalent for redirected writes — e.g. a
+  fixture whose split drops a write, and confirm the trace diff shows it.
+- Re-run the three existing regression fixtures (the `set -e` function-tail
+  abort, the self-referencing nameref, the valued-stub branch test) so the
+  change does not weaken what already works.
+- `shellcheck` + `shfmt -d` clean, **fixing** findings rather than suppressing.
+- `pre-commit run --files <changed>` before committing; the real `git commit`
+  additionally runs jscpd.
+
+## Repo rules that will bite you
+
+- Every commit touching code needs an evidence JSON in
+  `docs/superpowers/evidence/`; **≥4 staged code files** additionally needs a
+  fresh `docs/superpowers/contracts/*.json`
+  (`python3 meta/scripts/validate_contract.py <file>`).
+- New `.sh` files need the executable bit or the commit hook rejects them.
+- Markdown needs `npx prettier --write` — prettier is pre-push only, so a file
+  that passes every per-commit gate can still fail the push.
+- Work directly on `main`; commit and push. `git stash` and branch creation are
+  blocked by hooks. Use `git worktree add --detach` for a clean baseline.
+- `git push` runs `ci-mirror` (clean-venv install + `pre-commit --all-files` +
+  pytest) and takes minutes. Never edit files while a push is running.
+- **Do not wire the file-length pre-commit hook.** It must land last, only once
+  `bash ~/utils/scripts/check_file_length.sh --all` exits 0, or every push
+  fails. See "The constraint that will bite you" in the handoff.
+
+## Read these first
+
+1. `meta/scripts/trace_shell_split.sh` — the thing you are changing.
+2. `docs/shell-split-verification.md` — how a green split is still broken;
+   includes the silent-stub trap, the symlink-entry-point trap, and the
+   existing limits of the harness.
+3. `refactor_claude_todo_resume.md` — current state (57 files over the cap),
+   the decisions already made, and the verifiability triage for picking targets.
+
+## Definition of done
+
+- `trace_shell_split.sh` can execute at least `install_usage_monitoring.sh` and
+  `nvidia_troubleshoot.sh` and produce a diffable trace.
+- The real `~/.config/systemd/user/`, `~/.local/bin/` and `/etc/modprobe.d/` are
+  provably untouched after those runs.
+- The three existing regression fixtures still behave.
+- A fixture proves the trace diff **detects** a dropped write.
+- Committed and pushed, with evidence (and a contract if ≥4 code files).