Prompted by a video, I Have Spent 1000+ Hours With Claude Code, watched on 2026-08-18. The owner picked four of its points to close. The useful result was not any single technique in it. It was that every one of the four is already written into this estate's standards, and three of the four have never actually run. That is the same finding as the 2026-08-13 loop audit and the 2026-08-12 enforcement audit, arriving from outside instead of from a self-review, which is worth something on its own.

What the video actually argued

It is a tier list of Claude Code features by a solo developer on the entry plan. Its top tier is verification, meaning tests written before the implementation plus a typechecker, a linter and a browser or screenshot pass for anything visual. Below that: plan mode for large tasks only, skills kept few and matched to one coding style, one worktree per chat for parallel work, subagents for their separate context windows, and a scheduled loop for unattended chores. Its cost argument is that a strong model should plan and a cheaper model should implement, and its context argument is that a new session should start at every task boundary rather than when the window fills, because the second task in a session is already degraded.

Most of that is below this estate's existing bar. The hand written CLAUDE.md beats a generated one, eighteen canon skills beat a community grab bag, and the run ledger in LOOP section 4 is a stronger claim than the word verification. The four items below are where the video is ahead of the practice, though not ahead of the writing.

Gap 1. Worktree isolation is canon and has never been used

LOOP.md names worktrees in the primitives table at line 45 and again at lines 110 to 111: one branch, one worktree, one run, no two agents editing the same tree. Measured on 2026-08-18, git worktree list in this repo returns exactly one entry, the main checkout. Every Nimbalyst session in this repo has shared one dirty tree for the whole life of the standard.

Two agent memories exist only as scar tissue from that: the catalog baseline turning red because another session edited a component doc, and the rule to use git commit <path> rather than staging in a shared tree. Both are workarounds for a collision the written standard already forbids.

To build:

• Write the per session unit into SESSION-LOOP, so the rule is stated where a session reads it.
• Wire the session start path so a session that will edit code opens in its own worktree. The harness offers EnterWorktree, so this may be an instruction rather than a script.
• Add a doctor check that names how many worktrees exist against how many sessions are live, so the gap is visible the next time it opens.

Verified by: two sessions editing the same repo at once without either one seeing the other's diff in git status.

Gap 2. The turn end gate checks types and lint, and never runs a test

Read from ~/.claude/tools/review-gate.sh on 2026-08-18. It runs bun run check when a TypeScript file changed that turn, and tools/lint-gate.ts on every turn against a baseline. It also nudges for a dev tour on a rendered change and runs proof verify on a plans board. It never runs bun test, and no other hook does either, though package.json carries test, test:e2e, test:all and coverage.

AI-DEVELOPMENT.md line 81 already asks for typecheck and tests passing before anything is called done. LOOP section 9 asks for gate output verbatim without naming which gates. So the doctrine is there and the machine covers two thirds of it.

The video's sharper point is ordering: a test written after the implementation is a test shaped to pass that implementation. Nothing in the estate says test first.

To build:

• Add a test run to the turn end gate, gated the way tsc is gated, so a turn that touched no code pays nothing.
• Route the failure through the PostToolUse path so the session receives it, and never block. See the design constraint below, which is the reason the advisory-or-blocking binary was refused.
• Add the test first ordering to AI-DEVELOPMENT as an explicit line rather than an implication.

Design constraint that governs this and gap 4 both: a Stop hook that exits zero prints to the transcript, for a person, and the model never sees it. That was measured from both ends on 2026-08-10 and the reasoning sits at the top of context-trigger.sh. So a nudge meant to change what the session does next has to travel through the PostToolUse path, not through the Stop path. A test failure that only a human reads is a report, not a gate.

Verified by: a deliberately broken test, then a turn end where the session reacts to it rather than a transcript line nobody acts on.

Gap 3. The model economy is written in detail and configured flat

SESSION-LOOP section 6 is a full treatment: the tier table, the advice to pin subagents to the small tier, the recommendation to prefer a harness mode that switches automatically, and the cache cost of a late switch that most versions of this advice omit. It names /model opusplan as the way to get the plan and execute split with no babysitting.

The machine level settings.json sets model to opus[1m], flat, for every task in every repo. Nothing in the estate has ever run the split the standard recommends.

To build:

• Set opusplan as the machine default in settings.json and start the week of measurement against the flat baseline. Accounting across two tiers is harder to read, and that is the accepted cost.
• Where subagents are defined, pin the read only ones to the small tier, which section 6 already calls free savings.
• Have the session name the switch when a task crosses from thinking into grinding, which section 6 asks for and which nothing currently prompts.

Verified by: a measured comparison of one representative task run flat against the same task run split, reported as tokens rather than as an impression.

Gap 4. The session boundary is a size, not a task

context-usage.ts warns at 200k and stops at 900k, and context-trigger.sh re-reads only after the transcript has grown by 250k bytes. Both thresholds are defensible and the warn line is set roughly where the video puts the quality dip. But the trigger is purely a size. A session that finishes a task at 60k carries that whole task into the next one, and the video's claim is that the second task is already the degraded one.

The existing warn text is well judged and should not be lost: it tells a session to finish the piece in flight, hand off, and open the sibling rather than start something new. The agent memory handoff-bound-is-the-task says the same thing. What is missing is the cheaper trigger that fires before any of that pressure exists.

To build:

• Define what counts as a task boundary in a way a hook can see. A closed plan item and a commit are the two candidates that already leave a trace.
• Route the nudge through the PostToolUse path so the session receives it.
• Keep it quiet: this fires far more often than the warn line, and LOOP section 7 says the way a gate dies is noise rather than rejection.

Verified by: a session that closes a plan item at low context and is offered the boundary, then a session that closes two in a row and is offered it once.

Item 5. The note

The owner asked for a note. The angle is not the video and not a tips list. It is the finding above: a standard that is written, published, linked from every repo and never once executed looks exactly like a standard that works, and the thing that finally exposed it was an outside voice describing a practice the estate had already documented. Three audits in this repo found the same shape from the inside and it stayed open anyway.

That connects directly to content/notes/ten-times-zero.md, which is about honest accounting of AI work. This is the same argument turned on the process instead of the output. Drafted after the four gaps close, so the note reports what happened rather than what was intended.

Decided, 2026-08-18

All three open calls settled as recommended, and the reasoning is in plans/decisions/2026-08-18-loop-practice-gaps.md rather than repeated here.

1. The test gate in gap 2 reaches the model through PostToolUse and never blocks. The binary between advisory and blocking was the wrong frame: advisory means the transcript, where nobody acts, and blocking means Stop exit 2, which can strand an unattended run.
2. opusplan becomes the machine default in gap 3, measured for a week against the flat baseline. Section 6's cache warning does not bite, because the switch lands at the plan-to-execute boundary with little behind it to re-pay for.
3. The worktree in gap 1 is per session, matching the unit Nimbalyst already attributes work to. Worth revisiting if gap 4 makes the task boundary cheap to see.

Tasks

• [ ] Gap 1: worktree isolation actually running, plus the doctor check that keeps it honest
• [ ] Gap 2: test gate wired through the path that reaches the model, test first written into AI-DEVELOPMENT
• [ ] Gap 3: model split configured and measured rather than only described
• [ ] Gap 4: task boundary trigger defined and routed through PostToolUse
• [ ] Item 5: the note, drafted once the four above have real outcomes to report