In a test where a Claude Code skill wrote instruction sheets for a human, the "what to report when done" line was missing all three days (unverified report)

What happened

[Needs review] This article is a record of a personal operations experiment, not an incident report from production. What is covered here is a failure within a verification described in the article (gaps in how human-facing instruction sheets were written).

The author presents itself as an AI, Claude by Anthropic. The article states that it was posted under the owner's account, with the AI operating the browser. The author hands tasks that need the owner's hands over in instruction sheets named ops/dayNN.md. The article sets out the full text of the skill daily-ops, which fixes how those sheets are written, and the points fixed in operation.

The skill is .claude/skills/daily-ops/SKILL.md (33 lines, MIT) in the free-version repository jisou-starter. According to the author, in the verification a new Claude Code session (desktop app, Sonnet) was opened in the kit folder and run through a fictional plan for three days, October 19 to 21, 2026 (v1.4; the daily-ops content is the same as v1.3). The sheets it wrote left out the last part of writing rule 5, which asks for each task in the form "where → what → what to paste or press → what to report when finished".

The chaos on the ground

In the author's record, the model never used the Skill tool on any of the three days; it opened SKILL.md directly and followed it. In grading, day 1 passed, though its tasks summed to 3 minutes against a stated 6. Days 2 and 3 failed, and on both, tasks 1 and 2 had no report sentence. Task 1 on day 1 lacked the same sentence, but grading did not flag it. A question phrased "if 2, you may skip task 1" also lets the human choose, and grading did not flag that either.

Later, a format-checking script (a draft check_ops.py for v1.5) was run on copies of the three days' sheets in a place without LEDGER.md. According to the author, it stopped all three days for the missing report sentence and stopped day 1 also for the duration mismatch and the "may skip" phrasing (the October 21 regression test record).

The author says the real instruction sheets in the challenge also differ from the skill. Those sheets were written by hand without loading the skill. Task 2 of Day 5 lacks the fixed sentence for screen work, as do tasks 0, 1, 5 and 6 of Day 4. The fill-in line has no "警告=" field, and Day 5's preview runs four lines against the skill's limit of three.

Root cause

The article does not identify a cause. It records only the following.

• Only the skill's name and description appeared in the startup list, and the model never used the Skill tool on any of the three days; it opened SKILL.md directly and followed it.
• The real hand-written sheets, which did not load the skill, also deviate from the rules in places. Whether the rules were reread when writing them is not in the record, and the author calls it unconfirmed.
• The check script was not run on the challenge's real sheets. It was used for the kit's verification and is not in distributions up to v1.4.

The author sums up that even with rules written down, hand-written sheets drift, and that sheets written by a model also omitted writing rule 5. Whether the skill is selected from its description has not been confirmed in real runs.

The fix

According to the author, v1.5, now delivered through note and Brain, is 54 lines and adds an item to confirm before handing over that every task has a "what to report when done" sentence, among other things. Distribution on note began on October 26, 2026, and for Brain the author confirmed the approval of v1.5's wording on October 27. It also includes scripts/check_ops.py for checking the sheet format, but that script does not check whether the fill-in line is empty. It is not in the free-version repository.

Confirmation that v1.5 fixes the problem is still in progress. As of October 21, only one real run of v1.5 had been done, on a day before the start; counting later runs, six days have now been run for real. In Day 3 on October 24, run with the same rules as the distributed version, the four items that failed on v1.4's day 2 (including the per-task report sentence) all passed, while reading outside the folder remained. On October 26, Day 1 with a different plan on Sonnet did not write the next day's sheet, and the report that passed the check differed from the one returned, so it failed. The author says the records are in VERIFICATION.md in the free-version repository.

As a lesson, the author says the check script needs to be run on the challenge's own sheets as well. In other words, handing over rules as text is not enough; a mechanical check of the sheets that come out should follow.