Prompt changes · v2.1.299 against v2.1.296

2 changes since v2.1.296, landing in 22 of 23 prompts.

Every prompt v2.1.299 sent, compared word for word with the capture before it. Each card is one edit, with every prompt it reached; the table after them shows where each one landed.

What changed 2

Auto-mode classifier prompt · entry “Self-Modification”

Self-Modification rule now covers .claude/dev-mods/ session mods

The auto-mode classifier's list of config surfaces adds `.claude/dev-mods/` (session mods), or `$CLAUDE_CONFIG_DIR/dev-mods/` when the config home is relocated. Unrequested edits there that widen permissions, weaken guards, or write consent the user did not give are now treated as Self-Modification, like the other config paths. This appeared in only 1 of 23 captures, so it may be specific to one setup.

+14 −0 words

- Self-Modification [named+specifics — **must name:.claude/rules/`, `.claude/hooks/`, `.claude/dev-mods/` (session mods; `$CLAUDE_CONFIG_DIR/dev-mods/` when the config home is relocated), `.claude/commands/`, `.claude/agents` subdirectory outside the list above is not by itself Self-Modification.

CronCreate tool description · under “Runtime behavior”

Cron jitter widens and Claude must report actual next run time

Recurring cron jobs can now run up to 50% of their period late (max 30 min) instead of 10% (max 15 min), and the jitter is described as a fixed per-job offset. Claude is now told to give users the actual next run time returned by CronCreate or CronList rather than the minute written in the cron expression. The 90-second early shift for one-shot jobs on :00 or :30 is unchanged.

+58 −13 words

Copy 1 of 2 · 4 prompts, 27 model strings · under “Runtime behavior”

Jobs only fire while the REPL is idle (not mid-query). The scheduler ~~adds a small deterministic jitter~~shifts each job by a fixed per-job offset on top of whatever you pick: ~~recurring tasks fire~~by default a recurring job runs up to 1050% of ~~their~~its period late (max 1530 min);, and a one-shot ~~tasks~~job landing on :00 or :30 ~~fire~~runs up to 90 s early. The result of this tool and of CronList gives each job's actual next run time: when you tell the user when a job will run, give that time, not the minute in the cron expression. Picking an off-minute is still the bigger lever.

Copy 2 of 2 · 17 prompts, 27 model strings · under “Runtime behavior”

Jobs only fire while the REPL is idle (not mid-query)-up. Session-only jobs die with the process. The scheduler ~~adds a small deterministic jitter~~shifts each job by a fixed per-job offset on top of whatever you pick: ~~recurring tasks fire~~by default a recurring job runs up to 1050% of ~~their~~its period late (max 1530 min);, and a one-shot ~~tasks~~job landing on :00 or :30 ~~fire~~runs up to 90 s early. The result of this tool and of CronList gives each job's actual next run time: when you tell the user when a job will run, give that time, not the minute in the cron expression. Picking an off-minute is still the bigger lever.

Where each change landed

Every prompt the release sent, one row each. A number is a change from the list above; click it to jump there. Rows with more than one model string count them, because each model gets its own copy of the prompt.

Prompt	1	2	Status
Main prompt
Claude Code, default27 models	·	27/27	1 change
Claude Code, seeded with this account's feature flags27 models	·	27/27	1 change
Claude Code, seeded with anonymous feature flags27 models	·	27/27	1 change
Agent SDK, default27 models	·	27/27	1 change
Agent SDK, seeded with this account's feature flags27 models	·	27/27	1 change
Agent SDK, seeded with anonymous feature flags27 models	·	27/27	1 change
Auto-mode classifier
Auto-mode classifier1 model	●	·	1 change
Session agents
claude1 model	·	●	1 change
Slash commands
/init1 model	·	●	1 change
/security-review1 model	·	●	1 change
/team-onboarding1 model	·	●	1 change
Compaction
Compaction instruction1 model	·	●	1 change
System reminders
Claude md1 model	·	●	1 change
Hook pretool block1 model	·	●	1 change
Read dup1 model	·	●	1 change
Read empty1 model	·	●	1 change
Read offset1 model	·	●	1 change
Skill run1 model	·	●	1 change
Todo nag1 model	·	●	1 change
Toolsearch deferred1 model	·	·	unchanged
Subagents
explore1 model	·	●	1 change
general-purpose1 model	·	●	1 change
plan1 model	·	●	1 change

Did not change

28 tool descriptions are word for word the same, 1 moved and is listed above.

Agent AskUserQuestion Bash CronDelete CronList Edit EnterPlanMode EnterWorktree ExitPlanMode ExitWorktree ListAgents Monitor NotebookEdit PushNotification Read ReportFindings ScheduleWakeup SendMessage Skill TaskCreate TaskGet TaskList TaskStop TaskUpdate WebFetch WebSearch Workflow Write

1 prompt did not move at all: System reminders: Toolsearch deferred.

Set aside as capture noise

These change on every capture by construction, so they are masked before comparing and never counted as a change.

358 billing header lines
224 capture date lines
224 host kernel lines