Prompt capture

v2.1.283 system prompt

SDK-CLI The system prompt of v2.1.283, as Agent SDK, print mode, seeded with this account's feature flags sent it to the model, read out of the request it put on the wire. The text comes first; what changed since the previous capture is further down, kept apart from it.

This arm was captured with the sandbox seeded from flags-account, this account's own GrowthBook payload, seeded into the sandbox .claude.json. 720 flags from data/flags/snapshots/flags-20260925T183005Z.json. What a seeded capture does and does not tell you.

Released25 Sep 26captured 25 Sep 26

Compared withv2.1.282the previous capture taken the same way

Prompt215 ln27,241 characters

Model strings274 copies of the prompt between them

Tool descriptions26 0 moved this release, 25 with a parameter schema

Captured 6 ways. Each is a separate prompt with its own history, and a change is only ever measured against the previous capture taken the same way, never across these.

Default settings

Claude Code, interactive mode Agent SDK, print mode

a clean sandbox with no stored flags, as a fresh install runs

Seeded with feature flags

Claude Code, interactive mode, seeded with this account's feature flags Claude Code, interactive mode, seeded with anonymous feature flags Agent SDK, print mode, seeded with this account's feature flags Agent SDK, print mode, seeded with anonymous feature flags

the same client started with a stored set of feature flags on, so it can send text a default install does not. What that tells you

Read it The whole system prompt → 215 lines, 27,241 characters, with a reading view and a copy button

As plain textthe same prompt as a .txt, to download or diff
26 tool descriptionsthe text the model is given for each tool
10 appended blocks80,524 more characters the client sends after the system prompt
4 copies by modelthe prompt is not one text; which models get which copy
18 prompts in this releasecompaction, subagents, reminders: each its own request

What changed since the previous capture

0 One line per document that moved, the system prompt and the tool descriptions in the same table.

No change to the system prompt since v2.1.282.

Not every arm was quiet. 2 other arms of this release moved: Claude Code, interactive mode (moved), Claude Code, interactive mode, seeded with this account's feature flags (moved).

PromptReachMovedThe first line that changed

Nothing in the system prompt or in any of the 26 tool descriptions moved between v2.1.282 and v2.1.283.

The prompt by model

4 The client does not send one system prompt: each model gets its own copy with its name and a few lines worked in. A row is one copy and how many of the 27 model strings get it, most first. Every copy opens with the same line: You are an interactive agent that helps users with software engineering tasks.

#CopyLengthSent toDocuments

1 Haiku, Sonnet and Opus, 17 model strings 215 ln · 27.2k 17 of 27 3 copies

2 Fable and Mythos, 2 model strings 75 ln · 10k 2 of 27 2 copies

3 Opus 5 family, 2 model strings 66 ln · 9.1k 2 of 27 1 copy

4 Opus 4.0 to 4.8 family, 6 model strings 52 ln · 5.7k 6 of 27 2 copies

Tool descriptions

26 The text the model is given for each tool, on this arm.

All 26 tool descriptions →

ToolDescriptionCopiesMoved

Agent 1,668 chars 2 ○same

Bash 1,315 chars 2 ○same

CronCreate 2,595 chars 1 ○same

CronDelete 100 chars 1 ○same

CronList 60 chars 1 ○same

Edit 360 chars 2 ○same

EnterWorktree 3,220 chars 1 ○same

ExitWorktree 1,923 chars 1 ○same

ListAgents 777 chars 1 ○same

Monitor 6,354 chars 1 ○same

NotebookEdit 619 chars 1 ○same

PushNotification 1,362 chars 1 ○same

Read 790 chars 2 ○same

ReportFindings 574 chars 1 ○same

ScheduleWakeup 3,396 chars 1 ○same

SendMessage 4,259 chars 1 ○same

Skill 1,417 chars 1 ○same

TaskCreate 2,146 chars 1 ○same

TaskGet 732 chars 1 ○same

TaskList 998 chars 1 ○same

TaskStop 378 chars 1 ○same

TaskUpdate 2,243 chars 1 ○same

WebFetch 469 chars 2 ○same

WebSearch 312 chars 2 ○same

Workflow 3,480 chars 1 ○same

Write 240 chars 2 ○same