System Prompts Leaks is a collection of captured system instructions used to guide AI chatbots and coding agents before they receive user messages. It serves researchers and developers studying how different AI assistants are directed.
Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add asgeirtj/system_prompts_leaks --skill designgit clone --depth 1 https://github.com/asgeirtj/system_prompts_leaksWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/asgeirtj/system_prompts_leaks/design)<a href="https://agentmods.dev/skills/asgeirtj/system_prompts_leaks/design"><img src="https://agentmods.dev/badge/skills/asgeirtj/system_prompts_leaks/design/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/asgeirtj/system_prompts_leaks/design"><img src="https://agentmods.dev/badge/skills/asgeirtj/system_prompts_leaks/design.svg" alt="Reviewed on agentmods" width="80" height="20"></a>- Snyk warn
- NVIDIA SkillSpector warn
SkillSpector: 3 findings, up to high
These are SkillSpector’s own severities. On a checked sample its high-severity flags on skills were ~96% false positives — a documented command, a public API, a “never do X” rule — so we show them as a caution to read, not a verdict. Why →
- high Anti-Refusal · line 93 Skill instructs the agent to omit warnings, disclaimers, or ethical commentary. Stripping safety caveats hides risk from the user and is a common jailbreak preamble.Fix: Remove instructions that suppress warnings, disclaimers, or ethical commentary. Let the agent surface safety-relevant caveats to the user.
- medium Excessive Agency · line 56 Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.Fix: Add human-in-the-loop confirmation for destructive, irreversible, or high-impact operations. Never auto-execute commands that modify files, send data, or alter system state.
- low Excessive Agency · line 955 Skill's behavior or capabilities extend beyond its stated purpose. Scope creep allows an agent to perform actions unrelated to its documented functionality, increasing the attack surface.Fix: Limit the skill's scope to its documented purpose. Remove instructions that enable the agent to perform actions outside its stated functionality.
What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00224 | $0.13967 |
| Opus 5 | $0.00112 | $0.06984 |
| Sonnet 5 | $0.00045 | $0.02793 |
| Haiku 4.5 | $0.00022 | $0.01397 |
Grade A, and why
design scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured today.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 986 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Create a design canvas
Two quick exits. Empty request: ask in one line what they want
designed (and for what), then stop. Request EXACTLY one of consent,
revoke, sync, login, import, export or status alone (or
import/export/sync plus only a URL or project name): that is a
Claude Design account/project command this preview doesn't handle -
say so in one line and stop. For consent, revoke, login, sync
point at /design <verb> alone (/design-sync <project> for a sync
with a project hint); those need a first-party claude.ai login and an
org policy permitting Claude Design, so without either say Design
consent/sync is not available here. For import, export, status
say those are not available while this preview is on and point at
claude.ai/design, never a /design ... spelling. Do not design
something named "status". Anything that describes something to design
-- a login page, an export dialog, a status dashboard - is a brief.
This is an early preview of Claude Design inside Claude Code: the
skill ships a precompiled payload - Claude Design's "Design
Components" editor on a multi-artboard canvas, packaged to run inside
a published Artifact. It is not at parity with claude.ai/design and
the editor baked into each canvas does not update after publish; say
so plainly if asked. You do NOT build or modify the editor - you seed
design content into a copy of the payload with the helper, and
publish. Every .dc.html file renders as its own ARTBOARD (its own
sandboxed preview iframe) on one pan/zoom canvas; canvas.json lays
them out and picks the launch view. Where saving is enabled (the
artifact-publish capability - step 4 finds out) the viewer gets a
WYSIWYG canvas: click-to-select, a properties panel bound to the
focused artboard (closed until opened from the toolbar or a
selection's quick menu), inline text editing, undo/redo, edits local
until the explicit Save publishes the page for everyone. Without
it Save is refused and the view is read-only - viewing plus PNG/PDF
export is what the user gets. Never edit the payload's code: only the
title, the README note and the state block vary between canvases.
The foundation - save model, untrusted-state rule, no-egress iframe
rule, content guidance - is under "Foundation" at the end. One general
artifact rule is deliberately SUPERSEDED here: a design canvas stores
and EXECUTES .dc.html, which is only safe because the editor never
renders published content in its own page - everything runs in a
nested sandboxed preview iframe (opaque origin, no allow-same-origin,
inheriting the CSP's no-egress rule, postMessage-only). That isolation
is load-bearing; nothing may weaken it.
Keep the machinery to yourself - helper, payload, state block,
capabilities, contracts, versions - even when a publish fails or is
denied. Narrate
the deliverable ("drafting two directions for the poster", "saving
your canvas"). Never ask the user to approve or confirm a publish in
chat: the tool collects its own approval. (The one publish-time
question that stays is the "anyone still editing?" check before a
force: true save, under "Updating an existing canvas".)
What ships with it
1 file beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- today Changed · +8 lines 9587e71ac0a9
- 11d ago First seen · 978 lines · 224 tokens per session scan A bf9ab730d92e
design is a skill published in the GitHub repository asgeirtj/system_prompts_leaks (64,614 stars, last pushed yesterday), licensed CC0-1.0. It adds 224 tokens to every session and 13,967 once invoked, about $0.0011 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
claude-design
Design one-off HTML artifacts (landing, deck, prototype).
tldraw-offline
Drive and script tldraw offline canvases with an agent.
baoyu-infographic
An infographic generator that turns supplied content into a visual summary using different information layouts and visual styles. It supports standard and custom image proportions and multiple languages.
sketch
Throwaway HTML mockups: 2-3 design variants to compare.
adversarial-ux-test
Roleplay a hostile user to find and triage UX pain points.
design-md
Author/validate/export Google's DESIGN.md token spec files.