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 duffelgit 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/duffel)<a href="https://agentmods.dev/skills/asgeirtj/system_prompts_leaks/duffel"><img src="https://agentmods.dev/badge/skills/asgeirtj/system_prompts_leaks/duffel/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/duffel"><img src="https://agentmods.dev/badge/skills/asgeirtj/system_prompts_leaks/duffel.svg" alt="Reviewed on agentmods" width="80" height="20"></a>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 | $0.00040 | $0.06376 |
| Opus 5.5 | $0.00016 | $0.02550 |
| Sonnet 5.5 | $0.00008 | $0.01275 |
| Haiku 4.5 | $0.00004 | $0.00638 |
Grade A, and why
duffel 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 19d ago.
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 — 552 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Duffel Flight Booking
For a user-facing flight search or booking, first read
/opt/hatch/skills/booking/SKILL.md,
/opt/hatch/skills/booking/references/flights.md, and
/opt/hatch/skills/booking/references/presentation.md. Those files define the
discovery, privacy, presentation, and commitment rules. Use this file for the
Duffel CLI contract. Do not emit HTML, call create_options, or duplicate
flight choices in another widget; the native flight widget owns flight
selection and uses the action chosen for the current flow.
Use only this public CLI surface:
search
seat-options
validate-booking
book
booking-status
cancellation-quote
cancel-booking
Do not invent identifiers, call undocumented endpoints, or expose off_, ord_,
or ore_ ids to users. Do not show a Duffel command. Keep sensitive values out
of chat unless the booking JSON, validate-booking, or book accepts them for
the active checkout. Never request payment-card data or account credentials.
Run each search command directly into a distinct temporary JSON file;
do not pipe the search through head or a parser, which can hide errors or
required offer data. After the direct search succeeds, a read-only command may
inspect the saved JSON to group and rank offers, but it must not overwrite the
file. Reuse the unchanged file for presentation. If an earlier search was
piped, truncated, reduced, or overwritten, do not repeat it in the same turn;
explain that a complete result was not retained and ask the user to continue in
a new turn. A hand-built summary is not sufficient input for the required
structured flight list.
Booking flow
- Start by searching the complete itinerary and exact passenger mix together.
- Present the shortlist through
/opt/hatch/skills/booking/references/flights.md. - If Travel Planning delegated the search in
planning onlyposture, return the selected itinerary to Travel Planning and stop. Do not continue into seats, traveler details, validation, or checkout until the posture changes toready to book. - Optionally show
seat-optionsand collect bookable seats. - Collect required passenger details and any loyalty account the user wants attached.
- Let
/opt/hatch/skills/booking/references/flights.mdchoose between the native path and the airline website. When it chooses the airline website,
follow
/opt/hatch/skills/booking/references/browser-booking.md. Recheck the itinerary and terms. Do not require the user to complete the checkout. - Otherwise continue through the supported native path without asking the
user to choose an implementation provider. Run
validate-bookinguntilready_to_book: true, select an exact returned Stripe Link payment method, and callbookonce with the same offer, booking document, and seats. - On success, report every carrier confirmation reference/PNR and ticket status. Do not add the Duffel name to the message. Any required provider attribution belongs to the provider-controlled approval or receipt surface.
What ships with it
2 files 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.
- 19d ago First seen · 552 lines · 40 tokens per session scan A a8faf6258e16
duffel is a skill published in the GitHub repository asgeirtj/system_prompts_leaks (70,113 stars, last pushed yesterday), licensed CC0-1.0. It adds 40 tokens to every session and 6,376 once invoked, about $0.0002 per session on Opus 5.5. 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 2026-10-05.
Other skills, from other repositories
comps-analysis
Build comparable-company valuation workbooks in Excel.
anatomy
Explain an idea as an interactive isometric machine figure.
pretext
Build creative browser demos with DOM-free text layout.
weights-and-biases
W&B: log ML experiments, sweeps, model registry, dashboards.
nemo-curator
Curate LLM training data: dedupe, filter, PII redaction.
subagent-driven-development
Execute plans via delegatetask subagents (2-stage review).