Spaces:

VirtuGuard
/

BOS_AI

Sleeping

BOS_AI / modules.py
Upload 9 files
bf4e44d verified about 1 month ago
26.9 kB
"""
Module content for the BOS AI Training Platform.
Edit copy here. Each module is a dict with a title, a one line summary,
a depth label, and a list of sections. Each section has a heading and
HTML body text. Use <b>label.</b> inline for section labels inside a
paragraph rather than bullet lists.
"""

IT_HELP = (
    "If anything in this training raises a question, or you are unsure whether "
    "something is safe to put into an AI tool, stop and ask BOS IT before you "
    "proceed. Submit a helpdesk ticket or reach out to the IT team on Teams. "
    "There is no such thing as a dumb question here, and asking first is always "
    "faster than cleaning up afterward."
)

MODULES = [
    {
        "title": "AI at BOS: Start Here",
        "takeaways": [
            "Claude Enterprise is the one approved AI tool for BOS and client information.",
            "A language model drafts, rewrites, summarizes, and reasons in plain language. It does not know BOS unless you tell it.",
            "You carry the responsibility for anything the tool helps you produce.",
        ],
        "summary": "What AI is, what it is not, and why BOS standardized on one tool.",
        "depth": "Foundation",
        "sections": [
            {
                "heading": "Why this training exists",
                "body": """
<p>BOS Security, including BOS Staffing, BOS Medical Staffing, and the VirtuGuard
remote video monitoring division, has adopted Claude Enterprise as the company
standard AI tool. Everyone now has access to the same assistant, working from the
same rules, under a company agreement that governs how our data is handled.</p>
<p>That is a real change from a year ago, when people were experimenting with
whatever free tool they found online. This training is meant to get every BOS
employee to a shared baseline: what these tools actually do, how to get useful work
out of them, and where the hard lines are on client and company data.</p>
<p><b>How to use this platform.</b> The six modules build on each other. Start at
the top and read in order. Module 1 and 2 give you the mental model. Modules 3 and
4 make you productive. Module 5 is required reading on security and data, and
Module 6 ties it to your day to day work and tells you where to get help.</p>
""",
            },
            {
                "heading": "What AI is, in plain terms",
                "body": """
<p>When people at BOS say "AI" today, they almost always mean a large language
model, or LLM. Claude is one. So are ChatGPT and Gemini. A language model is a
computer program that has read an enormous amount of text and learned the
patterns in it well enough to produce new text that fits what you ask for.</p>
<p>The practical result is a tool that can read, write, summarize, rewrite,
explain, translate, organize, and reason through problems described in ordinary
language. You type a request the way you would ask a capable colleague, and it
answers in kind.</p>
<p><b>What it is good at.</b> Drafting emails and documents, turning rough notes
into clean writing, summarizing long material, explaining unfamiliar topics,
brainstorming options, reviewing your own work for gaps, and helping you think
through a decision out loud.</p>
<p><b>What it is not.</b> It is not a search engine, though it can search when
that feature is enabled. It is not a database of facts that is always correct. It
is not a person, it does not know who you are unless you tell it, and it does not
understand BOS, our clients, or our procedures unless you give it that context.
It also does not carry responsibility. You do.</p>
""",
            },
            {
                "heading": "Why one tool instead of many",
                "body": """
<p>Standardizing on Claude Enterprise solves three problems at once.</p>
<p><b>Data control.</b> Free consumer AI tools generally have terms that allow the
provider to use what you type to improve their products. A company agreement puts
our data under negotiated terms instead. That matters when the conversation
involves anything about a client, a site, or an employee.</p>
<p><b>Consistency.</b> When everyone uses the same tool, IT can support it,
configure it, write guidance for it, and build shared workspaces and templates
that the whole team benefits from. Ten different tools means ten different sets of
rules nobody can keep straight.</p>
<p><b>Compliance.</b> BOS tracks security frameworks, and pursuing government work
brings additional handling requirements. A single approved tool with known terms is
something we can defend in an audit. A personal account on a random website is
not.</p>
<p>The rule that follows from this is simple: use the company Claude Enterprise
account for work. Do not put BOS or client information into any other AI tool,
including a personal Claude account, ChatGPT, Copilot in a personal Microsoft
account, or AI features baked into browser extensions and phone apps.</p>
""",
            },
        ],
    },
    {
        "title": "How AI Actually Works",
        "takeaways": [
            "The model predicts plausible text. It does not look facts up unless given a tool to do so.",
            "Its working knowledge is the context: your messages, its replies, and attached files.",
            "Every number, date, name, and compatibility claim gets verified against the real source.",
        ],
        "summary": "Prediction, training data, context, and why the tool sometimes makes things up.",
        "depth": "Foundation",
        "sections": [
            {
                "heading": "It predicts, it does not look up",
                "body": """
<p>The single most useful thing to understand about a language model is that it
works by prediction. Given everything in the conversation so far, it produces the
most plausible next piece of text, then the next, then the next. There is no step
where it checks a fact against a trusted source unless it has been given a tool to
do that and told to use it.</p>
<p>This is why the same tool that writes a flawless email can also confidently
state a wrong date, invent a statistic, or cite a report that does not exist.
The output is plausible by design. Plausible is not the same as true.</p>
<p><b>Training.</b> The model learned from a very large body of text, most of it
public. That knowledge has a cutoff date and has gaps, especially for small
companies, niche products, local ordinances, and anything that changed recently.
It knows the general shape of the security industry. It does not know our pricing,
our clients, or what a specific camera model does with our platform.</p>
<p><b>Context.</b> Everything the model can see in a given conversation is called
its context: your messages, its earlier replies, and any files you attach. That is
its entire working knowledge for that chat. If the answer you want depends on
information that is not in the context, the model will either guess or ask.
Giving it the right material up front is the whole game.</p>
""",
            },
            {
                "heading": "Memory, sessions, and what carries over",
                "body": """
<p>By default a language model has no memory between conversations. Start a new
chat and it knows nothing about the last one. Some products add a memory feature
on top of the model that saves notes across chats. Whether that is turned on, and
what it saves, is a product setting, not a property of the AI itself. BOS IT
controls how these features are configured on the company account.</p>
<p>Inside a single conversation, the context has a size limit. Very long chats,
or many large attachments, can push earlier material out of view. If a long
session starts giving worse answers or forgetting instructions you gave earlier,
that is usually why. Starting a fresh chat with a clean summary of where you are
often fixes it.</p>
""",
            },
            {
                "heading": "Why it makes things up, and what that means for you",
                "body": """
<p>When the model does not have the information it needs, it does not stop. It
fills the gap with something that looks right. People in the field call this
hallucination. It is not a bug that will be patched next month. It is a
consequence of how prediction works, and every language model does it to some
degree.</p>
<p>The risk is highest for specific facts: numbers, dates, names, quotes, legal
requirements, product specifications, and anything about our own company that the
model has no way to know. The risk is lowest for tasks where you supply the facts
and ask it to work with them: rewrite this, summarize that, organize these notes,
check this draft for tone.</p>
<p>The operating rule at BOS is that the AI is a drafting and thinking partner,
and a person is always the last check before anything leaves the building. Every
number gets verified against its source. Every claim about a regulation, a
certification, or what a device is compatible with gets checked against the
authority for it, which for device compatibility is the Immix integration matrix,
not the AI's opinion.</p>
""",
            },
        ],
    },
    {
        "title": "Your First Conversations",
        "takeaways": [
            "Confirm you are in the BOS organization workspace before any work use.",
            "Good requests carry context, specifics, material, and a next step.",
            "The first answer is a draft. Iterate two or three times and keep your own voice.",
        ],
        "summary": "Getting started in Claude Enterprise and the basics of asking well.",
        "depth": "Practice",
        "sections": [
            {
                "heading": "Getting set up",
                "body": """
<p>You should have received an invitation to the BOS Claude Enterprise account
from BOS IT. Accept it, sign in with your company credentials, and confirm that
the workspace shown is the BOS organization, not a personal account. If you did
not receive an invite, or you are not sure which account you are in, stop and ask
BOS IT before you use the tool for work.</p>
<p>Spend your first ten minutes on something low stakes. Ask it to explain a term
you have heard but never looked up. Paste in a paragraph you wrote and ask it to
tighten the wording. Ask it to lay out the pros and cons of two approaches to a
routine task. You are not trying to produce anything. You are getting a feel for
how it responds and how much detail it wants from you.</p>
""",
            },
            {
                "heading": "The four things every good request has",
                "body": """
<p>Most disappointing AI results come from a request that would have confused a
human colleague too. The fix is the same in both cases.</p>
<p><b>Context.</b> Tell it who the output is for and what it is for. "Write an
email" gets a generic email. "Write a short email to a property manager who is
not technical, letting them know a scheduled camera check moved from Tuesday to
Thursday" gets something you can send.</p>
<p><b>Specifics.</b> Say what good looks like. Length, tone, format, what to
include, what to leave out. If you want three sentences, say three sentences. If
you want plain prose with no bullet points, say that.</p>
<p><b>Material.</b> Give it what it needs to work with. Paste the notes, attach
the document, include the original email you are replying to. The model cannot
read your screen or your inbox. Anything it should know has to be in the chat,
subject to the data rules in Module 5.</p>
<p><b>A next step.</b> Treat the first answer as a first draft. Tell it what to
change. "Shorter." "Less formal." "You dropped the part about the invoice." Two
or three rounds of this is normal and is how experienced users work.</p>
""",
            },
            {
                "heading": "A few habits worth building early",
                "body": """
<p>Read the whole answer before you use any of it. Skimming is how a wrong
detail slips into a client email.</p>
<p>When you ask about something factual, ask it how confident it is and what
it based the answer on. It will often flag its own uncertainty if invited to.</p>
<p>Keep your own voice. The tool is good at producing clean, safe, slightly
generic writing. If everything you send starts sounding the same, your contacts
will notice. Use it to get past the blank page, then make the result sound like
you.</p>
<p>Start a new chat for a new task. Mixing unrelated work in one long
conversation degrades results and makes it harder to find things later.</p>
""",
            },
        ],
    },
    {
        "title": "Getting Better Results",
        "takeaways": [
            "Give the model a role and a quality bar, and paste an example of work you were happy with.",
            "Break big deliverables into an outline and sections. Ask it to answer only from attached documents.",
            "Check with BOS IT before creating a shared project; one may already exist.",
        ],
        "summary": "Techniques that separate a casual user from someone who saves hours a week.",
        "depth": "Practice",
        "sections": [
            {
                "heading": "Give it a role and a standard",
                "body": """
<p>Telling the model who to be and what quality bar to meet changes the output
more than almost anything else. "You are reviewing this proposal as a skeptical
purchasing manager who has been burned by vendors before. What would you push
back on?" produces sharper feedback than "review this proposal."</p>
<p>The same works for writing. "Write this the way an experienced operations
manager would, direct and plain, no marketing language" steers it away from the
inflated tone these tools default to.</p>
""",
            },
            {
                "heading": "Show, do not just tell",
                "body": """
<p>If you have an example of what you want, paste it. One good sample of a past
email, report, or summary that you were happy with teaches the model your format
and tone faster than any description. "Here is how I wrote the last three of
these. Write the next one in the same style using these notes." This is called
giving examples, and it is the single most reliable technique for consistent
output.</p>
""",
            },
            {
                "heading": "Break big tasks into steps",
                "body": """
<p>Asking for a finished twenty page document in one shot invites a mediocre
result. Instead, ask for an outline first. Review it, adjust it, then have it
write one section at a time. You catch problems early, and each piece gets the
model's full attention.</p>
<p>The same applies to analysis. "Before you answer, list the factors that
matter here, then work through each one, then give me your conclusion" produces
better reasoning than asking for the conclusion alone. Making it think out loud
also lets you see where its logic went wrong, which is far more useful than a
confident answer with no visible reasoning.</p>
""",
            },
            {
                "heading": "Working with documents",
                "body": """
<p>Attaching a file puts its full text into the context, which makes the model
dramatically more accurate about that material. Long policy documents, meeting
transcripts, spreadsheets exported to text, and reference material all work well.
Ask it to answer only from the document and to tell you when the document does
not cover something. That one instruction cuts hallucination sharply.</p>
<p>Before you attach anything, check it against Module 5. A file can carry
information you did not notice: a client name in a header, a site address in a
footer, an account number in a hidden column. Look before you upload.</p>
""",
            },
            {
                "heading": "Shared workspaces and reusable instructions",
                "body": """
<p>Claude Enterprise supports shared projects where a team can store standing
instructions and reference material so every chat in that project starts with
the same context. This is how a team gets consistent output without each person
retyping the same setup. BOS IT can help set these up for a department, and will
tell you which ones already exist. Ask before creating your own, so we do not end
up with a dozen slightly different versions of the same thing.</p>
""",
            },
            {
                "heading": "Common mistakes",
                "body": """
<p><b>Accepting the first answer.</b> The first draft is the starting point, not
the deliverable.</p>
<p><b>Asking it to remember something it cannot.</b> If it needs to know it, put
it in the chat.</p>
<p><b>Trusting numbers.</b> Any figure it produces that you did not give it needs
a source you can point to.</p>
<p><b>Asking yes or no questions about compatibility, licensing, or
regulations.</b> It will answer confidently and may be wrong. These go to the
authoritative source, and where BOS has one, that is the answer.</p>
<p><b>Letting it write in its voice instead of yours.</b> Watch for words like
leverage, robust, delve, and streamline, and for every list having exactly three
items. Cut them.</p>
""",
            },
        ],
    },
    {
        "title": "Security and Data",
        "takeaways": [
            "CUI, FCI, export controlled material, credentials, personnel data, and raw incident material never go into any AI tool.",
            "External material refers to clients generically unless attribution is approved in writing.",
            "Legal notices, subpoenas, claims, regulatory inquiries, and incident disputes go to leadership, not to a draft.",
        ],
        "summary": "The hard lines. Required reading for every BOS employee.",
        "depth": "Required",
        "sections": [
            {
                "heading": "Why this module matters more than the others",
                "body": """
<p>BOS earns its living on trust. Clients give us access to their cameras, their
sites, and their incidents. Employees trust us with their personal information.
Government work, when we pursue it, comes with handling rules that carry legal
weight. An AI tool is just another place data can go, and the rules about where
data can go do not change because the destination is new and convenient.</p>
<p>Read this module in full. If you are ever unsure whether something can go into
the tool, the answer is to ask BOS IT first, not to try it and see.</p>
""",
            },
            {
                "heading": "Never enter this into any AI tool, including the company one",
                "body": """
<p><b>Controlled and regulated information.</b> Controlled Unclassified
Information, Federal Contract Information, and anything export controlled must
never be processed, summarized, restated, or stored in an AI tool. If you are
handling material that may fall into these categories, stop and route it through
the approved handling process. When in doubt, treat it as controlled until BOS IT
or leadership tells you otherwise.</p>
<p><b>Personnel and payroll data.</b> HR records, disciplinary documentation,
compensation, medical information, and anything from a personnel file stays out.
This applies even when the task seems harmless, such as reformatting a document
or checking grammar.</p>
<p><b>Credentials and access details.</b> Passwords, API keys, VMS logins,
network configurations, IP addresses of client systems, and remote access details
never go into a chat, an attachment, or a screenshot.</p>
<p><b>Raw incident material.</b> Incident narratives, dispatch records, video
descriptions, and details of law enforcement interactions stay inside operations
systems. If you need help writing about incident trends, summarize at the metric
level without tying a specific incident to a specific site.</p>
""",
            },
            {
                "heading": "Client confidentiality in practice",
                "body": """
<p>Anything you create with AI assistance that could leave the company follows
the same rule as everything else we publish: no client names, site addresses,
camera locations, account numbers, or contract values. Refer to clients
generically by vertical and size, for example "a multi site auto dealer group in
the Southeast," unless the client has approved attribution in writing.</p>
<p>Inside the company account, using a client name in a working draft is a
judgment call. The safer habit, and the one we recommend, is to work with
placeholders like "the client" or "Site A" and fill in specifics after the AI
step is done. It costs you thirty seconds and removes an entire category of
risk.</p>
""",
            },
            {
                "heading": "The approved tool and only the approved tool",
                "body": """
<p>Company and client information goes into the BOS Claude Enterprise account
and nowhere else. Not a personal Claude account. Not ChatGPT. Not the free tier of
anything. Not AI features inside a browser extension, a note taking app, or a
phone keyboard, many of which send your text to a third party you have never
heard of.</p>
<p>The reason is not that other tools are bad. It is that we have not reviewed
their terms, we cannot see what they retain, and we cannot answer for them in an
audit or to a client. If you find a tool you think BOS should adopt, bring it to
BOS IT. That is how good tools get approved.</p>
""",
            },
            {
                "heading": "Accuracy is a security issue too",
                "body": """
<p>An invented statistic in a proposal, a made up citation in a report, or a
confident claim that a camera model works with our platform when it does not,
damages the company as surely as a data leak. Never let AI generated numbers,
sources, model numbers, firmware versions, or compatibility claims into a
document without checking them against the real source. For device compatibility,
the Immix integration matrix is the authority. For licensing, insurance,
certifications, guaranteed response times, and regulatory compliance, the claim
comes from the source document, not from the AI.</p>
<p>And some things do not get drafted with AI at all. Responses to legal notices,
subpoenas, insurance claims, regulatory inquiries, and client incident disputes go
to leadership. Do not draft them yourself with or without the tool.</p>
""",
            },
            {
                "heading": "Watch for manipulation",
                "body": """
<p>Language models follow instructions, including instructions hidden inside
material you give them. A document, email, or web page could contain text
designed to redirect the AI into doing something you did not ask for, such as
revealing earlier parts of the conversation or drafting something misleading.
Be cautious about pasting content from unknown senders, and if the tool ever
does something strange or unexpected in response to a file, stop and report it
to BOS IT.</p>
""",
            },
        ],
    },
    {
        "title": "Everyday Workflows and Getting Help",
        "takeaways": [
            "Each role has good uses and hard lines. Know yours.",
            "Finished work reads like a person at BOS wrote it: prose, bold inline labels, no em dashes, no emojis.",
            "Ask BOS IT before you paste, not after.",
        ],
        "summary": "What good use looks like in your role, and where to turn when you need help.",
        "depth": "Applied",
        "sections": [
            {
                "heading": "What good use looks like by role",
                "body": """
<p><b>Operations and the Command Center.</b> Turning rough shift notes into a
clean, consistent report format. Rewriting a procedure so a new agent can follow
it. Explaining an unfamiliar camera or VMS term. What stays out: incident
specifics tied to sites, video descriptions, dispatch and law enforcement
details.</p>
<p><b>Sales and channel partners.</b> Drafting and tightening outreach, preparing
plain language explanations of remote video monitoring for a prospect who has
never used it, organizing meeting notes into follow ups. What stays out: client
names and contract values in anything that could be forwarded, and any claim
about compatibility, certification, or response guarantees that has not been
verified against the source.</p>
<p><b>Staffing and medical staffing.</b> Job posting drafts, interview question
frameworks, candidate communication templates. What stays out: anything from an
individual's application, resume, or personnel record without checking with BOS
IT and HR on the approved workflow first.</p>
<p><b>Administration and leadership.</b> Summarizing long documents, drafting
policies and announcements, thinking through decisions with a structured pros and
cons pass. What stays out: compensation, personnel matters, and legal or
regulatory correspondence.</p>
""",
            },
            {
                "heading": "Writing that sounds like BOS",
                "body": """
<p>Whatever the AI produces, the finished piece should read as if a person at
BOS wrote it. In practice that means prose rather than bullet lists in client
communications, bold inline labels where a document needs structure, no em dashes
and no emojis, plain language for property managers and general contractors, and
full technical depth preserved for integrators and engineers. Spell out BOS
Security on first reference. VirtuGuard is one word with a capital V and a
capital G. Tell the tool these rules at the start of a session, or better, ask BOS
IT whether a shared project already carries them.</p>
""",
            },
            {
                "heading": "A short checklist before anything leaves your hands",
                "body": """
<p>Did I read every line of the output, not just skim it? Did I verify every
number, name, date, and source against where it actually came from? Did I remove
or generalize any client, site, or account detail that should not travel? Did I
make it sound like me? Would I be comfortable if the recipient knew exactly how
this was produced? If any answer is no, it is not done.</p>
""",
            },
            {
                "heading": "When to bring in BOS IT",
                "body": """
<p>Reach out to BOS IT when you are not sure whether information can go into the
tool, when you did not receive or cannot access your Claude Enterprise invite,
when the tool behaves unexpectedly or produces something concerning, when you
want a shared project or template set up for your team, when you have found a
tool or feature you think the company should adopt, or when you want help
building an AI step into a workflow you repeat every week.</p>
<p>IT would much rather hear from you before you paste something than after. A
two minute question on Teams is the cheapest thing in this entire process.</p>
""",
            },
        ],
    },
]