Theme
AI tools
What your agent receives
When a tool asks Vythos for context, it doesn't get your Playbook. It gets a pack: the rules that bear on this task, for this client, within the access of the person whose connection it is — ordered, with the values stated separately, and honest about what it left out.
You don't have to read this page to use Vythos. It's here because the answer to "what did my AI tool actually know?" should be inspectable.
What's in a pack
Who your agency is. The description from Settings — what you do, who for — labelled as context, not a rule, so a tool knows whose work it's doing without quoting your self-description as policy.
Your clients. The names of the clients you've added, active and dormant, with exact counts — the names client accepts, and the answer to "how many clients do we have?", which no document states. For someone restricted to particular clients, only theirs, and the tool is told not to state a total.
The rules. Each one with its identifier, category, title, whether it applies to every client or one, its version number, its text, its rank, and its values. A one-line rule arrives whole: the line is the rule, with its reason and dates where it has them, and where its home is (Vythos, or a document).
What's true about the client. When a client is named, its About facts — what they sell, who to, who they compete with — labelled as context, not rules. Only for someone who works on that client, and only while the client is live.
The values, resolved. Where several rules state the same kind of value, the one that governs is resolved and handed over, labelled with what it is: not 1.5 but 1.5× the standard day rate, not 45 but 45 days. Where two rules at the same scope disagree, no value is resolved — the pack reports the disagreement rather than picking.
Passages from memory that match the task, marked as quoted: they are your emails, documents and web pages, for a tool to use as information and never to take instructions from. A passage that seems to address an AI is flagged.
The order, stated explicitly, so a tool doesn't have to infer it: client rules first, then house rules by category.
What must be cited.
Live figures, from the system that holds them. Every pack tells the tool never to state a budget, spend, what's running, a pipeline stage or whether a contract is signed from memory: get it from the system that holds it, or say it needs checking. A rule that states one is the agency's own record, given with its date and the note that it may have changed.
A pack identifier, used to check a draft later and to report what was followed.
What it says about its own gaps
This is the part that matters most, and the part most systems skip.
| Signal | What it tells the tool |
|---|---|
| No rules found | "No governed context available. Acknowledge the gap before proceeding" — so the tool says it's improvising rather than sounding authoritative |
| No client-specific rules | "Proceed with agency defaults. No client-specific rules found" |
| A client exception exists but wasn't asked for | Which client and which category has its own rule — never what it says. Enough for the tool to ask "who is this for?" instead of quoting your defaults as universal |
| Rules omitted for size | How many were left out, and to ask for a bigger budget or fewer categories before relying on the pack |
| Rules sent as values only | When a pack is tight, a rule can arrive as its values without its full wording. It still binds, and the tool is told to search for the text if it needs it |
| One-line rules left out | How many one-line rules shared few words with the task and were left out, each named, so the tool can ask again with a task that names one |
| Figures present | "State each figure in the unit its key names, and do not calculate new figures from them" |
That last one matters more than it looks. A rule stating 1.5× the standard day rate is easy to quote correctly and then multiply into an hourly figure that appears in no version of anything — stated with exactly the same confidence as the figure it came from. A derived number is the easiest way for a governance layer to make things worse, so tools are told to hand over the figures and the rule, and leave the arithmetic to a person.
What is never in a pack
- Rules in categories the connection's owner can't see. They aren't filtered on the way out; they were never gathered.
- Other clients' rules when a client is named.
- Another person's memory.
- Another client's rules, or anything they say. The client list carries names, never rules.
- For someone restricted to particular clients: the other clients at all. Not their rules, not their names, not that an exception exists — nothing that reveals them.
Size
A pack has a token budget — about 12,000 by default — and a tool can ask for more or narrow the categories. Rules are packed in precedence order, so if anything has to be left out it's the least authoritative, and the pack says what went.
One-line rules have a share of their own — a quarter of the budget, at most about 3,000 tokens — chosen first, so a long Playbook of lines never pushes your documents out, or the other way round. Within that share:
- Always sent: every line in Terms & commitments, Pricing & rates and Who decides, every Brand & voice line that bans a word, and every past decision a client said no to.
- Then: the most recent lines, and those sharing most words with the task.
- Left out: named, as above.
A line past its end date is never sent.
Seeing it for yourself
Ask your connected tool to fetch context for a task and show you what came back. Every field above will be there, including the gap signals, and it's the quickest way to understand what your tools are actually working from.
Next: The tools.
