Working with a pack from chat

Once your client is connected, you talk and the assistant picks the tool. This page covers what to say, what comes back, and what to do when the pack says no. Each section maps to the CLI verb of the same name in Ask, Teach, Task.

Not sure where to start? Ask these

1. “What does the growth-metrics pack cover, and what kinds of questions can it answer?”
2. “What documents and data sources is it built from?”
3. A real question from your work this week.

Name the pack

Every verb acts on one pack. In a new conversation the assistant calls list_packs and picks the one you mean. If you have several, name it:

Ask the growth-metrics pack how many swaps happened last week.

It stays on that pack until you name another.

Ask

Ask a question in plain language. You get one answer with the evidence it drew on. You never pick knowledge versus data and never write a query.

You: How many active wallets did we have last week?

Assistant: 4,812 active wallets in the week ending Sept 21, down 3% from the week before. Active means at least one swap in the trailing 30 days, per the pack’s definition. Sources: Wallet activity definitions, wallet_daily.

The assistant waits about a minute. A hard data question can take longer; the ask keeps running and the assistant follows it with action_get until it settles, relaying progress as it goes.

To see a source, ask for it. The assistant reads the cited document with read_document and can quote it or check a claim against it.

To see why an answer came out the way it did, ask the assistant to explain it. It calls explain_answer, which shows the plan, sources, and definitions behind the answer, with a link to open it in edisyl.

Answer a clarifying question

An ambiguous ask comes back waiting with a question from the pack:

Assistant: The pack needs one thing pinned down first. Do you mean wallets that traded, or wallets that signed in?

You: Traded.

Assistant: 4,812 wallets traded last week. …

Answer it and the assistant asks again with your answer folded in.

Declined vs. failed

A pack that can’t answer says so instead of guessing. Two outcomes look similar and mean different things:

• Declined: the pack understood the question and can’t or won’t answer, with a reason. “Lacks a definition” means a knowledge gap. “No rows” or “couldn’t resolve” means a data gap. Rewording won’t fix either; pass the reason to whoever maintains the pack. Improving answers covers what they do next.
• Failed: something broke. Retry, and report it if it repeats.

Teach

Teach the pack your business knowledge, such as a fact, a decision, or a definition. You can also attach a file.

You: Teach the pack that “active wallet” means any address with a swap in the last 30 days.

Assistant: I’ll add: “An active wallet is any address with at least one swap in the trailing 30 days.” Teach that to growth-metrics?

You: Yes.

Assistant: Done and indexed. The pack uses that definition from the next ask on.

Make the assistant confirm first

A wrong fact affects every future answer, and a task spends real agent time. The assistant should show you the exact text or objective and wait for a yes before teach or task. If yours doesn’t, tell it to.

A teach finishes when the content is indexed. To undo one, say “forget that”: the document and everything learned from it are removed.

Task

Give the pack’s agents an objective in plain language: research a topic and file the findings, audit what the pack covers, model data an ask couldn’t resolve. It runs in the background.

You: Have the pack research the new grant program and add what it finds.

Assistant: Objective: “Research the new grant program, summarize eligibility and deadlines, and add the findings to the pack.” Go ahead?

You: Go.

Assistant: Started. Currently: searching the open web…

The assistant relays progress and reports the summary when it finishes: what it found and what it wrote into the pack. Say “cancel” to stop it; a task mid-step finishes that step first.

Brainstorm objectives

Objectives are the questions a pack has to get right, which its calibration tests first. MCP tools call them target questions. Ask the assistant to brainstorm them and it runs an interview about what the pack is for and who uses it, then proposes questions rooted in the pack’s sources:

Brainstorm objectives for the growth-metrics pack. We use it for the weekly growth review.

It loads only the questions you confirm. Needs steer access on the pack. See Objectives.

Schedule it

Give an ask or task a time and the pack runs it on its own:

Every Monday at 9, ask the pack which hypotheses moved last week.

Each fire is its own ask or task you can read back. List, pause, resume, trigger now, or delete schedules from chat. Cron schedules run in UTC and fire at most every 30 minutes. A scheduled task spends agent time on every fire. See Scheduling recurring work for the CLI side.

Get better answers

• Be specific. Name the period, business unit, and measure: “Q3 on-time delivery for EU suppliers.”
• Put the context in the question. The pack works only from the question you send.
• Answer the clarifying question. Pick the definition you mean, and the same request continues.
• Check the sources. Every answer lists them. Open one when the number matters.
• Report a missing definition. Tell the pack’s admin, or teach it. Every user gets the fix.
• Expect a new pack to answer more slowly until its first calibration finishes. Ask the pack’s admin whether it’s ready.

Not available from chat

Connecting a warehouse, editing PACK.md, and publishing a pack are done in the CLI. If an ask keeps declining for “no modeled data,” that’s the message to pass to whoever set the pack up.

Reporting a problem

report_bug files straight to edisyl engineering and works for edisyl staff only. Everyone else: send what you asked, what you expected, what happened, and the action id if the assistant has one.