Skill instructions
conversation-triage2.73 KB
View saved version →
---
name: conversation-triage
description: >-
Find and prioritize the customer conversations a teammate should follow up on:
ones that are still unresolved and ones the AI escalated for the team to
review. Read a sample, summarize what happened, note any that got a negative
rating, and suggest a next step for each. Use when the user asks what
needs a reply, which chats to review, what to follow up on, or to sort the
support inbox. Do NOT use for aggregate stats (use performance-review) or to
change the Agent (use improve-agent).
metadata:
author: quickchat-ai
version: "1.0"
---
# Conversation triage
Surface the customer conversations a teammate should follow up on, and make each
one actionable.
## Before you start
- Call `list_scenarios` to resolve the Agent. One Agent -> use it; else ask.
- Default window: the last 7 days.
## Steps
1. Use `get_insights` with insight_type="flagged" for the conversations the AI
escalated for the team to review. It defaults to the most recent matches
across ALL time, so pass `start_date`/`end_date` to constrain it to the
window rather than filtering all-time results yourself, and page with
`next_cursor` if needed.
2. Use `list_conversations` with `resolution_status="open"` plus
`start_date`/`end_date` for the window to find open conversations.
3. Open `get_conversation_detail` on the most important few to read the
transcript, see what happened, and note any negative rating the customer left.
4. Where a reply looks wrong or an action seems to have failed, call
`get_message_diagnostics` on that conversation: it shows the tools the AI
actually executed (with errors) and the Why AI Said That analysis when one
exists (`analysis_status=not_generated` means it has not run yet).
5. Cross-reference: a conversation that is both escalated and unresolved is the
top priority.
6. For each item the user approves, act on it with `update_conversation`:
`assign_to` to route it to a teammate (or to 'me'), or `status='resolved'` to
close it out (add `assign_to='me'` if it is not yours, `confirm=true` to resolve).
## Guardrails
- Assigning and resolving are real writes on the customer's Inbox, and resolving
ends the conversation and can ask the visitor for a final rating. Ask before
each one and never batch them without approval. Both need SUPPORT access or
above; if you do not have `update_conversation` on this connection, recommend
the next step and tell the user to take it in the Quickchat Inbox instead.
## Output
A prioritized list, most important first. For each: the customer's request in one
line, why it is on the list (unresolved or escalated for review), any negative
rating it received, and a suggested next step. End with the single top item to
handle first.
Referenced files: 1 — file references are metadata, not package contents.
deploy-agent6.65 KB
View saved version →
---
name: deploy-agent
description: >-
Put a configured Quickchat AI Agent in front of real users: report which
channels are live, then connect the ones the user wants. The website widget,
public chat page, the Agent's own MCP endpoint, Discord, Telegram, Instagram,
WhatsApp, Slack and Email all start through the tools here, with browser
consent and any required setup choices. Use when the user asks to deploy,
publish, go live, share the Agent, embed it on a site, or connect or integrate
it with Telegram, Discord, Instagram, Slack, email, WhatsApp or any other
channel. Not for building or configuring the Agent itself (use launch-agent or
improve-agent).
metadata:
author: quickchat-ai
version: "1.1"
---
# Deploy an Agent to its channels
Connecting a channel is tool work, not dashboard homework. Every channel below
has a first-class path through `connect_agent_channel`; only the ones under
"Dashboard-only channels" need the Quickchat dashboard, and you say so exactly
once instead of guessing.
## Before you start
1. Resolve the Agent: call `list_scenarios` and match the name to its
`scenario_id`. If several qualify, ask which one.
2. Call `get_deployment_info` and read `channels` — that is the verified state.
A channel listed in `connected` has configured routing;
`configured_but_inactive` is set up but not live right now;
`not_configured` is a definite no;
`not_checked` (or absent from all four lists) means you could not verify it,
so say that rather than guessing.
A configured WhatsApp route does not prove current Meta permissions or
delivery. Ask the user to check an actual reply in WhatsApp.
3. A channel name is never a website. When the user says "connect Telegram" or
"put it on Instagram", that is a `connect_agent_channel` call — do NOT
scrape telegram.com or instagram.com as a knowledge source, and do NOT
describe dashboard menus for a channel this tool connects.
## Connecting each channel
All of these are `connect_agent_channel` with the right `channel` value. Relay
the returned `instructions` faithfully — they say who may open a link and for
how long it works.
- **Telegram** (`channel`=telegram): switches on the Agent's direct-message
link and returns it. The link is durable and shareable — ideal for a personal
assistant. Telegram groups are the one Telegram case that needs the
dashboard.
- **Discord** (`channel`=discord, optional `nickname`): returns a one-time
link. The user opens it in a browser where they are signed in to Quickchat,
picks the server, and approves. It works once, only for them, and expires in
about 10 minutes (`expires_in_seconds` is authoritative) — if it lapses,
mint a fresh one instead of resending the old link.
- **Instagram** (`channel`=instagram): returns a one-time link like Discord's.
It forwards to Instagram, which always asks the user to sign in (even with a
session already open) so they can choose the professional (business or
creator) account the Agent should answer DMs for; they sign in with that
account and approve message access. This cannot be completed for them: give
them the link and tell them which account to sign in with.
- **WhatsApp** (`channel`=whatsapp): returns a one-time link, valid for about
10 minutes and only for the requesting Quickchat user. They open it signed
in to Quickchat, finish Meta's WhatsApp Embedded Signup and consent, and
complete any business account or phone selection when Quickchat reopens.
Never ask for Meta access tokens in chat. After they finish, read
`get_deployment_info.whatsapp`: `pending_account_selection` means they still
need to choose an account, even if a previous route remains in
`channels.connected`. Ask them to message
the connected number from WhatsApp to check the Agent's reply. Testing with
`send_message_to_agent` does not test WhatsApp delivery.
- **Slack** (`channel`=slack): returns a workspace-install link. It needs no
Quickchat sign-
Preview of the first 4,000 characters. Read full instructions →
Referenced files: 1 — file references are metadata, not package contents.
improve-agent3.99 KB
View saved version →
---
name: improve-agent
description: >-
Audit a Quickchat AI Agent and improve it: read its current settings and
knowledge base, mine recent conversations for gaps (unanswered questions, wrong
answers, off-brand tone, missing knowledge), then propose specific
configuration edits and knowledge-base additions. Apply changes only after the
user approves each one. Use when the user asks to improve, tune, optimize, fix,
or update an Agent, add knowledge, or change its persona, greeting, language,
or profile picture.
metadata:
author: quickchat-ai
version: "1.0"
---
# Improve an Agent
Read first, change second. Diagnose from real data, propose the smallest fix, and
never write without explicit approval.
## Before you start
- Call `list_scenarios` to resolve the Agent. Config tools require EDITOR-or-above
on that Agent; if the user lacks it, say so and stop.
- Ask whether to PROPOSE changes only (default) or also APPLY approved ones.
## Diagnose
1. `get_assistant_settings` — persona, profession, creativity, greeting, language,
reply length, KB descriptions.
2. `list_knowledge_base_articles` — what the Agent already knows. `content` here
is a preview, not the article; `get_knowledge_base_article` returns the body.
3. `get_insights` plus a `list_conversations` sample read with
`get_conversation_detail` to find where it struggled. Classify with
`references/failure-taxonomy.md`.
## Propose
For each recurring problem, propose the SMALLEST change that fixes it:
- Missing knowledge -> a new KB article (draft the exact text).
- Wrong tone/persona -> a specific `personality` (an enum id, not a scale) or an
`ai_commands` guideline.
- Too long / short / robotic -> `reply_length` or `creativity_level`.
Present proposals as a before -> after diff, then STOP for approval.
## Apply (only after approval)
- `update_assistant_settings` — pass ONLY the fields that change; effective
immediately, no retrain. `ai_commands` REPLACES every existing guideline, so
read the current list first and send it back with the change. The list is
capped at 10,000 characters in total: near the cap, merge or shorten
guidelines rather than appending one more, which can only be rejected.
- `add_knowledge_base_article` — content required; embedded automatically in the
background, no retrain.
- `update_knowledge_base_article` — replaces the article's WHOLE body, so read it
with `get_knowledge_base_article` first and send the full edited text.
`previous_title` + `previous_content` restore the previous title and
body, unless `previous_content_truncated` is true.
- `delete_knowledge_base_article` — irreversible; read the article first, since
the response echoes only the first 20,000 characters for recovery. It requires
`expected_title` and `expected_added_at` from
`list_knowledge_base_articles` (they fail the call if the article changed
underneath you) plus confirm=true.
- `set_assistant_avatar` — profile picture from an image URL, a base64 data URI
(png/jpg/gif/webp, max 5 MB), or `use_default=true` to reset. It permanently
replaces the previous image and the legacy launcher icon, and needs a plan
that allows widget customization. It returns the resulting `avatar_url` —
there is no settings field to re-read it from.
- After each other write, re-read with `get_assistant_settings` /
`list_knowledge_base_articles` and confirm the change landed.
## Guardrails
- Never call a write tool before the user approves that specific change.
- Never build an edit out of a preview — correcting one line still means sending
the whole article, and a preview sent back silently drops the rest.
- `personality` and similar fields are enum ids — use the ids from the tool
schema, not free text.
- Some fields (e.g. `ai_commands`) need a plan feature; if a write is rejected for
that reason, report it plainly instead of retrying.
## Output
A short report: top issues (ranked, one example each), the proposed changes as a
diff
Preview of the first 4,000 characters. Read full instructions →
Referenced files: 2 — file references are metadata, not package contents.
launch-agent6.7 KB
View saved version →
---
name: launch-agent
description: >-
Configure a Quickchat AI Agent from scratch, end to end: set up the empty
Agent the account already has (or provision one when every Agent is already
configured), build it from a website URL or from a short interview when there
is no site, then run a confirmed customer question and show the actual reply.
Use when the user asks to create, build, launch, set up, or spin up a new
Agent, to make an Agent from a website or URL, or when they have just
connected Quickchat and have nothing set up yet. Not for editing an
established Agent (use improve-agent).
metadata:
author: quickchat-ai
version: "2.3"
---
# Launch a new Agent
Stand up a working Agent in one guided flow. Most people running this have just
connected Quickchat and already have an empty Agent, so configure THAT one and
build something before you report anything.
An OAuth connection alone is not a request to build or run a test. Start this
flow when the user asks for it.
## Before you start
1. Call `list_scenarios`. Connecting Quickchat already created an Agent, so if any
entry has `configured: false` AND a `role` of EDITOR or above, use that
`scenario_id` for everything below and do NOT call `create_assistant` (creating
another leaves theirs blank forever). Ignore unconfigured Agents you only have
VIEWER on: they are someone else's and onboarding them will fail. Only when
every Agent you can edit is already configured, or the user asked for an extra
one, call `create_assistant` with confirm=true and use the scenario_id it
returns. If more than one Agent qualifies, ask which to set up.
2. Ask whether they have a website to build from. That answer picks the path below.
3. Their Agent is on the FREE tier (no charge, no card); say so if you create one.
## Path A — they have a website
1. Use the scenario_id from "Before you start". Set the display name with
`update_assistant_settings` if they gave one.
2. Call `onboard_assistant_from_url` with that scenario_id, the URL and
confirm=true — all three are required, and the call is rejected without the
confirm. Say out loud that this takes 20-60 seconds while it scrapes the
site, writes the persona and embeds the content — silence here is where
people give up. It OVERWRITES persona and settings and adds the scraped
content to any existing knowledge, so run it on an Agent that holds nothing
yet; on an already-configured Agent, ask first.
3. Poll `get_assistant_settings` until `onboarding_from_url_completed` is true.
Stop if `onboarding_from_url_blocked` appears; explain its `message` and ask
for a usable source instead of rerunning the rejected URL.
If it reports `onboarding_progress_tracked` false there is no completion
signal. Read the settings once after about 60 seconds and verify the persona
and knowledge before testing; do not claim a pending or failed build is ready.
4. Show what was generated: the name, the main prompt, the guidelines and the
language it picked.
If the URL is rejected as a placeholder or marketplace domain, ask for their own
business website rather than retrying the same URL.
## Path B — no website
Do NOT create an empty Agent and stop. Interview first, in one short round of
questions:
- What does the business do, and what is it called?
- Who will be talking to the Agent (customers, members, players, staff)?
- What tone should it take?
- What are the top 3 questions it must answer?
Then configure their Agent in a single `update_assistant_settings` call on the
scenario_id from "Before you start". Do this on every no-website run, including
when you had to create the Agent: creation happens before this interview, so it
cannot have carried answers you had not collected yet, and skipping the write
leaves the new Agent blank.
- `name` and `one_word_description` — what the Agent is called
- `short_description` — the main prompt: role, what it helps with, what it must
not guess at
- `ai_c
Preview of the first 4,000 characters. Read full instructions →
Referenced files: 1 — file references are metadata, not package contents.
performance-review3.09 KB
View saved version →
---
name: performance-review
description: >-
Produce a performance review of a Quickchat AI Agent: pull conversation
volume, resolution rate, CSAT, handoff response time, and topic trends over a
period and turn them into a prioritized set of recommendations. Use when the
user asks how an Agent is doing, for a weekly or monthly review, a health
check, KPIs, or to compare one period against another. Do NOT use for reading
individual conversations (use conversation-triage) or changing settings (use
improve-agent).
metadata:
author: quickchat-ai
version: "1.2"
---
# Performance review
Turn a Quickchat AI Agent's analytics into a short, decision-ready review.
## Before you start
- Call `list_scenarios` to resolve which Agent the user means. Exactly one Agent
-> use it; otherwise ask which one.
- Confirm the period. Default: the last 7 days vs the 7 days before it.
## Steps
1. Use `compare_periods` to compare the current period with the previous one of
equal length. Pass the PREVIOUS (older) window as `period_a` and the CURRENT
window as `period_b`: deltas are computed as period_b minus period_a, so this
makes a positive delta mean "up versus the previous period." Returns both
overviews plus per-metric deltas.
2. Use `get_topics` for the customer-intent split, and read `topics_by_day` from
the overview for free-text themes that are rising.
3. Use `get_csat` for satisfaction and `get_ttfr` for human handoff speed (only
if the Agent hands off).
4. For any count, rate, total, or trend, ALWAYS use the analytics tools above.
Never page through `list_conversations` to compute an aggregate.
## Read the numbers correctly
- `resolution_rate` already combines confirmed and assumed resolutions; report it
as a percentage of conversations with a question (`total_resolved` plus
`total_unresolved`), not of all conversations. Small talk (`total_small_talk`)
is not counted. A missing `resolution_rate` means nothing was counted in that
period, not 0%.
- `get_csat` empty means no CSAT was received, NOT a low score — say "no CSAT
data" rather than implying dissatisfaction.
- `get_ttfr` measures HUMAN responders after a handoff, in seconds, not
business-hours adjusted; prefer the median and note overnight gaps inflate the
average.
- See `references/metrics-glossary.md` for the full field reference.
## Output
1. One-line headline: better or worse, and why.
2. A compact table: this period vs last, with deltas, for volume, resolution
rate, CSAT, and handoffs.
3. Rising topics worth attention.
4. 2-3 concrete, prioritized recommendations.
Keep it scannable; lead with the metric that moved most.
For period comparisons, `compare_periods` supplies only overview metrics. Call `get_topics`, `get_csat` and, when applicable, `get_ttfr` separately for EACH period. Pass `start_date=period_a_start`, `end_date=period_a_end` for the previous period, then `start_date=period_b_start`, `end_date=period_b_end` for the current period, using the same date values as `compare_periods`. If either period has no CSAT ratings, report the comparison unavailable; never treat missing data as zero.
Referenced files: 2 — file references are metadata, not package contents.