Install
$ agentstack add skill-seryozh-agent-skills-design-the-conversation ✓ scanned · ✓ verified, works with Claude Code, Cursor, and more.
Security review
✓ PassedNo issues found. Passed automated security review. · v0.1.0 How review works →
- ✓ Prompt-injection patterns
- ✓ Secret / credential exfiltration
- ✓ Dangerous shell & filesystem operations
- ✓ Untrusted network calls
- ✓ Known-malicious package signatures
What it can access
- ✓ Network access No
- ✓ Filesystem access No
- ✓ Shell / process execution No
- ✓ Environment & secrets No
- ✓ Dynamic code execution No
From automated source analysis of v0.1.0. “Used” means the capability is present in the source — more access means more to trust, not that it’s unsafe.
Verified badge
Passed review? Show it. Paste this badge into your README, it links to the public security report.
Reliability & compatibility
Declared compatibility
Compatibility is declared by the source manifest. End-to-end runtime verification is coming, see below.
We're building live execution health for every listing: tool-call success rate, median latency, uptime, and last-checked timestamps, measured, not self-reported. It isn't live yet, so we don't show numbers we can't stand behind.
How agent discovery & health will work →About
Design the conversation
A message that ends in a question is not finished. It is finished when you can say what happens after each answer, and where the exchange lands if it goes well.
The failure this prevents is the research interview wearing a sales suit: a smart, specific, well-researched question whose answer teaches the sender something and leaves the reader with nothing to do. The reader senses it, and the reply never comes. When it does come, the sender is holding a fact and no next move.
The method
1. Name the destination. Not the reply. The real thing: a signed agreement, a funded case, a decision made, a door opened. Write it in one sentence with who has to say yes.
2. Write the closing turn first. What is the message that gets that yes, and what must the reader already know and believe by the time it arrives? That belief list is the work the earlier turns have to do.
3. Work backwards to turn one. Each turn earns exactly one thing. If a turn is not moving the reader closer to the belief list, it is decoration and it costs you a reply.
4. Branch-test the question. Write out every plausible answer, including the discouraging ones and the one-word ones. For each, write your actual next sentence.
- If any branch leaves you with nothing to say, the question is wrong. Replace it.
- If every branch leads to the same next move, the question is theater. Skip it and make the move.
- If a branch leads somewhere better than your planned path, the question was better than your
plan. Follow it.
5. Make the answer worth giving. Their reply should unlock something they cannot get on their own: a number computed from their own figures, an honest comparison with what they already use, access, a decision they have been unable to make. Say so in the message, in one clause, so the reader can see what their answer buys them. "Whichever way you answer, I can tell you X" turns a question into a trade.
6. Keep the cost of replying near zero. One question per message, answerable in a single line from a phone. A question that requires them to look something up, poll their team, or compose a paragraph is a question you will read about in your unanswered-mail folder.
7. Put the heavy ask where it has earned its place. A first message that asks for a meeting is asking a stranger to spend time before they know why. The ask belongs in the turn after their answer, when their own words have made it obvious. If you cannot wait, you are not designing a conversation, you are running an ad.
The test before sending
Say out loud: "They reply with ___. Then I say ___. Then they ___." If any blank cannot be filled with a specific sentence, do not send it yet.
Worked example
A software support team publishes a four-hour first-response promise. A queue-triage consultant wants to find out whether one small paid pilot would help.
- Destination: the support lead approves a one-week pilot on one queue.
- Closing turn: "Send one anonymized week of that queue, and I will return the routing plan on
Friday."
- What they must believe by then: that the delay is concentrated enough to diagnose, the sample is
safe to share, and the pilot does not require a system migration.
- Turn one question: "Which ticket type most often misses the four-hour target?"
- Branch A, a ticket type: show how that type would be separated and ask for the one-week sample.
- Branch B, "none": do not force the pilot. Ask whether resolution time, rather than first response,
is the actual constraint.
- Branch C, "we do not track it by type": offer a lightweight measurement sheet before discussing a
pilot.
- Each answer earns a different immediate move. The conversation may still converge later, but the
first question is doing real work rather than decorating a predetermined ask.
Guardrails
- Designing the path never licenses a claim you cannot support. No invented exclusivity, no
fabricated urgency, no numbers the reader did not give you.
- Never write a branch that depends on the reader misunderstanding you.
- A conversation the reader would not want if they saw the whole map is a manipulation, not a
design. If the map embarrasses you, change the offer, not the wording.
Source & license
This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.
- Author: Seryozh
- Source: Seryozh/agent-skills
- License: MIT
Install and usage instructions live in the source repository linked above.
Reviews
No reviews yet, be the first.
Write a review
Versions
- v0.1.0 Imported from the upstream source.