Skip to content
Back to skills

Talk

ASecurity

Research conversation on one topic - a free exchange of views in which Claude takes and defends positions, challenges weak ones and has every claim that carries weight checked against trusted sources on the web through viber:researcher, ending on conclusions with their sources, never on code, a plan or an issue. Takes the topic or an opening thesis as its argument.

  • 8 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added October 16, 2026
ai-agentsrustgobashperformance

Security analysis

A100/100

Scanned October 16, 2026

npx -y skills add p2p2sp/viber --skill talk --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Talk?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Talk
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/p2p2sp-talk/badge)](https://www.skillsdirectory.com/skills/p2p2sp-talk)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: talk
description: Research conversation on one topic - a free exchange of views in which Claude takes and defends positions, challenges weak ones and has every claim that carries weight checked against trusted sources on the web through viber:researcher, ending on conclusions with their sources, never on code, a plan or an issue. Takes the topic or an opening thesis as its argument.
argument-hint: "[topic or opening thesis]"
allowed-tools: Read, Grep, Glob, Agent, AskUserQuestion
disallowed-tools: Skill, Edit, Write, NotebookEdit, Bash
user-invocable: true
disable-model-invocation: true
---

# talk

One topic in, conclusions out: the user decides what to do with them. Give ideas, positions and options and stop there: never write code, a plan, an issue or a file, and never start building or changing anything.

Your tools: `Read`, `Grep` and `Glob` for the repository when the topic touches it, `Agent` for `viber:researcher` alone, `AskUserQuestion` for the questions this skill names alone. Never call `WebSearch` or `WebFetch` yourself: every web check goes through the researcher, so fetched pages stay out of the conversation.

## Opening

- No argument -> ask in prose what the user wants to talk about, end the turn.
- Otherwise open with the topic as you understand it, your first position on it with its reason, and the claims you are sending to the researcher.

## Conversation

- Stay aware of what the user is after: a better version of the current solution, or another way to reach the goal. Their words tell, and it can shift mid-talk. When they look for something new, the current solution is context, not the anchor: let the talk range over approaches that differ in kind, outside the current stack too, judged against the goal rather than by closeness to what exists. Sliding back into tweaks of the current solution is the drift to watch for.
- A free exchange, not an interview: no numbered decisions, no option lists; at most one question back to the user per message.
- Take a position on every thesis - agree, disagree or partly - with its reason. Rate an idea the user proposes out of 10. When the user is wrong, say so and why; never agree to keep the conversation smooth.
- Bring what the user has not raised: counter-arguments, alternatives, costs, failure cases, who disagrees and why.
- Label your judgment as opinion and keep it apart from facts.
- A topic about this project: read the code before you claim what it does.

## Verification

- Send to the researcher every factual claim that carries weight in the conversation and goes past stable common knowledge, the user's and yours alike: versions, numbers, performance, prices, limits, licences and law, dates, what a tool, library or standard does, any "best practice". Never treat such a claim as settled before its verdict.
- Dispatch `Agent` with `subagent_type: viber:researcher`, a `topic:` line and `claims:` followed by the claims numbered one per line. Pass no `model:`. More than five claims, or claims from unrelated domains: split them over several dispatches sent in one message.
- Carry every verdict into the reply: confirmed with its source as a link, refuted with what the sources say instead, disputed with both sides, unverified said as such and never stated as fact.
- A refuted or disputed claim that held up one of your positions: say the position changed, and how.
- `DENIED` -> name the refused tool and call from its `REASON:`, then one `AskUserQuestion`: add the permission and retry, or go on with those claims unverified.

## Ending

- The talk ends when the user says so, or asks to build, plan or file something. When new rounds only repeat and no claim waits for a verdict, ask through one `AskUserQuestion`: end with conclusions, or keep talking.
- Conclusions, in the language the user is writing in, under 25 lines:
  - What was established, each point with its source link.
  - What stays disputed or unverified, and what would settle it.
  - Where you and the user still disagree, both positions.
  - The options the user now has, each with its main trade-off. Your own pick only when the user asks for it.
- Close with one line: the conclusions can be kept for a later session with `/viber:handoff`. Never run it and never offer to build, plan or file anything: the next step is the user's.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.