agbossjs

triage

Conveyor stage 1 → status Triage. Takes a requirement (or an existing issue), clarifies its business side through quizzes — why, what it does, what it does NOT do, constraints, business acceptance criteria (AC1…) — and creates or updates GitHub issues, splitting into a parent with sub-issues when useful. Use whenever the user brings a new requirement, idea, request, bug report or an issue number/URL and wants it clarified, triaged, "заведи тикет", "разбери задачу" — before any technical spec or code.

agbossjs 1 Updated 1w ago
GitHub

Install

npx skillscat add agbossjs/agboss/triage

Install via the SkillsCat registry.

SKILL.md

triage

Conveyor: Triage → Backlog → In progress → Test → Done.
This stage: the business side only. No architecture, no code, no implementation plan.
Output: issue(s) whose body answers why / what / what not / constraints / acceptance criteria (AC1…), in status
Triage.

Questions follow .claude/skills/QUIZ.md — read it before the first question.

Input

  • A requirement in free text → new issue(s).
  • An issue number or URL → update mode: gh issue view <N> --comments, keep what is settled,
    ask only about gaps and contradictions.
  • Several at once is fine; each gets its own quiz headers.

Steps

  1. Context. scripts/wiki.sh pull, read the relevant Wiki pages and skim the code the
    requirement touches. Existing answers become facts inside quiz situations, not questions.
  2. Quiz rounds until every item has an answer or a recorded assumption:
    • Why — which problem, who has it, what it is worth, what happens if we do nothing.
    • What it does — the main scenarios as concrete examples (who does what, what they get).
    • What it does NOT do — explicit exclusions; nearby things people will assume are included.
    • Constraints — deadline, cost / budget, providers, data, compliance, compatibility.
    • Acceptance criteria — numbered AC1, AC2, …: each an observable outcome in business
      terms (who does what, what they see or get), testable, with no implementation terms.
      Together they define "done"; the spec later proves each one with a test or a manual check.
  3. Shape. One issue, or a parent issue with sub-issues when parts deliver value on their own,
    touch different areas, or together exceed about a day of work. Propose the tree as a quiz
    (split / keep whole / other split, each with its price).
  4. Write. Body via stdin (template below):
    • new: gh issue create --title "<outcome, imperative>" --label "type:<feature|bugfix|chore|epic>,prio:<mvp|stretch>" --body-file -
    • update: gh issue edit <N> --body-file -
    • hierarchy: scripts/ticket.sh sub <parent> <child>; the parent gets type:epic.
  5. Status. scripts/ticket.sh status <N> Triage for every created or updated issue.
    Unresolved open question that blocks the spec → add label blocked and say why in a comment.
  6. Report to the user: issues with links, the tree, assumptions, and the next step:
    spec-session for these issues.

Issue body — triage part

## Why

<problem, who has it, value>

## What it does

- <scenario as a concrete example>

## What it does NOT do

- <explicit exclusion>

## Constraints

- <deadline / cost / provider / data / compliance>

## Acceptance criteria

Business terms only — what a person does and what they see. Each is testable; together they
define "done".

- **AC1** — <who> <does what> → <what they see / get>
- **AC2** — …

## Assumptions

- <decided without an answer — flag for review>

## Triage log

- <question> → **<choice>** — <why; the price accepted>

Specs, plans and docs never go to repo files: issues hold the what and why, the Wiki holds how
the system works now.