# Business Analyst

> Interview-driven requirements discovery — turn raw client input (emails, notes, transcripts, briefs) into a validated, traceable requirements package before solution architecture. Use when the user asks to analyze client requirements, run discovery on a lead, clarify a vague request, build a BA requirements package, or prepare requirements for architecture and estimation.

- Skill: `vukhanhtruong/business-analyst` (Agent Skill, multi-file: 18 files)
- Install (CLI): `npx skillmds add vukhanhtruong/business-analyst`
- Raw SKILL.md: https://api.skillmd.com/api/skills/vukhanhtruong/business-analyst/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Product & Planning
- Author: vukhanhtruong (https://skillmd.com/u/vukhanhtruong)
- Updated: 2026-09-10
- Page: https://skillmd.com/skills/vukhanhtruong/business-analyst

---


# business-analyst

Turn incomplete client input into two artifacts in the lead directory:
`requirements.md` (five parts, human-readable) and `requirements.json`
(machine handoff for analyze-requirements, estimate, and proposal).

## Hard rules

1. Never invent requirements — every material fact carries a label
   (`confirmed` | `assumed` | `recommended`) and a source. A
   recommendation never renders as a confirmed requirement.
2. Problem before solution — no technology recommendations, no
   architecture. This skill stops at WHAT; HOW belongs to
   analyze-requirements and later skills.
3. Unknowns render as open questions, never silently filled. Distinguish
   unknown (client doesn't know) from undecided (client must choose).
4. Never block on unanswered questions — write the package with an honest
   status and readiness; a re-run refines it, never restarts the interview.
5. Every requirement carries a stable ID and traceability links;
   `node scripts/validate.mjs` must exit 0 before the package is final.
6. The human reviews requirements.md before status may become
   `READY_FOR_ARCHITECTURE`.

## Flow

1. **Detect evidence**: input documents? an existing `requirements.json`
   (re-run — see Re-run)? greenfield or existing system? State findings;
   the user can override.
2. **Depth**: ask QUICK / STANDARD / DEEP first
   (`references/interview.md` §6).
3. **Extract**: pull every known goal, actor, process, rule, constraint,
   integration, and assumption from ALL inputs before asking anything.
4. **Gap analysis**: extracted knowledge vs the nine layers
   (`references/interview.md` §4) → prioritized P1/P2/P3 gap list.
5. **Select frameworks**: from `references/frameworks.md` — justify each
   pick in one line; never apply all.
6. **Interview**: per `references/interview.md` — grouped, adaptive,
   contradiction-challenging, example-hungry. Solution involves AI or
   agents → also work through `references/ai-extension.md`.
7. **Write**: requirements.md + requirements.json per
   `references/writing.md`.
8. **Validate**: `node scripts/validate.mjs --json <dir>/requirements.json
   --md <dir>/requirements.md` — fix findings, re-run until clean. No Node
   in the environment (e.g. the claude.ai sandbox) → run
   `python3 scripts/validate.py` with the same flags; identical checks.
9. **Fresh-eyes review**: dispatch a subagent per `references/review.md`;
   apply findings, re-validate; one cycle max.
10. **Human review**: show Part 5 (readiness report); the human confirms
    the status. Only they can promote it to `READY_FOR_ARCHITECTURE`.

## Re-run

`requirements.json` already exists → diff the new answers and material into
the registers (answer open questions, confirm assumptions, add findings),
keep every existing ID stable, recompute readiness, and advance the status.
Never restart the interview; ask only what is still open.

## Handoff

Downstream skills treat `requirements.json` as detected evidence — never a
prerequisite. FR/BR/ASM/Q ids are stable so architecture decisions and
estimates can cite them.

## Dependency

Node ≥ 20, or Python ≥ 3.10 where Node is unavailable (`scripts/validate.py`
is a parity-tested port of `scripts/validate.mjs`). Scripts are
dependency-free either way.