discovery

Files (1)

• SKILL.md

SKILL.md (first lines)

# Discovery — Requirements Discovery Interview

## What is Discovery

Discovery is a systematic interview process that clarifies requirements before building. Instead of asking open-ended questions and hoping for complete answers, Discovery proposes answers based on available context and asks the user to confirm or correct.

**Core mechanic:** One question at a time with a **proposed answer**. The user confirms, corrects, or expands. This is faster and more precise than open-ended questions.

Example:
```
Q: What is the primary deployment target?
Proposed: Docker containers on a Linux VPS (based on your existing Hetzner infrastructure).
→ User: "Correct" or "No, this runs on Vercel"
```

---

## Interview Modes

Select mode based on task complexity.

### Quick Mode (5-10 questions) — L1-2

For small fixes and enhancements where scope is mostly clear.

**Covers:**
1. Goal — What are we building/fixing?
2. Constraints — Time, tech, compatibility limits
3. Done criteria — How do we know it is finished?

### Standard Mode (15-25 questions) — L2-3

For features that affect multiple files or introduce new behavior.

**Covers everything in Quick, plus:**
4. Users & stakeholders — Who uses this? Who cares about it?
5. Edge cases — What happens when inputs are unexpected?
6. Dependencies — What existing code/services does this touch?
7. Alternatives — What other approaches were considered?

### Deep Mode (25-50 questions) — L3-4

For major features, new systems, or anything with business impact.

**Covers everything in Standard, plus:**
8. Business model — How does this affect revenue or costs?
9. Scale — Expected load, growth trajectory
10. Security & privacy — Data sensitivity, compliance requirements
11. Integration — External APIs, third-party services
12. Migration — How to transition from current state

---

## Source-Material Re-Read Rule (comparative research tasks)

**When building a candidate list for comparative research (technology selection, framework evaluation, etc.), re-read ALL raw materials the user provided at task inception** — source files, URLs, prior dialogs, wiki entries. Cross-check your assembled list against them before finalizing.

Do NOT rely on session memory or your own enumeration. The user pointed at specific files for a reason — they contain candidates, context, or constraints you may miss otherwise.

Source: prior incident reflection — a top-3 candidate was in a user-supplied source file but omitted from the candidate list because the agent built the list from memory instead of re-reading the file.

---

## Research Insights Awareness

If `datarim/insights/INSIGHTS-{task-id}.md` exists (created by `/dr-prd` Phase 1.3), **read it before starting the interview**. Use research findings to:
- Formulate better proposed answers (based on verified versions, docs, best practices)
- Skip questions already answered by research (e.g., "What version of X?" if insights already captured it)
- Surface risks or compatibility issues discovered by the researcher agent

Source: prior incident — research phase produces insights before discovery interview runs.

---

## Scope Live-Grep Rule (multi-artefact tasks)

**When a task touches multiple files of the same kind (commands, agents, skills, templates, pages, services, migrations, etc.), grep the filesystem for the actual count before fixing scope in the PRD.** Do not rely on memory or prior PRD numbers — they drift. The PRD's "scope" section must reflect the live filesystem at PRD-write time, not a stale enumeration.

Common live-counts:

```sh
ls ${DATARIM_RUNTIME:?}/commands/dr-*.md | wc -l   # current dr-* command count
ls ${DATARIM_RUNTIME:?}/skills/*.md | wc -l         # skills (single .md file)
ls ${DATARIM_RUNTIME:?}/agents/*.md | wc -l         # agents
ls ${DATARIM_RUNTIME:?}/templates/*.md | wc -l      # templates
git ls-files '*.service.ts' | wc -l          # services in monorepo (example)
```

Source: prior incident — PRD § Scope said "15 commands"; actual count was 17 (`/dr-help`, `/dr-publish` were missed). In compliance Step 1 this surfaced as "unplanned addition" and forced a scope-vs-actual reconciliation in the QA report. Cost: ~5 minutes diff-explanation. Avoidance: `wc -l` before PRD draft.

## AC-Feasibility Rule (operating-model check)

**Every measurable Acceptance Criterion in a PRD must be reachable under the current operating-model.** Before user-approval (PRD Phase 3), dry-run each AC against live state:

- AC says "exit 0" — run the command, check actual exit code on a representative working tree.
- AC says "file X exists" — verify the path resolution rules produce that path.
- AC says "build succeeds" — confirm the build command exists and the target environment is intact.
- AC says "runtime in sync with repo" — check whether the sync mechanism (curate / git pull / symlink) is active and applicable.

If an AC is not currently reachable due to operating-model state (e.g. symlinks vs real copies, git submodules vs flat repos, container vs host runtime), reformulate as **"X OR documented invariant satisfied by construction"** with the documented exception path.

Source: a prior incident formulated an AC as "drift-detector exits 0" under a tool that returned a non-zero symlink-guard signal under the symlink-first operating model. The AC was technically unreachable but semantically over-satisfied; the mismatch surfaced in QA only and should have been caught in PRD Phase 3.

---

## Codebase-First Rule

**Before asking any question, check if the answer already exists in code or docs.**

Sources to check (in order):
1. `datarim/projectbrief.md` — project overview
2. `datarim/productContext.md` — existing product requirements
3. `datarim/techContext.md` — technology decisions
4. `datarim/systemPatterns.md` — architecture patterns
5. `package.json`, `Cargo.toml`, `go.mod`, etc. — tech stack
6. `README.md`, `AGENTS.md` — project conventions
7. Existing code structure and tests

**If you can look it up, don't ask.** State what you found and ask the user to confirm:

From Arcanada-one/datarim, MIT.

More from Arcanada-one/datarim