Drafti Architect

drafti-architect — Technical Design PRD

Makes design decisions from technical problems and produces an implementation-ready PRD, without a planning document.

Bash Convention

Each Bash tool invocation is a fresh subshell. Every bash block re-declares HARNISH_ROOT="${CLAUDE_PLUGIN_ROOT}" inline; no persistent variables.

Skill Chain

Can be invoked independently. Follow-up: "start implementation after review" → impl, or "/galmuri:ralphi to verify PRD consistency" → galmuri:ralphi (sibling plugin).

Step 1: Problem Clarification

Inspect 5 items from the user input. Skip items that already have answers.

• Only ask about items marked ❌. Skip items marked ✅.
• Ask 1~2 items at a time, complete within 2 rounds total. Mark unanswered items as "[unconfirmed]" and proceed.

Step 2: Existing Asset Query

Extract 3~5 tags from the problem (tech stack → problem domain → task type order).

HARNISH_ROOT="${CLAUDE_PLUGIN_ROOT}"
if [[ -n "$HARNISH_ROOT" ]]; then
  bash "$HARNISH_ROOT/scripts/query-assets.sh" \
    --tags "{extracted tags}" --format inject \
    --base-dir "$(pwd)/.harnish"
fi

• Assets found → reflect guardrails in §7, decisions in §2 reinforcement, failures in §5
• Empty result → proceed without assets

Step 3: Design Alternative Exploration

Generate at least 2 alternatives. Create alternatives even if there is an "obvious right answer."

Alternative discovery: Existing tools? Build from scratch? Architecture change? Status quo? Phased approach?

For each alternative:

## Alternative {A/B/C}: {name}
| Aspect | Evaluation |
|------|------|
| Pros | (quantify if possible) |
| Cons | (cost, risk, limitations) |
| Implementation difficulty | Low/Medium/High + reason |
| Suitable situation | When is this the best choice |
| Rejection condition | When should this NOT be used |

Details: references/design-decision.md

Step 4: Selection + PRD Writing

Selection Rationale — must be conditional:

Bad example: "A is better" Good example: "Team has 2 years React experience + 80% existing code → zero learning cost. Vue requires 3 weeks of learning. Therefore React."

Required: State current situation → what the selection gains → validity conditions ("revisit if this condition changes")

PRD Scale Assessment → Section Decision

If unclear → ask the user: "Scale: small (1-2 days), medium (1-2 weeks), or large (1 month+)?" No silent assumption. Read references/prd-template.md and write accordingly.

Step 5: Save + Asset Recording

HITL (before any file write):

"PRD draft ready: §{sections} / {scale}. Save to docs/prd-{slug}.md? (y / n / edit-slug)"

If docs/prd-{slug}.md already exists, the prompt becomes: "docs/prd-{slug}.md already exists. (overwrite / n / new-slug)" — y is not offered. Treat overwrite as explicit destructive confirmation.

• n → end. PRD not saved.
• edit-slug → ask for slug, then y (re-check existence after slug change).
• y → proceed with save below.
• overwrite → existing file is replaced without backup (user has explicitly confirmed destructive write).
• new-slug → ask for a different slug, then re-check existence; repeat until a free slug is chosen or user picks n/overwrite.

Save PRD (only after y or overwrite):

mkdir -p docs/
# Write PRD content to docs/prd-{slug}.md

Asset recording (harnish ecosystem mode):

HARNISH_ROOT="${CLAUDE_PLUGIN_ROOT}"
if [[ -n "$HARNISH_ROOT" ]]; then
  # Decision recording
  bash "$HARNISH_ROOT/scripts/record-asset.sh" \
    --type decision --tags "{tags}" \
    --title "{one-line decision}" --content "{selection rationale}" \
    --base-dir "$(pwd)/.harnish"

  # Guardrail recording (when derived constraints exist)
  bash "$HARNISH_ROOT/scripts/record-asset.sh" \
    --type guardrail --tags "{tags}" \
    --title "{one-line rule}" --content "{consequence of violation}" \
    --base-dir "$(pwd)/.harnish"
fi

Step 6: Completion

✅ PRD complete: docs/prd-{slug}.md
Includes: §4 Implementation spec / §6 Test criteria / §7 Guardrails
Next: "start implementation" after review, or /galmuri:ralphi for consistency check.

Distinction from drafti-feature

Decide based on presence/absence of a planning document. If uncertain, ask the user "Do you have a planning document?"

Context Budget

Load at most 1 reference at a time; switch when moving phase.

Prohibited

• Finishing with only 1 alternative (minimum 2)
• Groundless selection like "~is better"
• Finalizing a decision without validity conditions
• Saving PRD without explicit user confirmation in Step 5
• Saving over an existing docs/prd-*.md without explicit overwrite confirmation
• Silently assuming PRD scale when unclear
• Loading 2 references simultaneously