vbcherepanov/total-agent-memory/design-explore

design-explore

Use this skill when the user asks for "design", "architecture", "approach", "options", "trade-offs", "how should we do X", or when task classifier returns level >= 3.

Workflow

1. Discover prior art — memory_recall(query="<task keywords>", mode="index", limit=10) + analogize(query="<task description>", exclude_project=None) — check if similar decision was made before.

1. Generate 2-4 options for the design space. For each option, fill:

• Core approach (1 sentence)
• Implementation sketch (3-5 lines)
• Pros (3 bullets)
• Cons (2-3 bullets)
• Unknowns (what we don't yet know)

1. Build criteria matrix — list 3-5 criteria that matter (e.g. performance, maintainability, migration risk, team familiarity). Rate each option 1-5 per criterion.

1. Pick one with justification referring to matrix. Name what was discarded and why.

1. Save — save_decision(title, options, criteria_matrix, selected, rationale, discarded, project).

1. Only then proceed to implementation — with a clear decision in memory for next session.

Output template

## Decision: {title}
 
### Prior art
{memory_recall/analogize results — 1-2 sentences}
 
### Options
 
#### 1. {option name}
-**Approach:** {1 sentence}
-**Implementation:** {sketch}
-**Pros:** {3 bullets}
-**Cons:** {3 bullets}
-**Unknowns:** {bullets}
 
#### 2. {option name}
... (same structure)
 
### Criteria matrix
 
| Criterion | Option 1 | Option 2 | ... |
|---|:-:|:-:|:-:|
| {criterion} | 5 | 3 | ... |
 
### Selected: {option name}
 
{rationale — 2-3 sentences referencing matrix}
 
### Discarded
 
-**{option name}**: {1-sentence reason}

When NOT to use

• L1/L2 tasks (quick fixes, straightforward additions) — overkill, just do it.
• Single-option situations (no real trade-off).
• Implementation bugs (debug first, design later).