[H1][DESIGN-SYNTHESIS]

Dictum: Design decisions require grounded context before implementation.

Synthesize research findings into design decisions via light codebase investigation.

Workflow:

• §INGEST — Load research artifact, parse original request

• §SCAN — Light codebase investigation via parallel-dispatch (3-4 agents)

• §EXPLORE — Generate 2-3 approaches with trade-offs

• §SELECT — Commit to best approach with rationale

• §OUTPUT — Structured design document

Dependencies:

• parallel-dispatch — Agent orchestration for codebase scan

• Research artifact — External findings from deep-research

Input:

• Research : Path to research artifact (research_{slug}.md )

• Request : Original user request/intent

[1][INGEST]

Dictum: Grounded context prevents speculative design.

Load and parse inputs:

[INDEX] Source Extract

[1] Research file Findings, confidence levels, key sources

[2] Request Intent, scope boundaries, success criteria

Parse research structure:

• [1][FINDINGS]

→ Domain knowledge by category

• [2][CONFIDENCE]

→ High/Medium/Low ratings

• [3][SOURCES]

→ Attribution for decisions

[IMPORTANT]:

• [ALWAYS] Extract high-confidence findings as primary input.

• [ALWAYS] Note low-confidence areas as design risks.

• [NEVER] Proceed without understanding request intent.

[2][SCAN]

Dictum: Pattern awareness prevents reinvention.

Dispatch 3-4 agents via parallel-dispatch for codebase context.

Agent Assignment:

[INDEX] [AGENT] [SCOPE] [RETURNS]

[1] Patterns Similar implementations in codebase Conventions, reusable patterns, prior art

[2] Constraints Project rules, architecture limits Hard boundaries, style requirements

[3] Interfaces Entry/exit points for feature area Touch points, consumers, integration surface

Agent Prompt Template:

Scope: [Specific investigation area] Objective: Surface [patterns|constraints|interfaces] relevant to: [request summary] Output: Bullet list of findings with file paths Context: Research indicates: [key findings summary] Exclusions: Do NOT analyze implementation details or specific file contents

[CRITICAL]:

• [ALWAYS] Dispatch ALL agents in ONE message block.

• [ALWAYS] Scope to patterns/constraints/interfaces—NOT implementation.

• [NEVER] Deep-dive into file contents—that's plan's job.

[3][EXPLORE]

Dictum: Comparison reveals optimal trade-offs.

Generate 2-3 distinct approaches from research + scan findings.

Per Approach:

[INDEX] Aspect Content

[1] Strategy High-level implementation direction

[2] Alignment How it leverages research findings

[3] Patterns Which codebase conventions it follows

[4] Trade-offs Pros and cons

Approach Generation Criteria:

• Approach A: Most aligned with existing patterns (conservative)

• Approach B: Best leverage of research findings (optimal)

• Approach C: Simplest implementation path (minimal) — optional

[IMPORTANT]:

• [ALWAYS] Ground approaches in scan findings—no speculation.

• [ALWAYS] Include trade-off analysis per approach.

• [ALWAYS] Apply YAGNI—cut unnecessary scope from all approaches.

• [NEVER] Generate approaches without codebase evidence.

[4][SELECT]

Dictum: Committed direction enables focused planning.

Select best approach via weighted criteria:

[INDEX] Criterion Weight Evaluation

[1] Pattern alignment High Matches existing codebase conventions

[2] Research support High Backed by high-confidence findings

[3] Simplicity Medium Minimal moving parts

[4] Risk profile Medium Low-confidence areas minimized

Selection Output:

• Selected approach name

• Primary rationale (1-2 sentences)

• Key trade-off accepted

[CRITICAL]:

• [ALWAYS] Commit to ONE approach—no hedging.

• [ALWAYS] Document trade-off accepted.

• [NEVER] Defer selection to downstream phases.

[5][OUTPUT]

Dictum: Downstream consumers require predictable structure.

Produce brainstorm.md with structure:

[H1][DESIGN]: [Title]

Dictum: [Build target—refined from request]

<br>

Research Summary: [Key findings relevant to design]

[1][APPROACHES]

[1.1][APPROACH_A]: [Name]

[1.2][APPROACH_B]: [Name]

[2][SELECTED_APPROACH]

[3][DESIGN_CONSTRAINTS]

[4][KEY_DECISIONS]

[CRITICAL]:

• [ALWAYS] Include all sections—downstream depends on structure.

• [ALWAYS] Table format for approaches and decisions.

• [NEVER] Prose paragraphs—tables and lists only.

[6][VALIDATION]

Dictum: Incomplete synthesis cascades errors downstream.

[VERIFY]:

• Ingest: Research parsed, request intent extracted

• Scan: 3-4 agents dispatched in ONE message

• Explore: 2-3 approaches with trade-offs generated

• Select: ONE approach committed with rationale

• Output: All sections present, table format used

• YAGNI: Unnecessary scope cut from all approaches