Architecture Review

Architecture Decision Record — Subagent Debate Skill

Structured persona debate → synthesis → markdown ADR (+ optional HTML report) → registry index.

Works with any coding agent that can run subagents. No external API calls required.

Supporting files (read when needed):

• personas.md — persona and synthesis prompts
• adr-template.md — markdown ADR output format

Step 0 — Elicit the Decision

Ask the user (in one message, all at once):

1. What is the decision? (e.g. "FTS5 vs pgvector for legal document search")
2. What are the hard constraints? (latency, cost, team size, existing infra, timeline)
3. What is the context? (stage of product, scale, who will maintain this)
4. Do you have a leaning already? (optional — being honest about priors helps)
5. What does "good" look like? (what would make this decision clearly correct in 6 months)

Do not proceed until you have at least 1, 2, and 3.

Also resolve output settings (infer from request if stated, else use defaults). See ../OUTPUT.md for the shared convention across all skills.

Create the output directory if it does not exist.

Step 1 — Brief All Subagents

Construct a decision brief from the user's answers:

DECISION: [what is being decided]
CONSTRAINTS: [hard constraints]
CONTEXT: [product stage, scale, team, infra]
OPTIONS: [Option A] vs [Option B] (vs [Option C] if applicable)
SUCCESS CRITERIA: [what good looks like]
USER LEANING: [if stated, else "none stated"]
INCLUDE_SRE: [yes/no]

Step 2 — Run the Debate

Read personas.md. Run subagents in parallel when supported.

Always run (3 subagents): Pragmatist, Architect, Contrarian

Optionally run (4th subagent): SRE — when INCLUDE_SRE: yes

Pass each subagent:

1. The decision brief from Step 1
2. The matching persona prompt from personas.md
3. Instruction: respond using the format in the persona prompt

Keep subagent contexts isolated — no persona sees another's response until synthesis.

Step 3 — Synthesis Pass

After all persona subagents return, synthesize using the Synthesis prompt in personas.md.

Preferred: Run a synthesis subagent with all persona responses as context.

Fallback: Synthesize yourself as the orchestrating agent using the same format.

Step 4 — Write Markdown ADR

Always save a markdown ADR. Fill in adr-template.md from synthesis output and persona responses.

Path: {output_dir}/ADR-NNNN-[decision-slug].md

ADR numbering: Scan {output_dir}/ for existing ADR-*.md files. Use the next sequential number (ADR-0001, ADR-0002, …). Start at ADR-0001 if none exist. The date lives inside the document, not the filename.

Step 5 — Generate HTML Report (optional)

Skip if user chose ADR-only mode.

Path: {output_dir}/ADR-NNNN-[decision-slug].html (same NNNN and slug as the .md)

Link to the markdown ADR from the HTML header.

Design spec

Page bg:        #f8f7f5
Card bg:        #ffffff
Border:         1px solid #e8e5de
Primary text:   #111111
Muted:          #6b6b6b
Hint:           #a8a39a
Pragmatist:     #111111
Architect:      #4a5568
Contrarian:     #744210
SRE:            #2c5282
Synthesis:      #1a3a2a
Font:           Georgia/serif for debate sections + system-ui for chrome
NO gradients. NO purple. Light and premium.

Report sections

HEADER — "Architecture Decision Record" · date · link to .md · "Export PDF" (print)

DECISION TITLE — large, prominent

CONTEXT STRIP — 3 pills: constraints, product stage, success criteria

THE DEBATE — columns per persona (3 or 4). Each column: icon, POSITION, ARGUMENT, risks/ops burden, CONFIDENCE badge, CAVEAT in italics.

TRADEOFF MATRIX — rows = evaluation dimensions; columns = options; cells = ✓ / ~ / ✗ + note

DECISION DIAGRAM — SVG/CSS comparison cards with CHOSEN highlight

SYNTHESIS — background #f0ede8: VERDICT, REASONING, agreed/decisive/dissent blocks, CONDITIONS

SIGNED ADR — monospace card with all ADR fields

FOOTER — Architecture Decision Record · generated [timestamp]

Step 6 — Update Decision Registry

Maintain {output_dir}/index.md as a chronological registry of all ADRs.

If index does not exist, create:

# Architecture Decision Records

| ADR | Date | Status | Decision | Verdict |
|-----|------|--------|----------|---------|

After each new ADR, append a row (do not duplicate if re-running the same file):

| [ADR-NNNN: Title](ADR-NNNN-decision-slug.md) | YYYY-MM-DD | Accepted | [options compared] | [one-line verdict] |

Sort rows by date descending (newest first). If updating an existing ADR in place, update its row rather than appending a duplicate.

Step 7 — Output to User

1. List saved file paths (.md, .html if generated, index.md updated)
2. State the verdict in one sentence
3. Name the decisive factor
4. Note the review trigger
5. Mention whether SRE persona participated

Edge Cases

• More than 2 options: give all options to every persona
• Strong user leaning: tell the Contrarian explicitly; ask them to challenge it
• Ambiguous decision: output a decision rule ("use A if X, use B if Y"), not a fake pick
• Re-run request: "argue the other side" — re-run debate with swapped priors; write a new ADR or update status to Superseded on the old one
• Subagents unavailable: run personas sequentially in isolated turns, or simulate each persona before synthesizing (note lower quality to user)
• Custom output dir missing: create it; warn if path is outside the project
• Re-run same decision: write a new ADR number; mark the old ADR Status as Superseded with a link to the new one

Invocation Examples

"should I use FTS5 or pgvector for legal search"
"ADR: Electron vs web app for our desktop client"
"help me decide between Redis and BullMQ for job queuing"
"debate this: do we build our own auth or use Clerk"
"architecture decision — monorepo vs separate repos"
"include SRE — managed Postgres vs self-hosted"
"ADR only"
"save to skill-outputs/architecture-review"