Skill Architect

+1: Sage — The Creative Provocateur (optional)

Activate Sage when the user says "add Sage," "I want the creative take," "give me the wild ideas," "+1," or when you sense the problem space would benefit from a fundamentally different angle — especially for Architecture mode.

When Sage is active, spawn 6 subagents instead of 5. The synthesizer gives Sage's novel proposals their own section in the output — "Creative Alternatives" with tradeoff analysis, not blended into the convergence analysis.

Orchestration Workflow

Step 1: Detect Task Mode

Determine which mode the user needs: feedback, architecture, or testing. If ambiguous, ask.

Step 2: Gather Inputs

Depending on mode:

• Feedback: The user's existing skill (SKILL.md + any referenced files). Read the full skill first yourself to understand what you're handing to the assessors.
• Architecture: The user's description of what the skill should do, plus any reference documents or source materials they want the skill to work with. Ask clarifying questions if the scope is unclear.
• Testing: The user's skill (existing or draft) plus any existing eval files. If no evals exist yet, that's fine — the skill-architect will propose them from scratch.

Step 3: Prepare Input Briefing (conditional)

If the total input materials from Step 2 exceed ~500 lines, produce a structured briefing before spawning assessors:

1. Summarize each input file: purpose, key content areas, estimated size
2. Identify the most important sections and domain-specific details that assessors need to engage with
3. Note any relationships between files (e.g., "the tutorial covers the same content as the utils README but in a different structure")

Assessors receive the briefing as their primary input, with pointers to the original files for targeted deep-reads when they need specific details.

If input materials are under ~500 lines, skip this step and pass the original files directly to assessors.

Step 4: Spawn Assessor Subagents

Spawn the assessor panel in parallel:

• Default: 5 subagents (Riley, Morgan, Jordan, Alex, Avery)
• With Sage activated: 6 subagents (add Sage — The Creative Provocateur)

Each subagent is spawned using the agents/assessor.md agent definition. Each receives:

• Their persona file (from agents/personas/{name}.md)
• The task brief for the current mode (agents/task-briefs/{mode}.md)
• The user's input materials (skill files, reference docs, description — whatever was gathered in Step 2)

Assessors load their own reference materials (evaluation criteria, skill design principles, advanced progressive disclosure) as part of their assessment process. All subagents receive the same task description (mode + user context + what to assess). The differentiation comes from each persona's thinking style and rubric dimension assignments.

Step 5: Synthesize

Once all assessors complete, spawn the synthesizer using agents/synthesizer.md. The synthesizer receives:

• All assessor outputs
• The original user inputs
• The current task mode
• Whether Sage is active (if yes, the synthesizer adds a "Creative Alternatives" section)

The synthesizer produces the final deliverable as structured JSON, which is then used to generate an HTML report.

Step 6: Generate and Present Report

After synthesis, generate a styled HTML report:

1. Structure the synthesizer's output as JSON matching the schema expected by scripts/generate-report.py
2. Save the JSON to the skill's target directory as report-data.json
3. Run: python scripts/generate-report.py report-data.json --open
4. The report opens in the user's browser

The report format varies by mode:

After presenting, invite the user to discuss, ask questions, and iterate on the proposal.

Step 7: Consolidate and Act

After the user has worked through discussion and made decisions, automatically consolidate what was agreed upon:

1. Capture the final decisions — what was chosen, how open questions were resolved, any modifications from the original proposal
2. Write the consolidated output to a file in the skill's target directory
3. Notify the user: "I've captured the finalized recommendations at [path]."

Then offer the user a choice:

• "Implement now" — Apply the recommendations directly. What gets built depends on the mode.
• "Create a project plan for later" — Generate a phased project-plan.md with implementation steps sequenced by dependency and priority.

What "implement now" means per mode:

Fallback Behavior

If subagent spawning fails or the environment doesn't support it:

1. Run each assessor's analysis sequentially (not in parallel) — same personas, same task briefs, but executed one at a time within the main conversation.
2. Note to the user that this is a degraded mode with less independence between assessors.
3. Still run the synthesis step to produce a unified recommendation.

This is less rigorous (later assessors have context of earlier ones, reducing independence), but still provides multi-perspective value.

Prerequisites

• Skill-creator skill (optional): The skill-architect designs and evaluates; the skill-creator builds, tests, and iterates. Typical workflow: skill-architect produces recommendations or blueprints, then hands off to skill-creator for implementation and eval runs. In testing mode, the skill-architect saves a testing-strategy.md file to the skill's directory; the skill-creator checks for this file when writing evals and uses it to guide eval design (coverage priorities, metrics, flakiness expectations). If no strategy file exists, the skill-creator writes evals from scratch.

Skill Directory