meta-research

Meta-skill that chains both research atomic skills in order, taking a project from open questions to a countered, evidence-backed synthesis.

The main chat context stays lean — all heavy research work happens in a dispatched Opus subagent.

Chain

research-execute Phase 0 (plan) -> [approval gate] -> write prompt -> dispatch Opus -> present results

Inputs

Input	Source	Required
project-context.md	Project root	Yes
Existing code	src/ directory	No
Prior research	artifacts/research/ directory	No

Instructions

Phase 1: Research Plan

Run Phase 0 (plan) of ../research-execute/SKILL.md.

Analyze the project context, extract research topics, categorize by lane (Academic / Code / Both), map to connectors, prioritize (P0/P1/P2), and self-counter the plan.

Exit condition: artifacts/research/{NNN}/research_plan.md exists with all topics categorized, prioritized, and mapped to connectors. The plan has been self-countered. The user has approved the scope.

Transition: Proceed to Approval Gate.

Approval Gate

Present this choice to the user:

"Research plan is ready with N topics (P0: X | P1: Y | P2: Z).

1. Execute all — run every topic through its mapped connectors
2. P0 only — run only the blocking topics, skip P1/P2
3. Cherry-pick — tell me which topics to run
4. Stop here — keep the plan, execute later with /research-execute"

Wait for the user's answer. Do not assume.

• Option 1: Proceed to Phase 2 with the full plan.
• Option 2: Proceed to Phase 2 but filter to P0 topics only.
• Option 3: Ask the user which topic numbers to include, then proceed to Phase 2 with the filtered set.
• Option 4: Confirm the plan is saved and exit.

Phase 2: Write Research Prompt

After approval, write artifacts/research/{NNN}/research-prompt.md — max 200 lines. This file is the ONLY input the Opus subagent reads to understand the full scope. Maximize conciseness without context loss.

Schema:

# Research Prompt — {NNN}

## Scope Decision
[Which option the user chose: all / P0 only / cherry-pick (list)]

## Research Plan Reference
artifacts/research/{NNN}/research_plan.md

## Topics to Execute
[List each topic with its priority, lane, and mapped connectors — copied from
the approved plan, filtered per the user's scope decision]

## Project Context Summary
[Compressed relevant sections from project-context.md — first 3 sections max]

## Source Counting Target
Target: 300+ sources scanned across all connectors.
Each connector subagent must follow the multi-query protocol (3-5 queries per
topic) and include a Source Tally table in its output.

## Output Configuration
- Research folder: artifacts/research/{NNN}/
- Summary destination: artifacts/research/summary/{NNN}-{topic-slug}.md
- Source tally: artifacts/research/{NNN}/source-tally.md

## Special Instructions
[Any user-specified constraints, focus areas, or modifications from the
approval gate discussion]

Phase 3: Dispatch Opus Subagent

Spawn exactly ONE Opus subagent. Read agents/opus-orchestrator.md for the prompt template — fill in the absolute skill path, NNN, and topic slug.

Do NOT pass the full research protocol in the prompt — the subagent reads it from the SKILL.md file. This keeps the dispatch lightweight.

Phase 4: Present Results

When the Opus subagent completes:

1. Read artifacts/research/summary/{NNN}-{topic-slug}.md
2. Read artifacts/research/{NNN}/source-tally.md
3. Present the executive summary with source counts to the user
4. Highlight gaps and low-confidence areas
5. Flag any counter-arguments that need human judgment

Then offer next steps:

"Research complete. {N} sources scanned | {N} cited across {N} connectors.

1. Build the plan — run /build-plan to create an implementation plan informed by these findings
2. Dive deeper — re-run specific topics with different connectors or broader queries
3. Go deep — run /meta-deep-research for exhaustive multi-model adversarial research on a specific question
4. Done for now — stop here and use the findings when ready"

Wait for the user's answer. If they choose 1, tell them to run /build-plan. If they choose 2, ask which topics and re-run Phase 2-3 for those topics only. If they choose 3, suggest a research question for deep research. If they choose 4, confirm and exit.

Error Handling

• If project-context.md does not exist, tell the user to run /project-context or /meta-init first. Do not proceed without it.
• If prior research exists in artifacts/research/, detect it and ask the user whether to build on it or start fresh.
• If the Opus subagent fails or times out, read whatever intermediate files exist in artifacts/research/{NNN}/ and present partial findings with the source tally so far.
• If the user already has a research_plan.md, ask whether to use the existing plan or create a new one. If using the existing plan, skip to Phase 2.

Examples

User: "Research this project"
Action: Read project-context.md. Run research-plan to extract and prioritize
        topics. Present the approval gate. On approval, write research-prompt.md.
        Dispatch Opus subagent. Present summary with source counts when done.

User: "/meta-research"
Action: Same as above — full chain from plan to countered synthesis via Opus.

User: "Do all the research but only the critical stuff"
Action: Run research-plan. At the approval gate, auto-select option 2 (P0
        only) since the user indicated they want only critical topics. Confirm
        with the user before proceeding.

User: "We already have a research plan, just execute it"
Action: Detect existing research_plan.md. Confirm with the user. Skip to
        Phase 2 (write prompt + dispatch Opus) directly.

---

Before completing, read and follow ../references/cross-cutting-rules.md.