Deepresearch

/deepresearch

Effectiveness-first deep investigation with live data — not stale training data.

MiniMax MCP is the primary research tool. Prefer mcp__MiniMax__web_search for live research whenever it is available.

MAXPARALLELAGENTS — ceiling for deep research lanes. Use only the number of distinct branches that materially improve the final investigation.

Use when: The user explicitly asks for deep research, high-quality investigation, landscape analysis, architecture research, due diligence, adversarial fact-finding, or any research task where the first search summary is not enough.

Research-First: Integrated into /workflow automatically. Use /deepresearch directly when the main job is investigation.

Swarm: "swarm deepresearch" → /deepresearch with an effectiveness-first research wave up to MAX_PARALLEL_AGENTS.

For narrower current-fact lookups, use /webresearch. For Claude, Claude Code, Claude.ai, Anthropic API, connector, plugin, skill, hook, MCP, subagent, availability, limit, or setup questions, use /claudeproduct as the official product-docs branch, escalating back here only when the answer is multi-surface, strategic, implementation-driving, or conflict-heavy. For older prompts or backwards compatibility, /browse should route to this protocol or the /webresearch protocol depending on scope.

---

Core Contract

Deep research is not a search dump.

It must:

• start from the current minmaxing time anchor when claims depend on "today",

"latest", "recent", "current", "SOTA 2026", or another relative date

• start with a collaborative research plan
• choose an effective budget up to MAX_PARALLEL_AGENTS
• run an iterative search -> read -> refine loop
• maintain a source ledger
• surface conflicting evidence
• do follow-up research when key unknowns remain

Temporal guard:

• Use the minmaxing temporal anchor injected by Claude Code hooks, or run

bash scripts/time-anchor.sh text, before making time-sensitive claims.

• Record the anchor date in the research output when the work concerns SOTA

2026, current providers/models/APIs, pricing, laws, benchmarks, docs, schedules, or news.

• Do not use pretrained memory as evidence for current or "best latest" claims.

If live verification is unavailable, mark the claim insufficient_data, stale, or unverified.

This protocol is effectiveness-first:

• more tracks only when they reduce uncertainty
• more depth only when it changes the plan
• more loops only when open questions remain decision-relevant

Investigation Modes

• standard — a few decisive branches, moderate depth, usually 1-3 loops
• comprehensive — multiple branches, high-stakes or strategic work, explicit pressure-testing before conclusions

Default to comprehensive when the user says "deepresearch", "investigate deeply", "top notch quality", "reverse engineer", "due diligence", or when errors from shallow research would be expensive.

Step 1: Memory Recall

bash scripts/memory.sh recall "[research topic]" --depth simple 2>/dev/null || echo "Memory recall: skipped"
bash scripts/memory.sh search "[topic]" 2>/dev/null || true

Step 2: Collaborative Research Plan

Before the first search wave, define:

• deliverable
• main research branches
• source classes to consult
• likely contradictions / unknowns
• stop condition for "research is sufficient"

Example:

### Collaborative Research Plan
- Deliverable: [what this research must unlock]
- Branches:
  - [branch 1]
  - [branch 2]
- Source Classes:
  - official docs
  - recent practitioner writeups
  - issues / discussions
- Stop Condition:
  - [what must be known before we move on]

Step 3: Effective Budget

Choose the smallest useful wave:

effective_research_budget = min(MAX_PARALLEL_AGENTS, distinct_branches, reviewer_capacity)

Good defaults:

• 1-2 lanes: narrow technical verification
• 3-4 lanes: architecture or implementation-pattern comparison
• 5+ lanes: due diligence, market / ecosystem / adversarial investigations

Do not fill the pool for theater.

Step 4: Discovery Wave

Launch only distinct tracks. Good track types:

• official documentation
• recent release notes / migrations
• GitHub issues / discussions
• expert writeups / case studies
• alternatives / competing patterns
• risk / failure-mode lookup

Step 5: Search -> Read -> Refine Loop

Minimum loop:

1. discovery
2. deep read
3. refine or pressure-test

Use more loops only when:

• evidence conflicts
• key claims remain weak
• the plan still depends on unknowns

Step 6: Source Ledger

Track:

• cited sources
• reviewed but not cited sources
• rejected or downweighted sources

This is required whenever external facts materially affect the conclusion.

Step 7: Synthesis

Before final synthesis, run /introspect pre-plan inline as a hard gate.

Challenge:

• weak sources or over-weighted secondary sources
• unresolved contradictions
• premature certainty after shallow evidence
• missing source classes
• conclusions that do not actually follow from the source ledger

If the introspection pass finds unresolved blockers, do follow-up research before presenting the final result.

## DeepResearch: [Topic]

### Investigation Mode
[standard / comprehensive]

### Collaborative Research Plan
- Deliverable: ...
- Branches: ...
- Stop Condition: ...

### Research Tracks
| Track | Query | Sources |
|-------|-------|---------|

### Loop Log
| Loop | What changed | Why it mattered |
|------|--------------|-----------------|

### Source Ledger
- Cited: ...
- Reviewed but not cited: ...
- Rejected / downweighted: ...
- Time Anchor: [local_time from hook or scripts/time-anchor.sh]
- Access Date: [YYYY-MM-DD for live sources]

### Conflicting Evidence
- Claim A: ...
- Claim B: ...
- Resolution / open uncertainty: ...

### Implications
- ...

### Coverage
- Research Tracks Used: [completed] / [effective budget]
- MiniMax MCP Searches: [count]
- Follow-up Research: [not needed / completed / blocked]

Quality Gates

• cite sources
• keep the collaborative research plan visible
• use MAX_PARALLEL_AGENTS as a ceiling, not a quota
• show what changed between loops
• surface conflicting evidence instead of smoothing it away
• do follow-up research before finalizing when uncertainty still matters
• run the introspection gate before final synthesis when the research will drive planning, implementation, or a high-stakes decision

Anti-Patterns

• one-shot search and summary
• slot-filling to hit MAX_PARALLEL_AGENTS
• no source ledger
• no conflict handling
• calling the result "deep research" when only one shallow wave ran
• finalizing with unresolved /introspect blockers

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: waitdeadai
• Source: waitdeadai/minmaxing
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.