# Deepresearch

> A Claude skill from waitdeadai/minmaxing.

- **Type:** Skill
- **Install:** `agentstack add skill-waitdeadai-minmaxing-deepresearch`
- **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior
- **Seller:** [waitdeadai](https://agentstack.voostack.com/s/waitdeadai)
- **Installs:** 0
- **Category:** [AI & ML](https://agentstack.voostack.com/c/ai-and-ml)
- **Latest version:** 0.1.0
- **License:** Apache-2.0
- **Upstream author:** [waitdeadai](https://github.com/waitdeadai)
- **Source:** https://github.com/waitdeadai/minmaxing/tree/main/.claude/skills/deepresearch

## Install

```sh
agentstack add skill-waitdeadai-minmaxing-deepresearch
```

Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent.

## About

# /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.

**MAX_PARALLEL_AGENTS** — 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
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:

```markdown
### 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:

```text
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.

```markdown
## 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](https://github.com/waitdeadai)
- **Source:** [waitdeadai/minmaxing](https://github.com/waitdeadai/minmaxing)
- **License:** Apache-2.0

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

## Pricing

- **Free** — Free

## Security capabilities

Automated source analysis of v0.1.0 — what this tool can access:

- **Network access:** no
- **Filesystem access:** no
- **Shell / process execution:** no
- **Environment & secrets:** no
- **Dynamic code execution:** no

*"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.*


## Versions

- **0.1.0** — security scan: passed — Imported from the upstream source.

## Links

- Listing page: https://agentstack.voostack.com/l/skill-waitdeadai-minmaxing-deepresearch
- Seller: https://agentstack.voostack.com/s/waitdeadai
- Browse the marketplace: https://agentstack.voostack.com/browse

---
Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.