Research

Research

Objective codebase investigation that produces factual findings blind to the intent expressed in the discovery brief. The lead reads the brief and generates neutral questions. Teammates investigate the codebase. The lead compiles findings. No teammate ever sees the brief, the ticket, or the user's goals.

Environment

Agent teams: !echo ${CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS:-disabled}

Prerequisites

1. If agent teams shows disabled above, stop immediately and tell the user:

   Agent teams are required for parallel research. Enable them by adding to your settings:

   { "env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" } }
   

   Then restart the session.

   Teammates also need an interactive session. Under claude -p, a named spawn runs as an ordinary subagent — the waves still work, but findings return as subagent results instead of SendMessage.

2. Locate the brief:

   • Discovery handoff — the brief is already in conversation context. Use it directly.
   • Standalone invocation — $ARGUMENTS names a brief, or ask the user to provide one. Read it into conversation context — the skill works from context.
   • If no brief exists in conversation or on disk, ask the user whether to run discovery first.

Process

Phase 1 — Plan and Dispatch

1. Read the brief from conversation context.
2. Identify the codebase scope clusters the brief touches (modules, directories, subsystems).
3. For each scope, formulate the questions worth investigating. Follow the question generation rules below.
   • Multiple questions per scope is normal. Group questions that share files to avoid duplicate reads across teammates.
   • Don't fix a question count up-front. Let the brief's complexity determine how many.
   • The brief's Questions for research list (facts discovery deferred) is mandatory input — every entry becomes an investigation question, rewritten through the same neutrality rules as any other.
4. Self-check questions against the bias validation gate.
5. For each scope cluster, create a task via TaskCreate using the task description format below, then spawn one teammate via Agent with subagent_type: "codebase-researcher", name: "researcher-{scope-slug}", and an intent-free spawn prompt (see Spawn Prompt below).
   • The name is what makes the spawn a teammate, and it is the address other agents use to message it. There is no team to create — the session owns one team — and team_name is ignored.
   • Task descriptions contain only questions and scope boundaries. Never the brief, ticket, or intent.
   • Explicit subagent_type ensures the teammate inherits the agent's tool restrictions and system prompt; without it, the platform falls back to the general-purpose agent.
   • If this session has no Task tools, skip TaskCreate and put the task description block in the spawn prompt instead. It carries only questions and scope, so the information barrier holds.

Phase 2 — Collect and Iterate

Research proceeds in waves. A wave is a set of teammates dispatched together. Never start a new wave while the current one is still running.

6. Wait until all teammates in the current wave have sent their findings. A teammate's idle notification carries no output — going idle is not a finding. The wave is in only when every question has an answer by SendMessage or an explicit "not determinable from files". If a teammate went idle with questions unanswered, message it by name and ask for the missing findings.
7. Once the wave is fully in, evaluate:
   • Are any questions still unanswered or have only partial answers?
   • Did any finding reveal a new scope or codebase area that wasn't initially identified?
   • Are there new questions that emerged from the findings (not from the brief) that need investigation?
8. If yes to any of the above, plan and dispatch a new wave: identify new scopes, formulate questions for them, run the bias self-check, create tasks, spawn new teammates with fresh names. Then return to step 6 for the new wave.
9. If no further investigation is needed, proceed to Phase 3.

Why waves: dispatching follow-ups while a wave is still running causes redundant investigation — a sibling teammate may already be answering the question you're about to ask. Batching by wave lets findings inform each other.

Phase 3 — Compile and Persist

By this point all waves are complete and no further teammates will be dispatched. The information barrier no longer applies.

10. Compile all findings into design-docs/NN-name.research.md as sole author using the document format below.
    • Organize by scope cluster, not by teammate.
    • Preserve concrete file paths, function names, and line references from teammate findings.
    • Flag contradictions or gaps explicitly.
    • Do not inject opinions, recommendations, or solution ideas.
11. If the brief was not yet written to disk (discovery handoff flow), write it now as design-docs/NN-name.brief.md.
12. Present the compiled research.md to the user for review.

Phase 4 — Shutdown

13. Ask every teammate to shut down, by name. A teammate can reject a shutdown request with an explanation — if one does, resolve what it reports before continuing.

    There is no team cleanup step. The session's team directories are removed when the session ends.

Transition

14. On user approval of research.md, offer: "Proceed to alignment skill?"

Question Generation Rules

Every question must request a fact, never an opinion, assessment, or judgment. Teammates can only return facts; asking them for evaluation produces speculation, not evidence. If the brief frames something as a problem to solve or a quality to improve, translate it into factual questions about what currently exists — never pass the framing through.

Verb Rules

Use: "document", "identify", "describe", "trace", "characterize", "examine", "list", "map"

Ban: "should", "must", "improve", "fix", "optimize", "enhance", "refactor", "recommend", "suggest", "better", "needs to"

Investigation Types

Each question targets one of these:

• Structure — what components exist, how they are organized, what their interfaces are
• Logic — execution paths, control flow, conditional branches as written
• Integration — what connects to what, dependencies, data flow across boundaries
• Patterns — conventions followed, architectural patterns used, consistency or inconsistency

Bias Validation Gate

Before presenting questions to the user, self-check:

1. Prescriptive language scan — does any question contain banned verbs or imply a desired outcome? Rewrite it.
2. Intent masking test — if someone read only the questions without the brief, could they infer the solution direction? If yes, the questions leak bias. Rewrite to ask about what exists, not what needs to change.
3. Balance check — do questions cover at least three of the four investigation types? If they cluster in one, add questions for underrepresented types.

Scope Clustering

Group questions by the codebase area they investigate, not by topic.

Heuristic

1. Identify the primary directories or modules each question targets.
2. Questions targeting the same directory tree or tightly coupled modules → same cluster.
3. Questions spanning the whole codebase → assign to the scope where the pattern is most concentrated, or create a dedicated cross-cutting cluster.
4. Name each cluster by its codebase area: "auth module", "API layer", "data persistence", etc.

Task Description Format

Each task description dispatched to a teammate must include exactly:

## Questions
1. [question text]
2. [question text]

## Scope
Directories: [paths]
Focus: [one-line description of the codebase area]

## Instructions
Investigate these questions by reading files within the scope above. Send one structured finding per
question to the team lead via SendMessage. Mark the task complete when all questions are answered.

Nothing else. No brief context, no intent, no background.

Spawn Prompt

The prompt parameter on the Agent call is visible to the teammate. The leak we care about is intent, not context. Distinguish:

• Context (allowed) — what the teammate is doing, where they're working, what role they play. The project name, the scope area they'll investigate, the fact that they're doing read-only codebase research. None of this biases the investigation.
• Intent (forbidden) — why we're investigating, what problem we're trying to solve, what change we're considering, what outcome we want, any framing that suggests a desired direction. This is what taints the research.

The teammate already gets stable context from its agent definition (role, tools, output format). The spawn prompt can add light orientation — project name, scope tag, "claim a task" routing — but every word should pass the test: would removing this change what the teammate looks for? If yes, it's intent. If no, it's safe context.

Research Document Format

# Research: {title from brief}

- **Date:** {date}
- **Teammates:** {count}
- **Scopes investigated:** {list of scope cluster names}

## {Scope Cluster Name}

### Files Investigated

- `path/to/file.ts` — {what it contains}

### Findings

**{Question text}**

{Compiled finding — concrete, factual, with file:line references}

### Patterns

{Patterns observed — include prevalence (e.g., "dominant in 12/15 files" vs "isolated instance")}

### Integration Points

{How this scope connects to other parts of the codebase}

### Open Gaps

{What could not be determined}

## Cross-Cutting Observations

{Patterns or findings that span multiple scopes — only if such observations exist}

## Unresolved Questions

{Questions that teammates could not answer from the files — only if any remain}

Rules

• Write the brief to disk only in Phase 3, after all waves complete. Teammates have Read/Grep/Glob and can encounter any file in the project. The brief contains intent — if it exists on disk during investigation, the information barrier is broken.
• Task descriptions contain only questions and scope boundaries — teammates never see the brief.
• Spawn prompts carry context, never intent. The prompt parameter on Agent is visible to the teammate. Project, scope area, role are fine. Goals, problems being solved, desired outcomes, or any framing that suggests a direction bias the investigation.
• You are sole author of research.md. No teammate writes to any file.
• research.md contains factual observations with file references — interpretation belongs in the alignment stage.
• An idle teammate is not a finished teammate. The idle notification carries no output. Count a scope answered only when its findings arrived by SendMessage.
• Batch by wave, not by urgency. A follow-up question that seems urgent may already be under investigation by a sibling teammate. Wait for the full wave before deciding what's still unanswered.