nemawashi-collect

Profile Collector

Collect data about a person from available MCP sources and create or update their profile in PROFILE_DIR/.

When to Use

• User says "create a profile for X"
• User says "update X's profile"
• User says "collect info about X"
• Another skill (e.g., nemawashi-reply) detects a stale or missing profile

Process

Step 1: Identify the Target

Ask the user:

• Who is the target person?
• What is their role/team? (if not already known)
• Any specific MCP sources to prioritize? (default: all available)

Step 2: Check Existing Profile

Read PROFILE_DIR/<person-name>/profile.md if it exists. Note what data is already collected and when it was last updated.

Step 3: Collect Data from MCP Sources (Adapter-Driven)

Read all *.md files in the adapters/ directory (relative to this skill). For each adapter:

1. Check the adapter's MCP Tools Required section against the tools available in this session.
2. If all required tools are present, follow the adapter's Collection Recipe and apply its Fact Extraction rules.
3. If any required tools are missing, skip the adapter and note the gap in the final report.

After running every known adapter, scan for additional MCP servers that look communication-relevant but are not covered by an adapter (e.g., Discord, Linear, Notion, Drive). For each uncovered source:

1. Identify search/read tools that can be queried by the target's identifiers (name, email, handles).
2. Perform best-effort collection following the same shape as the known adapters.
3. Tag fact entries with [<slug>] where <slug> is derived from the MCP server name (e.g., [discord], [linear]).
4. If a source proves useful repeatedly, recommend writing a dedicated adapters/<slug>.md file for it (see ADAPTER-CONTRACT.md).

Step 4: Write Profile Files

Create the directory PROFILE_DIR/<person-name>/ if it doesn't exist.

profile.md — Write with this structure. These are the manual sections that nemawashi-analyze preserves verbatim across re-analyses; the synthesis pass adds ## Core Pattern and ## Framework Summary underneath without touching these.

---
name: [Full Name]
role: [Role/Title]
team: [Team Name]
last_updated: [YYYY-MM-DD]
sources: [list of MCP sources used]
---

## Basic Info
- Title: [title]
- Reports to: [manager, if known]
- Tenure: [if known]

## Communication Patterns
- [Observed patterns from collected data]
- [Response time tendencies]
- [Preferred channels]
- [Tone and formality level]

## Active Channels
- [Slack channels, email threads, repos where this person is active]

## Work Patterns
- [Meeting cadence, focus hours, escalation patterns — fill what's observable]

Leave the analysis sections (## Core Pattern, ## Framework Summary) to nemawashi-analyze; the synthesis pass writes them. Per-framework files at frameworks/<slug>.md are also produced by nemawashi-analyze, not here.

relationship.md — Ask the user about their relationship:

---
person: [Full Name]
relation: [relationship description]
last_updated: [YYYY-MM-DD]
---

## Relationship with Me
- [User-provided context about their relationship]

## Approach Strategy
<!-- Populated after nemawashi-analyze runs -->

facts.jsonl — Append one observable behavior per line, following the canonical schema in FACTS-SCHEMA.md. New entries always go here; never modify a legacy facts.md if one is present (consumers read both).

Entry format rules:

• One JSON object per line (JSONL); no top-level metadata block — person info lives in profile.md.
• Required fields: date (YYYY-MM-DD or YYYY-MM), source (the adapter's output_tag or manual for user-provided facts), content (one observable behavior — what they said or did).
• Optional fields are adapter-specific — see FACTS-SCHEMA.md and each adapter file.
• Append-only. Don't read-modify-write the whole file just to add an entry.

Examples:

{"date":"2026-03-27","source":"slack","content":"Rear-guard criticism to subordinate about alert handling","url":"https://acme.slack.com/archives/C123/p456","channel":"#engineering"}
{"date":"2026-03-27","source":"gmail","content":"Accepted meeting invitation for design review as passive participant"}
{"date":"2026-03-26","source":"calendar","content":"Attended ops meeting organized by team lead","meeting_title":"Ops Weekly"}

Step 5: Suggest Next Steps

After collection, suggest:

• "Profile created. Run nemawashi-analyze to identify behavioral patterns?"
• If the profile already existed: summarize what was updated

Parallel Dispatch (Batch / multi-target Mode)

When you need to collect for multiple targets at once — invoked by nemawashi-discover Step 5, or directly by the user as "collect for X, Y, Z in parallel" — do not run the per-adapter recipes inline N times. Dispatch the profile-collector agent (defined at agents/profile-collector.md) once per target. Each agent runs in an isolated context and writes only that target's profile files.

The parent (this skill running in the main session) is responsible for:

1. Filtering adapters by which MCP tools are available in this session.
2. Composing a uniform prompt per target — see the prompt template below.
3. Dispatching N profile-collector agents in parallel (one Agent tool invocation per target, all in a single message).
4. Aggregating the one-line reports each agent returns. Surface failures and skipped adapters in the final summary.

profile-collector dispatch prompt template

target_name: <slug>
target_full_name: <Full Name>
target_role: <Role or "unknown">
profile_dir: ${PROFILE_DIR}/<slug>/
today: <YYYY-MM-DD>

relationship_hint: <one-liner from the user, optional>

Adapters available in this session: <comma-separated slugs>

Read the adapter files at skills/nemawashi-collect/adapters/<slug>.md for the
above slugs (plus ADAPTER-CONTRACT.md and FACTS-SCHEMA.md). Follow each
adapter's Collection Recipe + Fact Extraction. The recipes resolve the
TARGET's identifiers; "user's own handle / email / login" only matters
for a couple of recipe steps (e.g. GitHub's `@me` filter), which the
tooling resolves natively — no separate identity cache is needed.

Return one report block per the agent's documented output shape.

Sub-agents read the adapter files at the canonical path, so recipe drift is impossible.

Key Principles

• Collect, don't interpret — This skill gathers raw data. Leave analysis to nemawashi-analyze.
• Cite sources — Every fact in facts.md must have a source tag ([slack], [gmail], etc.)
• Ask about relationship — Only the user knows their subjective relationship with the person. Always ask.
• Respect privacy — Only collect data from sources the user has MCP access to. Do not speculate beyond available data.