Multi-Lens Thinking

A six-stage pipeline that runs a question through multiple independent perspectives, synthesizes them, and writes a memory patch for the next session to confirm.

When to invoke

Invoke when the user's question matches any of these patterns:

• Decision framing: "should I", "is it worth", "好不好", "值不值得", "適不適合"
• Future/risk framing: "未來", "趨勢", "風險", "機會", "前景"
• Career / relocation / investment / major life choice
• Geopolitics, market thesis, industry analysis
• Explicit request: "用多維度分析", "幫我從幾個角度看", "lens"

Do NOT invoke for:

• Pure factual lookup (definitions, conversion, current price)
• Code or how-to ("how to install X", "syntax of Y")
• Casual conversation
• Tasks already covered by a more specific skill (the user explicitly invoked another skill)

Pipeline overview

Question
  ↓
[1. Router]      — reads persona.md + memory.md (no search)
  ↓ produces: answer_mode, user_snapshot, search_hints, active_nodes, skip_log
  ↓                     ↑
  ↓             (answer_mode is THE most important field; it controls
  ↓              Personal lens depth and Synthesizer output style)
  ↓
  ├─ [2a. Macro]      — WebSearch
  ├─ [2b. Personal]   — no search; behavior depends on answer_mode
  ├─ [2c. Local]      — WebSearch → anysearch fallback
  └─ [2d. Historical] — hybrid (LLM + targeted WebSearch)
  ↓
[3. Synthesizer] — integrates outputs; output style depends on answer_mode
  ↓
Final answer to user
  ↓
[4. Memory Updater] — writes a patch file (no search)

Steps 2a–2d run in parallel. Each is a separate Task tool sub-agent so contexts stay isolated.

User data location

This skill has two kinds of files:

• Skill files (this directory): SKILL.md, prompts/, templates/, README.md. Read-only when installed as a plugin.
• User data files: persona.md, memory.md, patches/. Must live somewhere the agent can both READ and WRITE.

The user data location (USER_DATA_DIR) is resolved at runtime in this priority order:

1. Connected Cowork folder + multi-lens/ subdir — if the user has selected (mounted) a folder in Cowork, use <mounted_folder>/multi-lens/. This is the recommended setup.
2. Skill install dir — fallback for read-only access to persona.md and memory.md if the user copied them there during install. Patches CANNOT be written here; they fall through to (3).
3. Outputs folder — last-resort fallback for writing patches. The agent must tell the user the path and instruct them to move the patch into their actual data dir manually.

If neither (1) nor (2) yields a readable persona.md, the skill stops and asks the user to either connect a folder via request_cowork_directory or place persona.md in the skill dir.

Execution procedure

The main agent (you) executes this skill by following these steps in order.

Step −1 — Resolve USER_DATA_DIR

Determine where persona.md, memory.md, patches/ live:

1. Check whether a Cowork folder is mounted (env shows a selected folder). If yes, USER_DATA_DIR = <mounted>/multi-lens/. Create the subdir if missing.
2. Else, check whether persona.md exists alongside this SKILL.md. If yes, USER_DATA_DIR = <skill-dir> (read-only — patches will fall through to outputs).
3. Else, stop and tell the user: "I need a persona.md to run. Please either (a) select/connect a folder via Cowork (recommended — patches will then be writable) and place persona.md inside <folder>/multi-lens/, or (b) place persona.md alongside SKILL.md in the skill install dir (patches will go to outputs/ and you'll merge manually)."

Internally remember:

• PERSONA_PATH = <USER_DATA_DIR>/persona.md
• MEMORY_PATH = <USER_DATA_DIR>/memory.md (create empty from template if missing and writable)
• PATCHES_DIR = <USER_DATA_DIR>/patches/ if writable; else <outputs>/patches/
• PATCHES_WRITABLE_AT_DATA_DIR = true/false

Step 0 — Check for pending memory patches

List files in <USER_DATA_DIR>/patches/. If any patch files exist that haven't been merged, show them to the user one at a time and ask "approve / reject / edit" before merging into memory.md. Skip this step silently if patches dir is empty or unreachable.

Step 1 — Run Router

Read these three files:

• <PERSONA_PATH>
• <MEMORY_PATH>
• <skill-dir>/prompts/01-router.md

Apply the router prompt to the user's question. The router output is a structured JSON-like block — keep it in your working memory; do NOT show it to the user verbatim.

Router output schema:

{
  "question_type": "decision | thesis | comparison | exploration | other",
  "answer_mode": "analytical | personal_decision | framework | meta",
  "user_snapshot": "<2–4 sentence digest of which parts of persona/memory matter for this question>",
  "search_hints": {
    "macro": "<keywords / regions / actors to focus on>",
    "local": "<location + vertical hint, e.g. 'Sydney + jobs+infrastructure'>",
    "historical": "<comparable era / event keywords>"
  },
  "active_nodes": ["macro", "personal", "local", "historical"],
  "skip_log": [
    {"node": "<name>", "reason": "<one-sentence justification>"}
  ]
}

answer_mode is the most important field. It is set by the question verb, NOT by the user's professional identity. A writer asking "analyze X" still gets analytical — Personal lens then contributes only voice; Synthesizer delivers the analysis itself, not a writing brief.

If active_nodes is empty (router decided no lens helps), tell the user and answer directly without the pipeline.

Step 2 — Spawn parallel sub-agents

For each node in active_nodes, launch a sub-agent via the Task tool in the same message (parallel execution). Use the general-purpose agent type.

Each sub-agent prompt is built as:

<role prompt from prompts/02-{node}.md>

USER QUESTION:
<original question>

ANSWER_MODE: <one of: analytical | personal_decision | framework | meta>

USER SNAPSHOT (from router):
<user_snapshot>

SEARCH HINT (from router):
<search_hints[node]>     // omit if node doesn't search

OUTPUT FORMAT:
<for Macro/Local/Historical: existing 400-word format>
<for Personal: behavior depends on answer_mode — see prompts/03-personal.md>

Search-enabled nodes (macro, local, historical) must cite at least one URL. Personal node must cite at least one specific element of persona.md/memory.md it relied on (and in analytical mode, output drops to ~250 words and excludes any topic-redirection).

Step 3 — Synthesizer

After all sub-agents return, read prompts/06-synthesizer.md and apply it yourself. The synthesizer input is:

• Original question
• answer_mode from router (controls structure and forbidden patterns)
• user_snapshot from router
• skip_log from router
• All node outputs (verbatim)

The synthesizer produces the final user-facing answer. Behavior depends on answer_mode:

• analytical: deliver the analysis itself. Do NOT write a writing brief, article title, or "you should engage with this topic" meta-commentary. Personal lens contributes voice only.
• personal_decision: lead with what user actually cares about for their own decision; surface tradeoffs; end with concrete next step.
• framework: lead with the framework; lenses fill it.
• meta: minimal direct answer; skip lens scaffolding.

In all modes: surface real disagreements between lenses (don't paper them over); mention skip_log briefly at end; use the user's preferred language (Traditional Chinese unless the question was in English).

Step 4 — Memory Updater

After delivering the final answer, read prompts/07-memory-updater.md and apply it to produce a patch file at <PATCHES_DIR>/YYYYMMDD-HHMMSS.md. Do NOT modify memory.md directly — the user confirms next session.

• If PATCHES_WRITABLE_AT_DATA_DIR == true: silently write the patch and tell user: "已寫入候選 memory patch，下次開啟會請你確認。"
• If patches had to fall through to outputs (skill install dir is read-only): write to <outputs>/patches/ and tell user: "已將候選 memory patch 寫到 outputs/patches/YYYYMMDD-HHMMSS.md。因 skill 安裝目錄不可寫，請手動移到你的資料夾中、或下次跑時連接一個可寫資料夾。"

If the session produced nothing memory-worthy, write a "NO PATCHES PROPOSED" file (still useful as audit trail) — but don't surface it to the user beyond one quiet line.

Failure modes & guardrails

• Router output malformed: fall back to running all four nodes with empty search_hints; default answer_mode to analytical (the safer fallback — won't accidentally personalize the answer).
• Router puts a writer/analyst into personal_decision mode for an external question: this is the known failure mode (see Bug fix v1.1). The question verb determines answer_mode, not the user's profession.
• WebSearch fails / returns empty: that node should explicitly say "搜尋失敗" rather than hallucinate; synthesizer should down-weight it. If macro or local fails twice, try anysearch via scripts/anysearch_cli.py search "<query>".
• Sub-agent times out: skip it in synthesis and note in the skip_log.
• persona.md missing: tell the user this skill needs persona.md to work and stop.
• memory.md missing: create an empty one with the template header and continue.
• PATCHES_DIR not writable: write to outputs/patches/ and notify the user explicitly.

Files in this skill

• SKILL.md — this file
• prompts/01-router.md
• prompts/02-macro.md
• prompts/03-personal.md
• prompts/04-local.md
• prompts/05-historical.md
• prompts/06-synthesizer.md
• prompts/07-memory-updater.md
• templates/persona.md — copy to skill root on first install
• templates/memory.md — copy to skill root on first install
• patches/ — generated patch files awaiting user confirmation
• README.md — install & setup