git:20260621.90502c4 to git:20261031.9616b02

15 added, 1 removed. Audit A to A.

---

name: canvas-deep-research

- description: Source-backed research workflow for Pulse Canvas bootstrapping, due diligence, technical/product/market landscape analysis, current-state research, and any canvas task where Codex must gather reliable evidence before producing an artifact. Use when canvas-bootstrap needs a bundled research protocol, or when the user asks to research reliable sources before creating or organizing a canvas.

+ description: Source-backed research workflow for Pulse Canvas bootstrapping, due diligence, technical/product/market landscape analysis, current-state research, and any canvas task where the agent must gather reliable evidence before producing an artifact. Use when canvas-bootstrap needs a bundled research protocol, or when the user asks to research reliable sources before creating or organizing a canvas.

---

# Canvas Deep Research

Produce evidence-backed research briefs with explicit scope, source quality, uncertainty, and a reusable handoff for downstream artifacts such as reports, plans, or canvas workspaces.

## Operating Modes

- If the user directly requests research and has not provided scope or depth, ask for the missing scope and a depth choice before searching.

- If a parent skill supplies an approved scope, depth, and plan, proceed without re-asking the user.

- Use the user's language for questions and final synthesis unless a source quote or artifact format requires otherwise.

- Browse the web for current, unstable, niche, legal, medical, financial, market, product, standards, API, or vendor facts.

- Prefer primary and official sources for facts about products, APIs, policies, specs, prices, regulations, organizations, and people.

- Do not fabricate sources, dates, quotes, statistics, or consensus. Label inference as inference.

- This skill does not mutate canvas nodes or files by itself. Return structured research that another skill can use.

## Depth Levels

- `quick`: 1-2 research angles, a small set of high-signal sources, concise findings, and clear caveats. Use for orientation or low-stakes decisions.

- `standard`: 3-5 research angles, primary plus credible secondary sources, conflict checks, source ledger, and an actionable synthesis. Use as the default for canvas bootstrap work.

- `deep`: staged passes across primary sources, expert or institutional sources, recent developments, counterarguments, contradictions, and open questions. Use for high-stakes or strategy-shaping work.

When the user asks for "high quality" or "deep research" without a depth, default to `deep` only after confirming the time/cost tradeoff if the work is user-facing. Parent skills may choose `deep` from an approved plan.

## Workflow

### 1. Define the Research Contract

Capture:

- topic and decision context

- intended audience

- research depth

- date range and geography, if relevant

- must-include and must-exclude areas

- success criteria for a good answer

If user-facing, present the contract briefly and ask for approval before a long research pass unless the user already approved a plan.

### 2. Plan Before Searching

Draft a short research plan:

- core questions to answer

- source strategy by question

- expected source types, such as official docs, standards, datasets, research papers, company filings, reputable media, or local files

- contradiction checks and risk areas

- output shape, including whether a canvas handoff is needed

For `deep`, split the plan into multiple passes so findings from early passes can refine later searches.

### 3. Collect Evidence

For each source used, record:

- stable source id, such as `S1`

- title

- publisher or author

- date published or updated when visible

- URL or local path

- source type

- why it is relevant

Use short quotes only when exact wording matters. Otherwise paraphrase and cite the source id.

+ ### Search Tactics

+ Search quality decides research quality. For each core question:

+ - **Write several queries, not one.** Use 2-4 short keyword queries per question: the plain topic, the official name or product term, and a "vs", "pricing", "benchmark", "limitations", or "criticism" variant. Do not paste the user's sentence as a query.

+ - **Search in the language of the sources.** For a topic asked in Chinese, also search in English when the primary sources are English (most standards, APIs, papers, and global companies), and the reverse for China-specific topics. Write findings in the user's language.

+ - **Reach primary sources directly.** Restrict a search to the official domain when you know it (with Tavily: `includeDomains`), for example the vendor's docs, the standard body, or the regulator.

+ - **Ask for freshness when it matters.** For recent events, prices, releases, or "latest", restrict the period (with Tavily: `topic: "news"` and `timeRange`) and record each source's date.

+ - **Use the deeper mode for research.** With Tavily, use `searchDepth: "advanced"` and 8-10 results for `standard` and `deep` work; `basic` is for quick orientation.

+ - **Read before you cite.** Search excerpts are fragments. Before a source supports a key finding, read the page itself (with Tavily: `tavily_extract`, passing `query` to pull the relevant parts). Never cite a page you only saw as an excerpt.

+ - **Follow the trail.** When a good page cites a primary source, read that source and cite it instead.

+ - **Stop at saturation.** Stop a question when new queries return sources you already have, or when two independent strong sources agree. Note the question as open when they disagree.

+ - **Drop weak sources.** Skip SEO listicles, content farms, undated pages, and pages that only restate other sources, unless nothing better exists; then mark the finding `low` confidence.

### 4. Cross-Check and Synthesize

- Compare claims across source types before treating them as settled.

- Separate "what sources say" from "what this implies".

- Surface conflicts, stale data, weak evidence, and missing information.

- Prefer fewer strong findings over many thin observations.

- Assign confidence per major finding: `high`, `medium`, or `low`.

### 5. Produce the Research Brief

Return a brief with this structure when possible:

```yaml

research_brief:

scope: "What was researched"

depth: "quick | standard | deep"

executive_summary:

- "Most important takeaway"

key_findings:

- claim: "Evidence-backed finding"

support: ["S1", "S2"]

confidence: "high | medium | low"

implication: "Why this matters"

contradictions:

- issue: "Where sources disagree or evidence is weak"

sources: ["S3", "S4"]

source_ledger:

- id: "S1"

title: "Source title"

publisher: "Publisher"

date: "Published or updated date, if available"

url_or_path: "URL or local path"

note: "Why it was used"

open_questions:

- "Question that remains unresolved"

canvas_candidates:

frames:

- title: "Frame title"

purpose: "Why this grouping exists"

nodes:

- title: "Node title"

summary: "Content to place in the node"

source_ids: ["S1"]

edges:

- from: "Frame or node title"

to: "Frame or node title"

label: "Relationship"

kind: "flow | dependency | reference | contrast"

```

Keep `canvas_candidates` concise. It is a handoff, not a final layout.

## Canvas Handoff Rules

When another skill such as `canvas-bootstrap` will create a canvas:

- Organize findings into natural frames, not a fixed template.

- Suggest 3-6 frames for most topics.

- Suggest 2-4 substantial content nodes per frame.

- Include source ids inside each node summary so final canvas content can cite evidence.

- Include only meaningful edges; avoid connecting everything to everything.

- Mark draft, uncertain, or assumption-heavy nodes clearly.