/research

name: research
description: Explore social-platform topics and discover people, projects and adjacent leads; retain a broad map, selective depth and durable follow-ups.

Research

Use for sourcing, people/project discovery and open-ended landscape research. Simple facts, supplied-post summaries and single-tool requests usually do not need this skill. User instructions determine scope.

The runtime supplies short research operating principles on every request, including after compaction and during final/local delivery. This document is the strategy/tool reference; reload only when its details help, not every round.

Plan from the user's goal

Initialize research_update with objective, a few practical criteria and normally focus=explore. Use focus=verify only for explicit fact-checking/due diligence. Infer a useful plan from sparse requests and state assumptions briefly. Do not invent geography, check size, fundraising stage or a hard time cutoff.

The selected workflow supplies domain methods and deliverable expectations. Do not add investment criteria, a candidate roster or a direction map to tasks that do not call for them. Loading this skill does not increase the execution budget. Light mode makes a useful first pass; deep mode allows more complementary exploration and selective depth. Neither requires filling a quota.

Reserve time for delivery. Keep a short notebook of productive paths, new branches and next actions. Park questions that add little; revisit with a new clue. An unknown fact is not automatically unfinished discovery.

Tools and source coverage

• Search with one or two core concepts, an intact name or a short question.

Expand through adjacent concepts for breadth and observed entities/details for depth. Synonym swaps returning the same posts do not add coverage. Avoid stacking all criteria into one query or spending most searches on definitions.

• On XHS, preview to select unfamiliar useful posts, then read details/comments. Roundups

and image posts can contain many names. Use read_saved_notes(note_ids=[...]) and text pagination for saved body/OCR/comments that were truncated. For URLs only, section="links" returns a compact ID/title/URL index; avoid shell-parsing JSON or rereading all text. Omit note_ids to list saved sources.

• When explore_platforms is exposed, use it early for independent multi-platform

discovery. Assign complementary questions per platform with goal, known evidence and gaps. Merge findings before another focused round; avoid asking every worker to repeat broad searches. Workflow defaults guide automatic sources; explicit selections take precedence. Keep partial findings and distinguish unavailable platforms from searched platforms with no useful leads.

• When deep-mode instructions enable the implicit public-web worker, include a

`web` branch in the first explore_platforms call alongside selected platforms, even for single-platform research. Give it complementary questions and explicit user source restrictions. It searches DuckDuckGo in its own temporary tab and reads useful original pages; merge its sourced findings with platform evidence. Do not treat search snippets as full-page reads. Explicit user exclusions take precedence, and web access failures should not block useful platform findings.

• On XHS, use explore(branches=[{question: "..."}, ...]) for autonomous investigation of

concepts, people, profiles or posts. Up to three independent branches use separate tabs. Supply the goal, known context, observed IDs and tried queries; workers do not inherit your whole conversation. Default 15 detail posts per branch is adjustable with max_posts. One branch can deepen a single candidate; direct author_scan/get_notes also work. Prefer useful fresh profile posts over reopening known material. Avoid overlapping assignments and merge secondary findings as well as headline names before choosing the next directions.

• Alternate breadth with selective author/history depth. Stay within the selected source scope. In automatic mode choose useful available

platforms; in explicit mode do not collect from unselected platforms. Cross-platform profiles, articles, code and interviews can reveal new signals as well as enrich a lead; they are not mandatory verification for every candidate. Use each platform's actual tools; XHS explore cannot run other platforms. Do not guess many domains or repeatedly repair optional external sources.

• Preserve useful self-reports, tentative connections and secondhand/preview

leads with brief labels. A curator is distinct from a company they mention; several posts about one company are not several deals. Exclude irrelevance and duplicates, not useful unnamed builders. Do not invent visits or facts.

Incidental comment feedback in deep mode

While reading each post on platforms with comment sections (including XHS, X and Douyin), skim the first few available comments or replies, typically 3–5. Reuse returned/saved comments; otherwise read a small first batch with the platform's actual tools. Respect explicit user limits. Do not routinely paginate, expand reply chains or launch separate sentiment searches. If comments are unavailable, note the coverage limit and continue the main research.

Keep relevant praise, complaints, disagreements, unmet needs and repeated themes as supplementary evidence. Distinguish commenter views from author claims and link the original post plus an observed comment/reply URL when available. These small, platform-ordered samples are anecdotal, not representative sentiment; do not infer percentages or trends from isolated opinions. Fold useful signals into the synthesis, optionally in a brief comment-feedback section. Do not force a section without useful evidence. Pass this scope explicitly in XHS explore branch questions and retain comment findings from all workers when synthesizing. If the user explicitly requests reputatio