Conciseness

Lean throughout — no explanations of concepts Claude already knows, no padding; directives like 'Research is parallel; writing is single-point — never let multiple agents write report sections' and 'raw page content must never enter your context' earn every token. Matches anchor 5; a 4 would require trimmable over-explanation and none is evident.

Actionability

Concrete elements exist (run `date +%Y-%m-%d` via Bash, the workspace tree, the depth/budget table, the findings schema 'claim / quote / URL / date / confidence per item'), but the load-bearing artifacts — 'the locked template in [reference/subagent-prompt.md]', 'the free endpoints in [reference/sources.md]', and '[reference/report.md]' — are referenced yet absent from the bundle, so 'reproduce it verbatim' cannot actually be executed. Not 4 because these are key details, not minor gaps; not 2 because the orchestration steps that are given are specific and executable.

Workflow Clarity

Clear Step 0 → Phase 5 sequence with explicit validation checkpoints and feedback loops: triage ('Answerable with 1-2 searches? → STOP'), enforced budgets ('These are hard budgets... never exceed them'), the Phase 4 gap check with conditional delta-query rounds, failure handling ('If a sub-agent fails or returns thin results, note it and move on'), and a critique pass for deep mode. Matches anchor 5.

Progressive Disclosure

Section structure is clean and references are clearly signaled ('[reference/subagent-prompt.md](reference/subagent-prompt.md)', '[reference/sources.md](reference/sources.md)', '[reference/report.md](reference/report.md)'), but none of the three referenced files exist in the bundle — there is no reference(s) directory at all — so the one-level-deep disclosure chain is broken. Not 4 because the gap is material (every pointer dangles); not 2 because the body itself is well organized and nothing is inlined that belongs in a separate file.

Specificity

Names the mechanism concretely — 'using parallel sub-agents, the available search tool (websearch or a search MCP server), WebFetch and free APIs' plus the output 'a cited report' — which is several specific actions with only minor gaps (the deliverable itself is sketched only as 'cited report'). Not 5 because the enumeration covers tools/mechanisms more than the full range of concrete capabilities; not 3 because it clearly exceeds 1-2 actions.

Completeness

Explicitly answers both what ('Deep research on any topic using parallel sub-agents... WebFetch and free APIs' with 'a cited report') and when ('Use when the user asks for a thorough multi-source investigation with a cited report') with concrete trigger phrases, plus explicit exclusions. Matches the anchor-5 example structure; a 4 would require the 'when' clause to be less explicit, which it is not.

Trigger Term Quality

Includes multiple natural trigger phrases users would actually say, in both Chinese and English — "深度调研X"、"deep research"、"帮我全面研究一下"、"多方求证"、"写一份调研报告" — but common English variants like 'research report', 'deep dive', or 'investigate X' are missing. Not 5 due to those missing synonyms; not 3 because coverage clearly exceeds 'some relevant keywords'.

Distinctiveness Conflict Risk

Clear niche (thorough multi-source investigation with a cited report) with explicit negative boundaries — 'NOT for simple lookups (a single search suffices)' and 'NOT for academic literature surveys (use auto-research skill instead)' — actively disambiguating against sibling skills. Minimal conflict risk; matches anchor 5.