Agent deep research trigger
Portable /deep research trigger for Claude Code and OpenAI Codex. Use only when the user explicitly types /deep to start one bounded, host-led research session.From its SKILL.md
npx -y skills add jechiu16/agent-deep-research-triggerAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
2.9 KB, 691 tokens by cl100k_base, as published. Nobody here has run it
問題:{正規化後的問題} Query Brief:{決策、範圍、成功條件各一句} 建議:{light/standard/heavy},因為{一個理由} Light:deep {a}|search {b}|free unlimited Standard:deep {a}|search {b}|free unlimited Heavy:deep {a}|search {b}|free unlimited D1:{最低成本 ready provider;候選與資料外送範圍} 共通:背景執行;host 複驗並寫結論;交付 JSON + 繁體中文 HTML;超限即停並標註缺口 開始:light|standard|heavy|調整|取消
<!-- PURE_TRIGGER_CARD_END -->/deep
SKILL.md is the sole public protocol shared by Claude Code and Codex. On
explicit /deep, normalize the question from conversation context, run only
the local deep-research-state card command, and show exactly one completed
card with no prose around it. If the question is missing, set the recommendation
to 調整; do not start research.
Before Confirmation
Do not search, inspect the project, call a provider, or start a worker. Local
profile/registry reads for the card are allowed and make no external request.
The user's light, standard, or heavy reply confirms that printed count
vector and disclosed provider/egress set for one run. 調整 and 取消 spend
nothing. Re-card only when the vector, provider set, or egress scope changes.
After Confirmation
Read HARNESS.md, create the canonical package, and run in the background. The selected host is the Organizer and sole conclusion author. Provider reports, including D1/D2, buy breadth and structure only; they cannot support a canonical claim.
Use the cheapest ready D1 provider unless source fit or privacy gives a named
reason to choose another disclosed candidate. After each provider report, feed
its useful hypotheses, contradictions, citations, and prior session context to
the host. The host chooses targeted re-verification, consuming search; free
routes remain unlimited within physical/time safety bounds. Fix disproved
claims, mark unverifiable claims, and never withhold delivery for uncertainty.
Stop external calls at the confirmed count limit. Finish from existing materials and name the unresolved gap. Heavy may use a second deep call only when the host expects a material new angle, challenge, or expansion; no hard gate and no automatic provider bundle.
Delivery
Always deliver canonical state.json, events.jsonl, raw/, and
report.html. Human-facing narrative fields and HTML are Traditional Chinese;
exact excerpts, titles, URLs, IDs, hashes, and diagnostics stay unchanged.
Integrity failure remains unsafe; evidence insufficiency remains an annotated
epistemic status. Neither prevents producing the package.
What ships with it: 199 files
2312.0 KB alongside SKILL.md, 68 of them executable
docs/
- benchmarks/2026-07-11-exa-vs-brave.md1.1 KB
- superpowers/plans/2026-07-10-v2-runtime-foundation.md79.9 KB
- superpowers/plans/2026-07-11-exa-search-adapter.md4.4 KB
- superpowers/plans/2026-07-11-openalex-adapter-completion.md4.8 KB
- superpowers/plans/2026-07-12-beta5-productization.md11.8 KB
- superpowers/plans/2026-07-12-pure-trigger-tier-contract.md3.1 KB
- superpowers/plans/2026-07-12-zh-tw-human-report.md12.8 KB
- superpowers/session/2026-07-10-v2-decisions.md19.8 KB
- superpowers/specs/2026-07-10-adaptive-scientific-research-harness-design.md79.8 KB
- superpowers/specs/2026-07-10-provider-portfolio-design.md20.4 KB
- superpowers/specs/2026-07-11-async-deep-engine-boundary.md5.4 KB
- superpowers/specs/2026-07-12-beta5-productization-design.md7.0 KB
- superpowers/specs/2026-07-12-pure-trigger-tier-contract.md4.3 KB
- superpowers/specs/2026-07-12-zh-tw-human-report-design.md5.4 KB
examples/
- eval/annotations.example.jsonl1.3 KB
- eval/questions.jsonl2.7 KB
- eval/README.md3.6 KB
- field/01-duckdb-source-of-truth/confirmation-card.txt712 B
- field/01-duckdb-source-of-truth/session/events.jsonl4.1 KB
- field/01-duckdb-source-of-truth/session/raw/A1.txt165 B
- field/01-duckdb-source-of-truth/session/report.html10.9 KB
- field/01-duckdb-source-of-truth/session/state.json15.7 KB
- field/02-regime-staleness-boundary/confirmation-card.txt755 B
- field/02-regime-staleness-boundary/session/events.jsonl4.6 KB
- field/02-regime-staleness-boundary/session/raw/A1.txt337 B
- field/02-regime-staleness-boundary/session/raw/A2.txt549 B
- field/02-regime-staleness-boundary/session/report.html14.1 KB
- field/02-regime-staleness-boundary/session/state.json19.0 KB
- field/03-long-job-acceptance-test/confirmation-card.txt732 B
- field/03-long-job-acceptance-test/session/events.jsonl4.6 KB
- field/03-long-job-acceptance-test/session/raw/A1.txt557 B
- field/03-long-job-acceptance-test/session/raw/A2.txt839 B
- field/03-long-job-acceptance-test/session/report.html14.7 KB
- field/03-long-job-acceptance-test/session/state.json19.4 KB
- field/04-duckdb-concurrency-boundary/baseline.md2.1 KB
- field/04-duckdb-concurrency-boundary/confirmation-card.txt894 B
- AGENTS.md898 B
- CHANGELOG.md11.2 KB
- CONTRIBUTING.md1.4 KB
- .env.example826 B
159 more files not listed here. See all 199 in the repository.