Evidence-grounded desk research — portable across every repo.
After installing research, every question your agent takes on is grounded before it answers. `/desk-research` runs scoping, source curation, and synthesis in one session across four depth modes. For sustained investigations, the four `desk-research-project-*` skills run a multi-week lifecycle that accumulates a corpus and ends in a brief you can hand to a decision.
desk-research1 gateThe primary desk-research skill. Runs scoping, source curation, and synthesis in a single session, selecting depth from four modes — shallow through exhaustive.
source-mapMaps the canonical sources for a topic before evidence retrieval begins — prevents wasting citations on secondary sources that restate a primary.
build-outline1 gateBuilds a research outline from the source map, structuring the question before the agent fetches anything.
identify-perspectivesIdentifies the stakeholder perspectives and intellectual traditions bearing on a question before synthesis.
compare-hypothesesRuns the competing-hypotheses pipeline against a set of candidate answers, producing a scored matrix.
devils-advocateSteelmans the opposing case for a position — used after compare-hypotheses to stress-test the top candidate.
decision-archaeologyReconstructs why a prior decision was made from artifacts, commit history, and design docs — used when the answer is historical rather than open.
desk-research-project-start1 gateInitializes a research project folder with a scoped question, source list, and corpus skeleton.
desk-research-project-statusOrients to the current desk-research project at a glance — reads overview.md and surfaces phase, working hypothesis, stop-signal verdict, and what to do next.
desk-research-project-checkSnapshots progress: which sources are captured, what the corpus covers, and what remains.
desk-research-project-digestSummarizes the accumulated corpus into a digest artifact — the input to the final synthesis.
desk-research-project-synthesize1 gateSynthesizes the corpus digest into a brief graded by confidence, ready to hand to a decision.
You give the agent a question and select a depth mode. For a single-session query, the agent activates /desk-research and runs a scoping pass — identifying what kind of question this is (factual, comparative, historical, open-ended) and which sources to consult. For a project-mode investigation, it runs desk-research-project-start to create a corpus folder and an initial source list.
You: Read the scope statement the agent produces before it begins fetching. A bad scope leads to a confident answer to the wrong question — the cheapest fix is here, not after the synthesis returns. If the agent’s framing misses the real question, redirect with one sentence before the retrieval subagents run.
The agent runs source-map or its equivalent, identifying the canonical sources for the question domain. Two retrieval subagents — evidence-retriever and source-extractor — fetch and synthesize source material without polluting the main session context.
You: Watch the source list take shape. If a key source is missing — a specific industry report, a primary author’s original paper, an internal standard you know exists — name it explicitly. The agent doesn’t know what you know about your domain. A one-sentence nudge here is faster than a post-synthesis correction.
After sources are captured, the agent synthesizes a brief graded by confidence (GRADE A–D). Each claim cites its source. Unsupported claims are marked explicitly as gaps, not silently omitted.
You: Review the synthesis at the G-synthesis gate. The confidence grades are the first thing to read — a GRADE-C synthesis needs a different follow-on action (narrow the question, run another retrieval pass, seek a domain expert) than a GRADE-A synthesis (act on it). If a claim lacks a source citation, check whether it’s an inference the agent labeled correctly or an assertion it presented as fact.
For each gate, everything you need to make a confident decision.
A specific, answerable question with a chosen depth mode and a clear success criterion — something a colleague could pick up and continue.
A question that can't be falsified or finished — 'what is the best approach to X?' with no scope boundary or success criterion.
The scope gate sets the direction for everything that follows. A bad scope means the agent searches the wrong space and returns a synthesis that looks complete but answers the wrong question.
A synthesis that names its sources, grades its confidence honestly, and directly answers the scoped question — including an explicit gap map where confidence is low.
A synthesis that sounds authoritative but can't be traced to primary sources. Or one that confidently answers a different question than the one you asked.
The synthesis is the final output. A confident but wrong synthesis is actively harmful. If the confidence grade is low, the right response is to narrow the question or run another retrieval pass — not to ship the draft.
agentbundle install --pack desk-research --scope user