Agent Skills
› mohitagw15856/pm-claude-skills
› desk-research-sprint
desk-research-sprint
GitHub一种限时桌面研究技能，通过分解问题、规划来源、实时捕获和强制停止规则，在有限时间内产出结构化答案与置信度评估，避免无效浏览。
Trigger Scenarios
需要快速调研特定市场、工具或主题
时间紧迫需高效获取知识
希望将研究转化为明确结论而非零散信息
Install
npx skills add mohitagw15856/pm-claude-skills --skill desk-research-sprint -g -y
SKILL.md
Frontmatter
{
"name": "desk-research-sprint",
"homepage": "https:\/\/mohitagw15856.github.io\/pm-claude-skills\/skill\/desk-research-sprint.html",
"metadata": {
"openclaw": {
"emoji": "🧠"
}
},
"description": "Run a timeboxed desk-research sprint that ends with an answer instead of forty tabs — the question decomposition, the source plan by question type, the capture discipline that prevents re-reading, and the stop rule that beats completionism. Use when asked research this market\/tool\/topic by Friday, I have two hours to get smart on X, structure my desk research, or I keep researching and never concluding. Produces the decomposed questions, the source plan, the capture format, and the timeboxed synthesis with confidence labels."
}
Desk Research Sprint Skill
Unstructured research expands to fill all available time and ends with tabs instead of answers — because "research X" was never converted into questions that can be done. The sprint fixes the shape: decompose into 3–5 answerable questions (each with what-good-enough-looks-like), plan sources by question type (market numbers, user sentiment, and technical claims live in different places), capture findings in one running doc at reading time (re-reading is the silent time-thief), and obey the stop rule — the timebox ends, the synthesis gets written from whatever's captured, gaps labeled honestly.
What This Skill Produces
- The question set — the vague topic decomposed into 3–5 answerable questions with good-enough bars
- The source plan — per question: where answers of that type actually live, and the source-triangulation depth it deserves
- The capture doc — one running format: finding → source → confidence → which question it feeds
- The synthesis — the answers at their earned confidence, the gaps named, the next-sprint questions if any
Required Inputs
Ask for these if not provided:
- The real question behind the topic — "research the CRM market" hides "which three CRMs should we demo?" — the decision the research feeds defines done (what-to-ask energy, applied to research)
- The timebox — two hours and two days are different sprints; the question count and depth budget follow
- What's already known — prior research, existing beliefs to test (stated as hypotheses, so confirmation bias gets a fence)
- The output's destination — a recommendation memo? A brief for the boss? The synthesis writes toward its reader from the start
Framework: The Sprint Rules
- Decompose to answerable: each question passes two tests — could a finding settle it? and what does good-enough look like? ("rough market size ±50% is fine" vs. "need the actual pricing tiers"). Questions without a good-enough bar recruit completionism; the bar is the permission to stop.
- Sources by question type: market numbers → industry reports, filings, the triangulation discipline · user sentiment → review sites, forums, communities (read for patterns, not anecdotes) · technical claims → docs and changelogs over marketing pages · pricing → the vendor's page plus the forum thread about what it actually costs. Typed source plans kill the generic-search spiral.
- Capture at reading time, once: every useful finding goes into the running doc as it's read — one line: the finding, the link, the confidence flag, the question it feeds. The alternative (read now, harvest later) reads everything twice and harvests half; the capture doc is also the synthesis's raw material, pre-sorted.
- The timebox allocates, the stop rule enforces: budget across questions up front (the decision-critical ones get double), and when a question's good-enough bar is met — stop researching it, even mid-interesting-article. At timebox end, synthesis happens with what exists; "one more source" is the lie completionism tells.
- Synthesize with confidence labels: each question answered at its earned level ("Q2: roughly $4–6B, single-sourced, fine for our purpose · Q4: couldn't verify — flagging as the open risk") — the labeled gap is a finding, and pretending coverage is the sprint's cardinal sin. The last section: what a second sprint would chase, if the decision warrants one.
Output Format
Research Sprint: [topic] → [the decision it feeds] · timebox: [T]
The Questions
| # | Question | Good-enough bar | Time budget |
|---|
Source Plan
[Per question: the typed sources + triangulation depth]
The Capture Doc (running)
[Finding · source · confidence · feeds-Q# — one line each, written at read-time]
Synthesis
[Per question: the answer at earned confidence · the labeled gaps · the recommendation if the destination wants one · next-sprint questions]
Quality Checks
- Every question has a good-enough bar set before searching
- Sources were planned by question type, not generic-searched
- Findings were captured at read-time into the one doc
- Questions stopped at their bars; the timebox ended the sprint
- Gaps are labeled as findings, never papered over
Anti-Patterns
- Do not research a topic — decompose to questions or inherit forty tabs
- Do not read without capturing — the second read is the sprint's biggest hidden cost
- Do not keep researching past the bar — good-enough was defined for exactly this moment
- Do not present echoed sources as confirmation — the triangulation rules ride along
- Do not end without the synthesis — captured-but-unsynthesized research is tabs with better formatting
Version History
- 471c606 Current 2026-07-22 13:21
