# Source Research > Source Research - Skill: `handsomeboy990/source-research` (Agent Skill, multi-file: 4 files) - Install (CLI): `npx skillmds@latest add handsomeboy990/source-research` - Raw SKILL.md: https://api.skillmd.com/api/skills/handsomeboy990/source-research/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Research & Search - Author: Handsomeboy990 (https://skillmd.com/u/handsomeboy990) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/handsomeboy990/source-research --- # Source Research Finding sources and reading them are different acts, and the second is where research is actually done. A list of search results is not research; a set of sources opened, read, and attributed is. This skill does the gathering, to the standard `research-core` sets. ## 1. Build the search strategy before searching ``` terms the vocabulary the answer uses, including the terms of art a domain uses that a layperson would not, and the synonyms a source might sources where the top-of-ladder answer lives: which registry, which journal, which official site, which primary corpus, not just a web search angles the question from more than one direction, so a single phrasing does not select a single viewpoint stop rule what saturation looks like: when new searches return sources already seen, the search is done, not when patience runs out ``` ## 2. Work down to primaries A web search lands on aggregators. The work is following them to the source. ``` land on rung 3 or 4: the article, the summary, the answer follow its citations to rung 1 or 2: the paper it cites, the filing it quotes, the documentation it paraphrases read the primary, not the paraphrase; the paraphrase is often wrong in the detail that matters record the primary as the source, noting the route that found it ``` A finding whose only source is an aggregator, with no primary behind it, is a finding to flag as unverified, not to report as established. ## 3. Actually retrieve and read each source The consulted rule from `research-core` is operational here. ``` retrieve open the source; a search snippet is not the source read the relevant part in full; a title and a snippet mislead locate record where in the source the finding is: the section, the page, the paragraph, the line, so it can be checked quote the exact words when the wording matters; paraphrase accurately when it does not, and never blur the two ``` ## 4. Extract findings with attribution ``` finding the specific thing the source establishes, in the researcher's words attribution the source, its rung, its date, and the exact location within it type fact, inference, or opinion, per research-core confidence how firmly the source supports the finding: directly, by implication, or as one interpretation ``` ## 5. Track coverage ``` map which parts of the question each source addresses, so gaps are visible gap a part of the question no consulted source answers is named, not hidden by the parts that are well covered saturation the question's parts are each covered by more than one independent source where they matter, or the gap is stated bias the sources consulted skew toward one viewpoint, one geography, one vendor: note it, and seek the missing angle ``` ## 6. Prohibitions - Never report a search result as a source without opening and reading it. - Never cite an aggregator's paraphrase when the primary was reachable and not followed. - Never record a finding without its exact location in the source. - Never blur an exact quote and a paraphrase. - Never stop at the first source that confirms the desired answer. - Never present a well-covered part of a question as if it covered the gaps. - Never let the set of sources skew to one viewpoint without noting it. ## 7. Protocol 1. Take the framed question from `research-core`. 2. Build the search strategy: terms, target sources, angles, stop rule. 3. Search, landing on whatever rung, then follow citations down to primaries. 4. Retrieve and read each source in full where it is relevant. 5. Extract findings with exact attribution, location, and type. 6. Map coverage against the question's parts; name the gaps. 7. Continue until saturation or a stated gap; stop then. 8. Hand the consulted sources and attributed findings to `synthesis-reporting`. ## 8. Auto-critique Score from 0 to 5: strategy built before searching, aggregators followed to primaries, every source actually retrieved and read, findings attributed with exact location, quote and paraphrase kept distinct, coverage mapped and gaps named, source set checked for skew, search stopped at saturation not fatigue. Threshold: no axis below 3, average at least 4. A reported source that was not opened, or a finding without a locatable attribution, caps the score until fixed. ## 9. Interfaces - Upstream: `research-core` frames the question and sets the evidence standard. - Downstream: `source-verification` checks the findings that matter most, `synthesis-reporting` turns the attributed findings into an answer, `competitive-analysis` and the career and business skills consume the gathering. - Lateral: `report-writing` renders the result for a decision. ## 10. Live sources This skill reads whatever sources the runtime can reach: web search and retrieval, document stores, code repositories, official registries. It never substitutes its own prior belief for a source it could have retrieved. When a needed source is unreachable (paywalled, offline, out of scope, or requiring a connector that is not configured), it names the source and the obstacle rather than inventing what the source would have said.