Research is complete when it becomes useful understanding. The aim is not to collect the most sources or produce the longest report, but to explain a worthwhile subject clearly, with enough evidence and depth to trust it.
This workflow applies to AI technology, products and markets, industry landscapes, and human capabilities. Use it with the Content standards. For a model study, also use the Model Deep Dive guide.
Define what the reader needs
Begin with a question, topic, field, paper collection or practical decision. Identify the reader, purpose and scope before broad research. Decide what understanding would make the work useful, and which questions are outside it.
Look for related explanations first. Extend an existing page or connect it to a focused companion when that is clearer than adding another isolated entry. Organize by the subject’s natural domain, not by a fixed technical template.
Build and challenge the evidence
Map the main dimensions and the evidence that could resolve them. Include a credible alternative explanation or skeptical direction. Inspect actual sources and preserve the conditions around their findings.
Match the evidence to the claim. A technical mechanism, a company’s current product, a market estimate and a firsthand account call for different sources. Keep published findings, calculations, interpretation and experience distinct.
Review consequential claims separately from drafting. Check source identity, scope, freshness, units and comparisons. Seek counterevidence instead of building an article from supporting examples alone. Keep missing information visible when it could change the conclusion.
Design the explanation
Give the reader a clear opening: what the subject is, why it matters, and what this page helps them understand. Start with intuition before introducing specialized vocabulary. Where an answer is established, state it early. Where the purpose is exploration, do not manufacture a tidy conclusion.
Let the structure fit the work. A mechanism may need a computation and its tradeoffs. A landscape may need comparisons. A study of communication may need a concrete situation, research findings and the limits of applying them. Not every page needs the same headings, numbers or figures.
Make important detail available without making the main path exhausting. Use focused sections, useful links and optional depth. Keep essential caveats beside the claims they qualify, even when secondary detail is folded away.
Give visuals a job
Choose a figure when it makes a mechanism, relationship, comparison or process easier to understand. Prefer a useful source figure with a documented reuse basis. Create our own explanation when an important view is missing or a focused comparison would be clearer.
Use the smallest visual that teaches the idea. Preserve meaningful labels, short numbered captions, attribution and text equivalents. Avoid decorative complexity, redundant illustrations and unreadable overview posters. Follow the media standards.
Keep one canonical explanation
For site work, the Markdown or MDX page is the canonical narrative. The site generates its HTML and text exports from that source. Research records support the article; they are not a second article to maintain.
Keep sources, claims, source versions, visual provenance and review results in
docs/research/<domain>/<topic>/. Preserve existing reports as research history.
A standalone report can be the canonical output when that is the requested
deliverable. Keep every repository record suitable for public sharing.
Review, refine and verify
Review evidence and presentation in separate passes, then read the page as a whole. Check the actual rendering, not only the source or a passing build.
- Can the opening orient a reader without prior knowledge of the topic?
- Does the explanation flow from motivation to useful understanding?
- Are important assumptions, disagreements and evidence gaps visible?
- Does each figure teach something and remain readable on mobile?
- Do captions, citations, equations and comparisons remain clear?
- Can repetition or clutter be removed without losing necessary meaning?
- Do text exports preserve the substance without requiring a visual or media player?
Fix concrete problems and inspect the affected output again. Record checks that actually occurred. An unavailable review remains a gap, not a passed check. Follow the validation checklist.
The repository skill
The project includes a versioned latent-light-deep-research skill under
.agents/skills/latent-light-deep-research/. It coordinates the work described
here and supplies research record schemas, initialization and validation.
Its instructions, references, scripts and tests are managed in this repository.
The About source pages remain authoritative for editorial and development standards. The skill does not authorize publication, external actions or changes outside the requested scope. Its technical files are not published as website pages.