Technical research

Technical Research

A claim without a source is a guess. Three sources that copy each other are still one source. The most expensive sentence in a technical article is the one that turns out to be wrong.

The research phase is where most technical articles win or lose. Drafts written from a thin research base produce the same telltale prose: hedged claims, vague benchmarks, "studies show" with no study, anecdotes presented as evidence. Drafts written from a strong research base produce the opposite: specific numbers, named sources, claims the reader can check.

This skill covers the pre-writing research workflow: who you're writing for, where the evidence comes from, how to verify it, and how to cite it.

Core principle

Triangulate before you cite. Specific before you generalize. Primary before secondary.

The discipline is asymmetric: it's cheap to verify a fact and expensive to retract one. A single wrong number in paragraph two costs the entire article's credibility. Spend the time upfront.

Five research moves

1. Profile the audience

Before researching the topic, profile the reader. The same article on "incremental adoption of TypeScript strict mode" reads differently for:

• A staff engineer at a 500-person company evaluating the migration
• A solo developer maintaining a 3-year-old side project
• A platform team writing internal guidance

Audience determines what counts as common ground (skip), what needs explanation (cover briefly), and what's the actual contribution (the core).

Profile across three axes:

Axis	Question
Knowledge level	What does the reader already know? What's "obvious" to them and what's "new"?
Job-to-be-done	Why are they reading? Decision support, implementation guide, conceptual model, entertainment?
Prior beliefs	What do they currently believe about this topic — and what does this article challenge or reinforce?

A piece that gets the audience profile wrong is unsalvageable in revision. Get it right before you outline.

See references/audience-profiling.md for templates and worked examples.

2. Tier the sources

Not every source is equal. Use a strict tier hierarchy and prefer the highest tier you can find:

Tier	Source type	Examples
Tier 0 — Primary	The thing itself	Source code, RFCs, specifications, raw data, the API you're benchmarking
Tier 1 — Peer-reviewed	Published research	ACM, IEEE, USENIX, peer-reviewed journals
Tier 2 — Authoritative practitioner	Named expert with reputation at stake	Engineering blogs from companies, named individuals with track records, conference talks with archived video
Tier 3 — Quality popular	Reputable publication, no peer review	NYT/FT/The Atlantic technology coverage, well-known tech publications with editorial standards
Tier 4 — Anonymous popular	No accountability	Reddit comments, anonymous blog posts, "I heard that…"
Tier 5 — Vendor / advocacy	Selling something	Vendor whitepapers, product blogs, sponsored content

Rules:

• Cite the highest tier available. If a Tier 4 blog post and a Tier 1 paper both make the same claim, cite the paper.
• Tier 5 is a last resort. Vendor sources can be used for "X claims Y" but never as the sole source for a load-bearing claim.
• Translate down. When citing a Tier 1 paper, also link to a Tier 2/3 explainer if one exists — the reader can choose their depth.

See references/source-tiering.md for the full hierarchy with examples and edge cases.

3. Triangulate load-bearing claims

A load-bearing claim is one your argument depends on. If it's wrong, the article falls down.

For every load-bearing claim, find three independent sources that confirm it. "Independent" is the hard part:

• Three blog posts that cite the same StackOverflow answer = one source.
• A primary paper, an independent benchmark, and an experiment you ran yourself = three sources.
• The vendor's docs, the vendor's blog, and the vendor's CEO's tweet = one source.

If you can't find three independent sources for a load-bearing claim, either weaken the claim, run your own experiment, or cut it. Don't stretch a single source into a confident-sounding sentence.

See references/triangulation-and-fact-checking.md for the full method, the "independence test," and what to do when sources disagree.

4. Inventory evidence types

A persuasive technical article uses a mix of evidence, not just one type. Build an inventory before you outline:

Evidence type	Best for	Failure mode
Quantitative data	Settling factual disputes; demonstrating scale	Numbers without methodology
Expert quotes	Establishing authority; introducing nuance	Quote-mining; cherry-picking
Demonstrations	Showing rather than telling; reproducible claims	Demos that work in slides only
Case studies	Showing a pattern in context; texture	One-off anecdote presented as pattern
Source code	Settling disputes about what software actually does	Code without commit hash or version
Logical argument	When evidence is unavailable; framing	Argument substituting for evidence
Lived experience	Texture, voice, stake	Ego in the way of insight

The strongest articles balance two or three types. A piece that's all numbers reads cold. A piece that's all anecdote reads thin. A piece that's all logic reads detached.

See references/evidence-types.md for when each type pulls weight and how to combine them.

5. Apply citation discipline

Citations are a contract with the reader: I claim X, here's where you can check.

Rules:

• Link to the most stable URL. Prefer DOI, archived versions, permalinks. Avoid blog posts that get edited.
• Capture link rot up front. Save every cited page to archive.org while researching. The Wayback Machine is the only thing standing between your article and broken links a year from now.
• Inline vs footnote. Inline links for the casual reader, footnotes for the deep reader, both for important claims.
• Quote with context. Never quote a sentence without surrounding context if the surrounding context changes the meaning.
• Date the source. "As of 2024-Q3" or version numbers — software evolves, claims age.
• Cite specifically. "Section 3.2 of RFC 9110" not "the HTTP spec."

See references/citation-discipline.md for citation formats, link-rot mitigation, and worked examples.

✅ Use for

• Researching a technical blog post or deep-dive before drafting
• Building a claim-evidence map for an in-progress article
• Auditing an existing draft's citations and source quality
• Profiling the audience for a piece
• Fact-checking a load-bearing claim
• Choosing between competing sources

❌ NOT for

• Code documentation research — what an API does is documentation-generator's job
• Line-level editing — use long-form-polish (or communication/clarity-editing for short-form)
• The article structure itself — use long-form-structure
• Product personas / UX research — use persona-definition
• Library lookup / API examples — use the context7 docs MCP

Anti-patterns

"Studies show" without a study

What it looks like: "Studies show that developers spend 80% of their time reading code."

Why it's wrong: No study cited; the reader can't verify; the number is folklore. The original Robert Martin claim is anecdotal, the 80% is a guess, and "studies" plural is a fiction.

What to do instead: Either cite the actual study (with link, year, methodology), weaken to "anecdotally" with attribution, or cut the claim. Never use "studies show" as a magic phrase that grants authority.

Single-source confident claims

What it looks like: A 3000-word article whose central thesis rests on one Hacker News comment.

Why it's wrong: Single sources fail. The author of the comment may be wrong; the comment may be misread; the comment may not survive scrutiny. The article's credibility is now hostage to that one link.

What to do instead: Triangulate. If the claim is interesting enough to anchor an article, it's interesting enough to find three independent confirmations.

Vendor-as-authority

What it looks like: "AWS is the most reliable cloud provider, with 99.999% uptime according to AWS's own status page."

Why it's wrong: Vendors grading themselves is not evidence. The status page is a marketing document.

What to do instead: Cite third-party uptime monitors (e.g., StatusGator, ThousandEyes) and academic papers on cloud reliability. Treat vendor sources as "X claims Y," never as ground truth.

Decoration citations

What it looks like: Three citations after a sentence that's already obviously true. Or citations to sources that don't actually support the claim.

Why it's wrong: Citations are a contract with the reader. If the reader clicks and the source doesn't support the sentence, every other citation in the article becomes suspect.

What to do instead: Cite the load-bearing claims. Don't cite the throat-clearing. And read the source you're citing — actually read it.

Profile-by-projection

What it looks like: "I'm writing this for developers" without specifying which developers, what they know, or why they're reading. Or worse, "I'm writing this for myself" — and then publishing it for an audience.

Why it's wrong: A vague audience produces vague writing. Every sentence has to make decisions about what to assume, and an unprofiled audience makes those decisions inconsistently.

What to do instead: Pick a specific reader (real person if possible) and write to them. Update the profile as the piece develops.

Workflow

Use this order for a new long-form piece:

1. Profile the audience. One paragraph: who, knowledge level, job-to-be-done, prior beliefs.
2. State the thesis. One sentence. If you can't, you're not ready to research.
3. List load-bearing claims. What must be true for the thesis to hold?
4. Find sources. For each load-bearing claim, find Tier 0–2 sources. Triangulate.
5. Inventory evidence. Build a map: claim → evidence type → source → link.
6. Identify gaps. What can't you source? Cut the claim, run your own experiment, or weaken to "anecdotally."
7. Archive everything. Save cited pages to archive.org. Capture screenshots of dynamic content. Note version numbers.
8. Hand off to long-form-structure. With the research brief in hand, structure the piece.

Skip steps at your peril. The most common failure mode is researching as you draft — which produces drafts shaped by what's easy to find, not what's true.

References

File	Contents
references/audience-profiling.md	Audience profile templates, worked examples, knowledge-level / JTBD / prior-beliefs frameworks
references/source-tiering.md	Six-tier source hierarchy with examples, edge cases, and "translate down" patterns
references/triangulation-and-fact-checking.md	The triangulation method, the independence test, sources-disagree resolution
references/citation-discipline.md	Citation formats, link-rot mitigation, inline vs footnote, dating sources, archive.org workflow

Related skills

• long-form-structure — once research is done, structure the piece.
• engaging-craft — proven techniques for turning research into prose that holds attention.
• long-form-polish — pacing and scan-ability after the draft exists.
• distribution-craft — title, dek, social pulls after polish.
• persona-definition (skillstack) — for product personas distinct from article-audience profiles.
• storytelling (skillstack) — for narrative arcs distinct from technical exposition.