How to Write a Literature Note
BraindumpThe objective is the contract
A literature note answers one stated question. The question is written before any other prose — it is the contract the rest of the note has to honour. A reader scanning the note must be able to tell, in one line, what the note is going to settle and what it is not. Without that contract, the note becomes a dump: there is no test for whether a given paragraph belongs.
The objective also bounds the work. A note finishes when its question is answered, not when the source has been read in full. Material that is interesting but off-topic earns its own note with its own objective; appending it here would dilute the contract this note is going to be audited against. Hereafter, audit refers to the act of reading the note carefully and checking it against each of the rules that follow.
Neutrality is key
Neutrality starts upstream of the prose, in the formulation of the objective itself. A question — “what does the spec require for X?” — admits any answer the source supports. A thesis dressed as a question — “why does the spec require X?” — has already chosen the answer, and every step downstream inherits the choice: what is read carefully, what is quoted, how the quote is paraphrased. The independent reader does not catch this, because they can only check the quotes the author selected, and the author selected to confirm. The objective is therefore phrased as an open question, and any framing that presupposes its own answer is rewritten before the source is opened.
The chain runs from objective to answer
Sections follow the path the reader needs to take from the question at the top to the conclusion at the bottom. Each one moves the argument by exactly one step: it introduces the next concept the answer depends on, or rules out a candidate the objective could have been confused with. A section that does neither is noise, regardless of how interesting it is in its own right.
The order of sections is not the order in which the reading of the source happened. The note is rewritten, after the source is digested, so the chain is the shortest path from objective to answer. If a detour was needed during research and is not needed in the final argument, it is dropped — its absence in the chain is the audit. A reader who has to reconstruct the argument from a meandering note is doing the work the note was supposed to do.
This chain is made explicit as an Argdown map, not left to prose: each statement and each inference is a node wired by a typed relation, and the prose only narrates the map (the node discipline is three kinds, nothing else; the relation vocabulary and tooling are in argdown in org-mode).
The reader’s flow is sacred
The reader should glide through the prose uninterrupted, clicking through to a claim’s source node whenever they want to verify it against the original. Everything else — detours, meta-commentary, secondary justifications — is noise and must be removed. If something is valuable but breaks the flow — a side comparison, a self-contained derivation whose result the argument will consume — move it to its own note and link to it from the main chain at the moment its result is used.
Brevity serves flow. The longer a sentence, the more its message dilutes — attention is narrow, and a convoluted clause already costs a glance back. Where a bulleted breakdown carries the idea more directly than a paragraph, prefer it.
Put a lot of links — to the canonical source URL of a cited text, to the matching map node. If the reader wants to follow a concept referred to and described somewhere else, they must be able to simply click and go there.
A new term enters the document where the document has just laid out what motivates it — not earlier. In a literature note this rule is acute, because the source vocabulary is dense and indexical: article numbers, case identifiers, doctrinal labels, regulatory acronyms. A section title like « 666 plutôt que 653 » breaks flow on the heading itself, because the reader has not yet been told what each one says or why one might apply. The fix is to name the question instead — « quelle présomption s’applique au grillage ? » — and let the identifier arrive in a sentence the body has already prepared. When the structure constrains the sequence, defer the reference: name the general concept the reader already has, and let the specific identifier arrive when its time comes. The rule applies everywhere the reader’s eyes go — section titles, body sentences, node titles, link anchors.
A literature note is a story, not an inventory. Three elements carry rigorous content through:
- stakes — the open question;
- tension — the chain of arguments building toward a result;
- payoff — the answer, or a documented unresolved disagreement.
The reader should feel the writer’s pull toward the answer; substance without narrative rarely lands.
Every semantic claim is an Argdown node
Every claim, every argument, every conclusion is a node in the Argdown map, wired by a typed relation; the prose only narrates the map. There are three kinds of node, and nothing else:
- sourced statement — a verbatim quote of its source: the node carries
+ [source]and reproduces the exact words, the source text living in its own node with the canonical link. Such a quote node earns its place only by atomising — lifting out of a longer source the single proposition the argument uses; when the source node is itself already that lone proposition, a separate quote node merely copies it, and the argument cites the source node directly. - argument — a PCS: premises →
----→ conclusion, the inference holding literally — if the premises, read to the letter, do not entail the conclusion, cut it (« il s’ensuit que » rescues nothing). - conclusion — the output of a PCS, or a premise that is the author’s own reading, explicitly marked unsourced.
Never an interpretation as a sourced node: a reading, gloss, qualification or
inference is a conclusion or a flagged unsourced premise, never a node
carrying + [source] as if the source had said it. One node = one atomic
claim — so a verbatim quote that only restates a source node already pared to
that single claim is not a node of its own but a copy, and collapses into a
direct citation of that source. Anything neither sourced nor derived this way
is a guess and does not belong.
Prose may carry an affirmation the map does not — a framing line, a bridging remark the chain never formally consumes as a node. This is allowed, but it does not escape sourcing: any such out-of-map affirmation takes a footnote holding the verbatim source and a link to the reference, exactly as a source node would. The footnote is the prose counterpart of the source node — same verbatim discipline, same canonical link. An affirmation stated outside the map and without that footnote is a bare guess in the narrator’s voice; it is footnoted or cut, never left standing.
Off-topic is exclusion, not appendix
Material that does not move the chain from objective to answer is cut, even when it is correct, even when it was hard-won during the reading. A literature note is not a record of what was read; it is the distilled answer to one question. A reader who scanned the objective and is looking for the answer must not have to skip past unrelated paragraphs to find it.
Sourced is not load-bearing. A correct, well-cited fact whose removal would not weaken the chain — a foundational article cited for atmosphere, a date that no later sentence consumes, a parenthetical clarification — earns no place by being true. The test is consequence: if the paragraph still follows when the sentence is deleted, the sentence was not chain-bearing and is cut. The same test applies to source nodes; an orphaned source node is the fingerprint of an orphaned claim and goes with it.
The cut material is not lost: it earns its own note, with its own objective, and is linked from any chain where it is actually load- bearing. This is also how the note set grows without each individual note becoming a swamp. A note that tries to hold every adjacent fact ends up holding none of them well, and the link graph between notes — which is where the value compounds — never forms.
Source quality is part of the claim
A source node pointing at the primary source — the specification, the standard, the maintainer’s own documentation, the source code of the system under discussion, or a blog post or forum answer from a well-known and respected actor of the domain — supports a claim differently than one pointing at a forum thread or a blog post by an unknown author. Both are citations; only one is evidence. The note prefers the primary source whenever it exists and is reachable, and names it as such.
Secondary sources are admissible only when they interpret a primary source the note cannot host — paywalled, deprecated, untranslated — and they are named as secondary, with the reason the primary is not cited directly. Anonymous forum posts, undated blog entries, screen- shots without provenance, and AI-generated summaries are not citations: they are unsourced claims wearing the costume of one, and they do not survive the audit. A note built on them collapses the moment the sourcing chain is followed.
The channel matters as much as the source. Each intermediary between the note and the source — a search engine, a reformatter, a third-party mirror, a cache — is a place where wording can drift or the wrong version can be served. A dedicated channel the source itself exposes — an official API, a structured archive addressable by a canonical identifier (Legifrance for French legal texts, an MCP tool for an agent) — is preferred whenever one exists: it returns the source’s own text under its canonical URL, so the source node points at exactly the document the quote came from. Web fetches and search engines are the fallback, with a higher burden of scepticism: the URL is verified, the version date is read, the text is compared against any independent reference before it is trusted as the primary it claims to be.