Format First

format-first

Most content advice optimizes the wrong variable. It picks a topic and varies the shape. This inverts that: pick a format and let the topic vary.

A format is visual grammar (how it is shot, framed, paced) x script structure (where the beats land). It is the thing a viewer recognizes before they process a single word — and it is the thing almost nobody chooses deliberately.

Epistemic stance (read this before quoting anything)

This skill separates these grades, and the separation is the point:

The single most important line: no source in the set searched here describes format-level clustering. That set is Meta's own published material — the Explore engineering writeup, the Reels Chaining system card, about.instagram.com, Mosseri's accounts, the Transparency Center — enumerated in references/evidence-map.md §1.

By this skill's own grades that is unverified, not refuted. Absence from an operator's own description of its system is real evidence and stronger than a failed keyword search, because those pages exist to enumerate ranking inputs. It is not proof: the systems are proprietary and nothing obliges the published descriptions to be complete. The positive mechanism — that consistency improves retrieval — is a working hypothesis with a plausibility argument. Say so, in both directions.

The loop

1. Choose a format, not a topic. Highest-leverage decision, almost never made deliberately.
2. Probe, do not guess. Watch several instances of a candidate format through to the end, then see whether the feed serves you more of that shape. If it does, the shape has a resident audience the recommender already serves. This is a behavioural test — valid regardless of whether the embedding hypothesis holds. (No source located here specifies how many instances; a handful is enough to notice the shift.)
3. Copy the skeleton, supply your own substance. Structure is fair game; content is not — and reposting others' content is the one thing that triggers a real account-level gate (Instagram's originality policy de-recommends primarily non-original accounts; it does not touch followers).
4. Repeat it. The counterintuitive part — and the part resting on the hypothesis above, not on a measured result. The argument is that consistent items are easier for a similarity-based retriever to match to an audience that already engaged. Treat it as the best available guess, and watch your own numbers.
5. Hold more than one format; rotate on measured decay. No source located here supports a specific count or interval — rotate on your own numbers. The wearout literature is laboratory stimulus research and transporting it to platform formats is an analogy, not a finding; what it does support is that liking rises then falls with repetition, and that the curve differs by modality (visual inverted-U; auditory came out U-shaped), so do not port repetition advice from video to audio.
6. Calibrate the hook to existing knowledge. A curiosity gap only fires if the viewer already has enough context to perceive it as a gap. This is why generic hooks fail on cold audiences — structurally, not for craft reasons.
7. Match the signal to the goal. Likes weigh slightly more for connected reach, sends slightly more for unconnected. "Would someone send this?" is the growth question; "was that satisfying?" is the retention question.

Expectations, calibrated

• Expect a heavy tail, not a hit rate. The operator whose method seeded this skill reported most of their own posts underperforming even after a breakthrough. That is one self-reported account, not a measured base rate — do not quote a ratio.
• An audience is not a delivery guarantee. In a study of 347 TikTok users and 9.2M recommendations, only 10.3% of viewed videos originated from followed accounts (https://doi.org/10.1145/3613904.3642433). Note what that does not say: it measures video origin, not delivery path, and it is TikTok rather than Instagram. It does not establish that reach is re-earned per item — it establishes that most of what people watch does not come from accounts they follow.

When the audience can verify

The default law is that reach selects for the legibility of expertise rather than its correctness — because most audiences cannot check. If yours can (technical, professional, scientific), that inverts: correctness stops being a moat choice and becomes closer to a distribution requirement. You cannot out-polish a full-time creator, and you do not need to. The differentiated format for a builder is usually the one that shows the work: take a claim everyone repeats, traverse to the primary source, report what survives.

Guard against the failure that pairs with it: an artifact that performs its own claimed mechanism on you and offers the sensation as proof. Feeling an effect is not evidence that the proposed explanation of it is correct.

The gate (deterministic)

python3 scripts/format_lint.py DRAFT.md            # ERROR on refuted claims
python3 scripts/format_lint.py DRAFT.md --strict   # warnings fail too
python3 scripts/format_lint.py - --json            # stdin, machine-readable

Severity follows the ledger's grade: of the claim rules, only refuted (contradicted by a primary source that was loaded) exits non-zero. Malformed lint controls — an unclosed fence, an unclosed or nested disable — also exit non-zero, because a silently disabled gate is worse than a noisy one. contested, unverified, folklore and hypothesis_as_fact are WARN — because "I could not find a source" is a statement about a search, not about the world, and a gate that conflates the two manufactures the false confidence it exists to prevent. Use --strict to fail on warnings too.

Sentences that negate, correct, or attribute a claim do not fire, so the linter does not punish the corrections it exists to promote. Fenced blocks and well-formed YAML frontmatter are exempt; a leading --- horizontal rule is not. Suppress narrowly with <!-- format-lint: allow=<rule-id> --> (that line only) or a disable/enable region — an unclosed or nested region is itself an ERROR, and any region that hides a finding is reported as one (suppressed-findings, naming what it hid and where).

The bypasses, stated plainly, because a gate whose escape hatches are undocumented is a gate you cannot reason about:

A disable region cannot hide anything quietly. Whatever it suppresses is reported as suppressed-findings, naming the rules and the line range:

WARN  L2  (suppressed-findings · lint_control)  «format-lint: disable»
      This region hides 2 finding(s) on lines 3-4: algorithm-punishes, three-second-hook.

It is the one guard here with no threshold in it.

That sample sits in a fenced block for a reason, and the reason is worth a sentence: a rule id such as algorithm-punishes contains the very string its own rule matches, so quoting the linter's output in prose makes the document flag itself. Fence the output, or mark the line — the second option is what the line you are reading does.

Its severity is inherited from the worst thing it hides. Suppressing warnings is ordinary editorial practice and stays a WARN; suppressing an ERROR is itself an ERROR, so wrapping a refuted misquotation in a disable region cannot turn a failing document into a passing one.

That is deliberate. A coverage-ratio guard used to live here, erroring when a region covered "most" of a document, and six consecutive review rounds defeated it with six different ways of padding the denominator: fenced code, frontmatter, lint's own markers, bare --- rules, - bullets, |---|---| separators, >, #, and finally ordinary HTML comments. Each round closed one shape and left the class open, because a ratio over "prose" invites an argument about what counts as prose and every answer was wrong at a new edge. Reporting the fact instead ends the argument: padding cannot change what a region hid.

The citation bypass is the one to watch. A WARN-grade rule asserts "this circulates with no located source", so a sentence that supplies a source has already done what the rule asks — "According to Instagram, its algorithm demotes primarily non-original accounts: https://creators.instagram.com/blog/rewarding-original-creators-on-instagram" is true, cited, and was flagged as folklore until this existed — the sentence you just read is itself suppressed by the rule it describes.

The citation must be a resolvable locator: an http(s) URL with a real host and a path, a DOI in either form, or an arXiv / PubMed / PMC identifier — arXiv:2301.00001 and PMID: 12345678 included, not only their URL forms. A bare example.com/page does not suppress, and neither does a naked https://, a malformed host like https://..com/path, or an identifier embedded in another word like NOTPMC1234. Note that references/evidence-map.md cites several sources as backticked bare domains — that house style is not a suppression marker, deliberately: a bypass should be narrower than a citation convention, not wider. The consequence is that a resolvable link in the same paragraph silences those two grades' claims in that paragraph.

The cut is per rule, declared as citation_resolves in the ledger, and the question it answers is narrow: does supplying a source answer THIS rule's complaint? Four of eleven rules say yes. The rest say no, for three different reasons:

• The literature disagrees (contested). Citing one side does not settle a disagreement.
• A mechanism is asserted as established (hypothesis_as_fact). A citation does not make it so.
• The complaint is about source QUALITY, not source existence. three-second-hook's finding is that every located source is a content farm — so a URL would satisfy the bypass while confirming the complaint. A regex cannot tell a content farm from a journal, so this rule does not opt in.

Two coarser cuts were tried and both were wrong. Keying on severity swept in contested and hypothesis_as_fact, which a cross-model review caught. Keying on grade still swept in three-second-hook and post-daily, which probing the fix caught. An absent key means no bypass: it must be opted into, never inherited.

The scope is the paragraph, deliberately narrower than the ±3-line window the precision rule uses. A line window let a URL in a different paragraph — even one above the claim — silence it, which turns "cite your source" into "put a link somewhere nearby". Block scope still spans a hard wrap, which is the only thing it ever needed to span.

Extend it by editing references/claims-ledger.json — pattern, message, and the correct replacement. Every rule needs both polarities in tests/.

What counts as a citation is code, not configuration. It used to be a ledger field, and guarding it turned out to be impossible in that shape: any finite set of "this must not count" examples can be evaded by a pattern written against it, and any finite set of "this must count" examples wrongly rejects a legitimately narrower policy such as DOI-only. The two pull against each other by construction. A resolvable locator — an http(s) URL with a path, a DOI, an arXiv, PubMed or PMC id — is now fixed in the code, because making it configurable was making the gate configurable into silence.

The ledger is validated strictly, including its key names. An unknown top-level key or rule key is an error, not an extension point: precision_without_sources, one letter off, would otherwise load cleanly and switch off an entire rule class while the run kept reporting documents as clean. Nothing may quietly change what is checked without changing what is said. Nor may a refuted rule opt into the citation bypass, nor a pattern or citation marker match the empty string — each of those is a way to configure the gate into silence, and each is refused at load.

Before you trust it, sweep your own archive.

python3 scripts/corpus_sweep.py ~/writing                       # what fires, per rule
python3 scripts/corpus_sweep.py ~/writing --compare old.json    # what a ledger edit changed

A rule that fires on a third of your existing work is a rule you will learn to ignore, and a widened pattern is where the next false positive comes from. --compare runs both ledgers over the same tree and prints only the delta; every added finding has to be a true positive or the coverage was bought with noise. That check is not decorative — the first draft of the word-form-multiplier rule added three findings across 3,328 files and two of them were the enumeration idiom ("the problem is threefold"), which is why the rule now requires a measurement context. Rerun the command to regenerate those numbers.

Anti-rationalization

Provenance

Derived from broomva/workspace BRO-2145 (PR #376, merged 195a9e54): a full ingest of a paid-acquisition funnel plus its 36:56 sales video, transcribed locally, with every load-bearing claim traced to a primary source. That upstream research arc was cross-reviewed over three rounds scoring 2/10 -> 5/10 -> 6/10, catching two genuine errors in it.

This skill's own review record is separate, and long. Twenty-two cross-model adversarial rounds, scoring 2, 5, 5, 5, 6, 6, 5, 5, 6, 6, 5, 6, 5, 6, 6, 6, 6, 6, 6, 6, 8 (one round timed out before rendering a verdict, though it produced a real finding), plus a dogfood pass against a real unseen article that found two defects all 47 tests had missed. It passed on the twenty-second.

Eighteen rounds sat at 5-6, and the reason is the most useful thing here. At round 21 the reviewer was asked directly whether that reflected a defective artifact or a surface a finite review cannot exhaust, and answered: defective — "twenty rounds finding repeated silent bypasses in the same core detector is not merely arbitrary-input inexhaustibility." It was right. Three consecutive rounds had each found one branch of the citation-locator alternation missing a boundary condition, and I had been adding the missing boundary each time. A rule spelled out per branch is a rule you can forget once per branch.

Replacing the per-branch boundaries with a single token-alignment invariant — applied to every match whichever branch produced it, so that adding a branch cannot reintroduce the bug — moved the score from 6 to 8 in one round, and the reviewer withdrew the diagnosis: "the defective per-branch shape is gone."

Coverage went 47 tests to 184, with a mutation proof for every fix. What the rounds found:

• Universal-absence phrasing — the defect this skill exists to prevent, in this skill. Sites throughout asserted absence as a property of the world rather than of a search: format clustering "documented nowhere", absent from "any PUBLIC documentation", "there is no per-format punitive state", "unevidenced", "every source is a content farm", "no source establishes". Every one is a claim about a search, which is precisely the grade this skill defines as unverified.

  It took four passes, and the reason is worth more than the count: the first three swept for the wording already found, so each pass caught only rephrasings of the last. The pass that converged swept for the shape — any sentence asserting absence without naming what was searched — and found sites the wording-based greps could not see. A count produced that way is a count of what the pattern happened to match, so none is quoted here; the sites that remain quoting the old phrasings are these, describing the defect.

• Fence tracking treated any ``` or ~~~ as a toggle, so a ``` inside a ```` block closed it early and the rest was linted as prose. A fence is now a character and a length, and a closing run must carry no info string.

• Frontmatter parsing, twice. A bare ``` in a YAML value opened a fence no --- could close, so the entire document went exempt and its claims were silently missed; frontmatter now resolves before the fence scan. And an *indented* --- inside a block scalar was mistaken for the closing delimiter, ending the frontmatter early and linting YAML as prose; the delimiter is now column-zero only.

• The citation bypass, which a reviewer's counter-example forced into existence and a later round twice narrowed: first from a ±3-line window to the claim's own paragraph (a URL in a neighbouring paragraph was silencing claims), then from "any WARN grade" to unverified and folklore only — because contested means the literature disagrees, and citing one side of a disagreement does not settle it.

• A vacuous guard, deleted. citation_markers was a ledger list no code path read, and a test asserted properties of it under the docstring "Regression on a real finding". It passed whatever the linter did. The field is gone; the test now exercises the regex the linter actually compiles, in both polarities.

• Paraphrase coverage, widened where it was cheap and honest: anthropomorphic verbs beyond "punish" (demote/suppress/throttle/shadowban/bury/deprioritise), cadence phrased as "each day" or "seven days a week" or in the past tense, the variance hypothesis in five more verbs, and word-form multipliers. Each widening carries a paired negative fixture and was swept against 3,328 real files.

The exit code is the only signal, so it is classified deliberately. 0 clean, 1 findings, 2 bad input. A ledger fails to load in three unrelated ways and they are not conflated: anything about the path or content you supplied (missing, permission-denied, a directory, not UTF-8, not JSON, schema-invalid, a pattern that will not compile) exits 2 with no traceback, because you fix the ledger; anything about the machine (out of memory, a failing disk) propagates with its real traceback, because calling that "bad input" would send you to edit a ledger that is fine; and KeyboardInterrupt is never caught at all.

Still known-open. A ledger is trusted configuration. Its patterns are validated at load, but a pattern that compiles and then explodes at match time ((a+)+$) loads cleanly and will hang the run. Detecting that is a halting problem in miniature, and a heuristic for it would be a weaker gate than the one it guards. Treat a ledger the way you treat a linter config: as code you are choosing to run.

The matcher is regex-over-text, so a claim stated in words the ledger does not list still passes; the widenings above narrow that gap, they do not close it. Block boundaries are line-structural, not a real Markdown parse. Do not read a clean run as proof a document is sound; read it as proof it contains none of the specific strings in the ledger. Full graded claim set: references/evidence-map.md.

Recheck the ledger after 2027-02-12 — platform documentation moves.