Semantic curation cycles.

1The problem

Generative AI made production cheap. It did not make deciding what to say any easier. Most brand content still comes from intuition, from whatever a competitor published last week, or from a trend that has already become consensus. The result is volume without a point of view: the same topic, the same angle, published by everyone at the same time.

Teams feel this as a set of recurring complaints:

• “We never have a backlog of ideas. Every week starts from zero.”
• “We publish a lot, but nothing has real depth.”
• “The voice changes from channel to channel and from writer to writer.”
• “We always arrive late to the conversation.”
• “Nobody can say why this piece was a priority and that one wasn't.”

All of these come from the same cause. The team has a production process but no curation process: no system that captures, filters and stores what is worth thinking about before anyone opens a blank document or an empty prompt.

2The pipeline

The method runs as a pipeline of five databases, built in Notion, each one feeding the next. Every stage has its own fields, filters and quality criteria, so editorial decisions become data instead of gut feeling.

Sources: a radar, not a reading list

The radar combines around 190 feeds curated in Inoreader (publications, newsletters, research outlets, YouTube channels and niche communities) with an automated opportunity monitor described in section 3. A Python pipeline rates and groups roughly 60 relevant sources a day into Notion, and Claude, connected through MCP, turns those groups into editorial briefs.

Sources are not equal: each carries a rank, and the system measures over time which ones actually turn into content.

Signals: positioned on the Futures Cone

No signal moves forward without a diagnosis. The curator places it on a Futures Cone, which decides how the brand should treat it:

Each signal also carries an urgency level, a potential rating, a creative application (how a brand could use it) and a guardrail (how not to use it, the misreadings and opportunistic takes that would damage credibility). That second field is what turns a trend list into a brief a team can act on safely.

Concepts: where the thinking lives

Signals are perishable; concepts are not. A concept records the thesis the brand defends, the tension the subject reveals, and the territory it belongs to. Because concepts accumulate, the repertoire grows richer with every cycle: new signals get connected to existing theses instead of starting from scratch.

This is the stage AI cannot replace. The model generates variations; the curator decides which story a data point should tell and which tension a subject reveals.

Production: one concept, a full trousseau of pieces

Once a concept is approved, production is fast and repeatable. A library of reusable Skills (structured instructions, templates and examples for the AI agent) turns one approved brief into 9 to 12 channel-specific drafts: newsletter, LinkedIn post, carousel, thread, article, video script.

Every draft starts from the same intellectual core, so depth and voice stay consistent across channels. An Agent Manual encodes the brand's identity, tone, quality standards and selection criteria, and every piece goes through human review and fact-checking before it ships. Once the repertoire is in place, a first publishable piece comes out in one working day.

3Listening to demand: the Opportunity Monitor

Supply-side curation (what the world is saying) needs a demand-side counterpart (what the audience is struggling with). The Opportunity Monitor is a Python system that runs three times a week through GitHub Actions, collects signals from six sources, rates them editorially and writes prioritised opportunities into Notion, each tagged by audience, funnel stage and territory.

Pain first, territory second

The monitor searches with the language of the audience's pain, taken from how each audience describes its problem before it knows the brand, and only then classifies results into the brand's editorial territories. It did not start that way, and the change came from data.

How a signal is rated (0 to 100)

Source weights come from observed yield, not opinion

Each source's weight reflects how often its items actually left the discard pile and entered the editorial funnel, measured on the same 778 records (April to August 2026):

• Reddithot posts in each audience's communities75% n=8
• Google Trendsrising and top related queries22% n=9
• Google "People Also Ask"real search questions14% n=240
• YouTube autocompletesuggestions with modifiers6.5% n=354
• Google Newsheadlines from the last 10 days3% n=167

Small samples are reported as such. The point is not the exact percentages but the discipline: the system is recalibrated against what editors actually used.

Memory: recurrence and grouping

1. 01

   Recurrence ledger

   Every item gets a signature (normalised title, sorted meaningful tokens). A known signature does not create a duplicate row; it increments a recurrence counter. A theme that keeps coming back is flagged as an evergreen candidate, and a rising recurrence rate warns that the search vocabulary is exhausted.

2. 02

   Grouping

   Different phrasings of the same question ("How do I recover Instagram reach?", "Why is my reach dropping?") are merged when they share a core (Jaccard similarity of 0.5 or more). The variants are kept as entry angles for the same piece.

3. 03

   Discards teach the system

   Nothing is deleted. A redundant but good item is linked to its canonical version instead of discarded, because the discard pile is the training signal for future calibration.

4Governance: where humans stay in the loop

The rule is simple: automation handles collection, structure and formatting; editorial judgement stays human. The agent does not replace the strategist. It executes what the strategist defined, with the same criteria every time.

5What it changes in practice

Measured results from running the method on my own operation

• 4 h → 20 min

  of production time for a recurring weekly intelligence report, once packaged as an automated template.

• 1× → 4×

  faster editing of my two newsletters with the "Curation Factory", a Make.com flow (Inoreader, full-text extraction, OpenAI translation and summary, Notion or Google Docs).

• 1 → 9 to 12

  drafts per cycle from one approved brief, all from the same intellectual core.

A public sample of the Curation Factory →