Scout

Scout sits upstream of Forge. Where Forge takes sources and produces a pack, Scout takes a topic and produces a list of candidate sources. It searches across 15+ curated open knowledge databases and presents the results as a reviewable candidate list. You pick what to keep, Scout fetches the content into a temporary folder, and the batch is handed to Forge.

Click the Scout button in the sidebar to open the panel.

why Scout exists

A pack is only as good as its sources. Going from a topic ("Stoic philosophy") to a curated list of authoritative sources is the hardest manual step in building a pack. Scout automates the discovery while keeping the curation step explicit — you always review and approve before Forge spends real money on the build.

The source roster

Scout queries across:

Wikipedia, Stanford Encyclopedia of Philosophy (SEP), Wikisource, Internet Archive, OpenAlex, PubMed, arXiv, Crossref, MathWorld, Project Gutenberg, GitHub, StackExchange, Hacker News, and more.

Sources are toggled on or off per run. An Auto-detect domain mode enables only the sources relevant to the topic's domain by default — uncheck it to query the full roster.

The form

Topic

A single text field. Examples: "Julius Caesar", "transformer attention", "CRISPR-Cas9", "Stoic philosophy".

Be specific. A vague topic produces vague candidates.

Scope chips

Six chips define the angle you want for this pack. Pick exactly one:

Chip	Use for
📖 Overview	Foundational overview — suitable for a top-level or intermediate node in your knowledge tree. Covers main schools, key figures, central concepts, historical context.
🎯 Deep dive	Specialised material on this exact topic — sub-topics, primary sources, scholarly debates.
⚖️ Critique	Debates, criticisms, controversies, modern reassessments.
🔄 Comparison	Compares the topic with adjacent traditions or alternatives.
🛠 Practical	Applied / how-to material — modern uses, exercises, real-world application.
✏️ Custom	Free-form. Follow the Brief literally with no scope template.

Brief

An optional free-text constraint: what should this pack cover, focus on, or exclude?

most powerful field

The Brief is the most powerful field in Scout because it filters every later step — query expansion, source ranking, and (when you click into Forge) concept extraction. A precise Brief produces a precise pack.

Suggest from topic

The ✨ Suggest from topic button asks a cloud model to fill the Brief and pre-select relevant sources, based on the topic and scope you chose. Costs about €0.001 per click. Recommended as a starting point — you can edit the generated Brief before running.

Language and Pack name

Language — Scout and the resulting pack will be built in this language. Default English. Switch only for locale-specific topics. One pack equals one language.
Pack name (target) — the pack identifier where the fetched sources will land when you click Run Forge →. Pre-fills if Scout was opened from inside Forge.

The three modes

Scout offers three discovery modes, with their estimated cost shown inline:

Wikipedia-first (€0)

Walks Wikipedia outward from the topic. Zero LLM cost. Best for well-covered encyclopedic subjects.

Options:

Walk depth — 1 / 2 / 3 hops.
Max pages — 15 / 30 / 60 / 100.
Link strategy — "See Also + categories" (focused) or "See Also + inline links" (wider, noisier).
Disambiguate via Wikidata first — resolves ambiguous names (e.g. Cesare → Giulio / Beccaria / Cremonini) to the right page before walking.

Multi-source (~€0.10)

Fan-out across the full source roster. An LLM expands your topic into 6 angled sub-queries and ranks the combined results. You review and pick.

Options:

Sources grid — checkboxes for each database. Auto-detect domain enables only the relevant sources by default; uncheck it to query all.
Per-source results — 4 / 6 / 10.
Final top-K — 20 / 40 / 80 candidates after ranking.

Full auto (€0.30 + Forge)

Multi-source + auto-pick top results + Forge + Lint, all chained. No human review until the end. Use this when you trust the topic to be well-defined and want a hands-off build.

LLM picker (Multi-source and Full auto)

Both LLM-driven modes show a picker:

Provider — Top-tier OpenRouter (recommended) or Local Ollama (free).
Model — populated automatically from the provider's available models.

The two stages — Find, then Fetch

Scout works in two distinct stages, and the buttons reflect that.

Stage 1: Find Sources

Runs the search. Candidates appear in the Candidates pane with a checkbox each. You can filter by source, by language, or select all. A live activity log streams below.

Stage 2: Fetch sources

Downloads or synthesizes the selected sources into local markdown files. Forge is not yet run. After fetch completes, a second action bar appears:

⬚ Open folder — opens the temp folder so you can inspect or hand-edit.
↺ Discard — throws everything away and lets you start over.
Extract model and Writing model pickers — same options as in Forge.
Run Forge → — hands the batch to Forge and starts the pack build.

why two stages

The two-stage flow is intentional: it lets you preview, prune, and curate before paying for the Forge run. Discovery is cheap; pack building costs real money. Insert a human review between them.

Typical workflow

A pack built end-to-end with Scout + Forge usually goes:

Open Scout, type a topic, pick a scope.
Click ✨ Suggest from topic. Edit the generated Brief.
Choose a mode (Wikipedia-first for general topics, Multi-source for technical / academic ones).
Click Find Sources. Review candidates, deselect anything off-topic.
Click Fetch sources. Optionally inspect the folder.
Click Run Forge →. Forge builds the pack with the selected Extraction and Writing models.
After Forge finishes, run Lint on the result.
Either keep the pack private or submit it to the community via Submitting a pack.