Bulk path

The user has a pile of content — notes, transcripts, pasted text, exported documents — that hasn't been organized yet. Rather than scanning a directory, the agent processes the pile through dialogue.

Signals you're on this path

• User pastes content directly into the conversation, or references specific file drops rather than a structured directory
• Content is mixed-format and mixed-topic — no apparent native organization
• Volume is conversation-sized: tens to low-hundreds of documents, not thousands
• User is actively present to react and disambiguate

Not this path if the volume is too large for one conversation. Large piles with no structure are a survey problem plus embeddings — see the Onboarding mode entry in ROADMAP.md.

Not this path if content already has structure on disk — that's survey.

Inventory rules

Dialogue-driven, not filesystem-driven. Ask the user to paste or describe three to five representative samples from the pile. Ask them what types of things they see (transcripts, notes, half-drafts, quotes, etc.).

Do not try to read every document before proposing structure — the conversation won't fit. Sample, cluster mentally, and propose.

Keep a running list of topics, entities, and recurring types as the user shares samples. Read back what you're hearing every few samples so the user can correct.

Sense-making

Because the content lacks native structure, the "what's your existing structure?" question at Level 2 is blunted. The sequence:

1. Ask about real-world structure, not content structure. Even if the notes are unorganized, the domain they're about has structure. "These are all notes about your consulting clients — do you already slice clients by industry, by stage, by account team?" Use whatever comes back.
2. Framework fallback. If the domain is net-new or the user can't articulate structure, offer frameworks.
3. Derived clusters. Propose 3–5 collections from the samples you've seen, grounded in the types of things the user pasted.

Seeding

1. Write the root index.md and collection skeletons as in other paths.
2. Write one exemplar node per collection before fanning out. See guardrails. With an unstructured pile, the exemplar-first rule is especially important — without it, parallel agents produce wildly different interpretations.
3. Work through the pile in batches with the user:
   • User pastes a document or batch
   • Agent classifies it into a collection, extracts name + description + any obvious links
   • Agent writes the node file
   • User confirms or corrects before the next batch
4. For documents that don't fit any collection cleanly, stash them in a _inbox/ directory (underscore-prefixed, not compiled) and revisit at the end.
5. When the pile is exhausted, compile and walk the graph.

Gotchas

• Paraphrase drift. An agent classifying documents at speed will paraphrase loosely and lose the user's voice. Read back descriptions verbatim every five to ten nodes so drift gets caught.
• Missed entities. Inline links to people, clients, projects that haven't been created yet as nodes produce dangling references. Either create those entities as you encounter them (preferred) or queue up a "create these" list and batch-write them after the main pass.
• No natural stopping point. An unstructured pile doesn't know when it's "done." Agree on a stopping criterion with the user upfront: time-boxed (30 minutes of processing), volume-boxed (first 50 documents), or coverage-boxed (until every major topic has at least one node).