Lesson 08 — Provenance: keeping sources, drafts, and your own words separable

Learning objective

(Bloom's: Apply) Apply claude.ai's actual context-boundary and version-history mechanisms — the project knowledge base, per-chat context isolation, and artifact version history — to keep The Last Mill's sources, your own prose, and Claude-drafted text traceable and separable, beyond a single ad hoc labeling habit.

Standard named: AIHC.1.C1 (Provenance). Provenance is kept as a first-class property of the work (X5), practiced as an ongoing filing discipline (P4) — not a feeling you have about your own honesty, and not a single feature Claude provides for you automatically.

Cycle 1 — What Claude actually keeps track of, and what it doesn't

SAY — Projects are "self-contained workspaces, each with its own chat history and knowledge base" [S03]. But there's a sharp edge: verbatim, "Context is not shared across chats within a project unless the information is added into the project knowledge base" [S04]. If you paste an interview note into one chat and never add it to the knowledge base, your other chats in that same project — and any later provenance check — can't see it. Provenance starts with knowing what's actually tracked.

SEE — A described diagram of Priya's real Project: a "knowledge base" panel on the left holding uploaded files, added via the "+" button in that panel [S04]; several chat bubbles on the right, with a dashed line showing which chats can see the knowledge base and a flagged chat bubble that only has a raw paste — no line connects it, meaning nothing else in the project can draw on it.

DO — Open your real Project. For your three real source files (the draft, interview notes, the pitch tracker), check which are actually sitting in the project knowledge base versus just pasted into one chat. Add any missing one via the "+" button now [S04].

⏸ Pause point. Knowledge-base membership checked and corrected. A complete, resumable step — stop here if your block is up.

Cycle 2 — Your draft can carry its own version history

SAY — Claude turns "significant and self-contained" content — typically over 15 lines, something you're likely to edit, iterate, or come back to — into an artifact, a dedicated window separate from the chat [S05]. Artifacts support "switch[ing] between artifact versions without disturbing Claude's memory of the original" [S05]. That version history is a real provenance mechanism: which draft passage existed when, before your next edit.

SEE — A described mockup of the sidebar "Artifacts" section [S05] showing "The Last Mill — Section 3" with three stacked, timestamped versions listed underneath.

DO — If your current draft section isn't already an artifact, make it one. After your next real edit, confirm you can step back to the prior version without losing it.

⏸ Pause point. One artifact, version history confirmed. Stop here if needed.

Cycle 3 — Test a risky rewrite without touching the record

SAY — The "Customize" button "opens a new conversation seeded with the artifact's content — edits there don't touch the original" [S06]. That is a clean way to try a bolder rewrite without risking the version an editor has already seen. Referent (true-crime podcasts, flavor only): a careful investigator doesn't scribble a new theory directly onto the original case file — they work the theory on a copy, and the original stays intact either way.

SEE — A described two-pane view: left pane, the original published artifact, untouched; right pane, a "Customize"-spawned conversation where Priya is trying an alternate opening paragraph.

DO — Use "Customize" [S06] on one real paragraph of The Last Mill you're unsure about, and try an alternate version there — leaving your working artifact untouched.

⏸ Pause point. One experiment run, original preserved. Stop here if needed.

Cycle 4 — Build your own three-way mark, on top of the real mechanisms

SAY — None of S03–S06 automatically labels whose words are on the page — that convention is yours to build, using the real mechanisms above as the filing system underneath it. A workable convention: mark every passage as your prose / Claude-drafted, you revised / source quote. One honest limit to know first: project files get "text extraction only, except multimodal PDFs (which keep visual content)" [S07] — so if an interview transcript is a scanned, image-only PDF, confirm Claude actually has the words, not just an unreadable image, before you cite it precisely.

SEE — A described half-page of Priya's real draft with bracket-tags written in by hand — [mine], [Claude-drafted, revised], [source: interview #2] — next to the corresponding lines.

DO — Mark one real page of The Last Mill with your three-way convention, in your Project, now.

Independent at-bat

Unscaffolded, full loop on your next real section: file a new interview transcript into the knowledge base (checking first that its file type is one Claude can actually read the text of, per S07); turn the resulting draft passage into a version-tracked artifact; apply your three-way marks to it. No checklist provided this time.

Exit ticket (climbing to Apply; graded against the doc-set)

1. (Remember) Name the doc-set feature that keeps content invisible to a chat unless you deliberately add it there. [S03/S04]
2. (Understand) Why does turning your draft passage into an artifact protect its provenance better than leaving it as plain chat text?
3. (Apply) You want to try a bolder rewrite of your lead paragraph without risking the version your editor has already seen. Which real feature do you reach for, and why?
4. (Apply) Your next interview transcript is a scanned PDF of handwritten notes, 40 pages. Per S07, what can you actually count on Claude having read, and what should you not assume it saw?
5. (Apply — objective level) Take your real pitch-tracker note for tomorrow's interview: file it correctly (knowledge base or a single chat?), and state which of your three marks would govern anything Claude drafts from it.

Ledger write

ledger_write:
  learner_id: L3-CONS-PRIYA
  lesson_id: L3-08-C1
  standards: [AIHC.1.C1]
  tags: [P4, X5]
  exit_ticket:
    score: 
    bloom_reached: 
  auto_score: 
  self_score: 
  calibration_gap: 
  journal_prompt: >
    Before this lesson, how did you keep track of what was yours versus Claude's versus a
    source's, in your draft? Name one place that habit would have failed you, and how the
    knowledge-base / artifact-version / three-way-mark combination closes it.
  structure_used: PBL
  referents_used: [true-crime-podcasts]
  next_lesson_seed: "carry the knowledge-base discipline forward into Lesson 09 (C2 Governance), where the question becomes what should never go into that knowledge base unredacted."

Rubric self-audit (R1–R13)

Escalation verdict: all load-bearing indicators PASS → clears; auto-ships.