version: "1.0.0" name: note description: Use this skill to record progress, partial findings, hypotheses, and page-anchored evidence while investigating a long document. Call Note aggressively — any time you read a page and confirm, narrow, or update your understanding of the user's question. Triggers include "jot that down", "save this finding", "note that page 7 says X", or (more often) implicit internal signals — you just learned something and want it recoverable later. Notes persist in the session and Review can select among them. Never paraphrase notes back from memory; write them as you go. license: MIT compatibility: Requires Python 3.10-3.13, the complete DocAtlas Skills bundle, and a DocAtlas session file. metadata: author: Microsoft Research version: "0.3.0"

Note — append a progress-analysis note

Evidence requirements

Every factual Note call must include evidence copied from the source. Prefer an exact quote in a type:"text" item, and include a physical page reference in source, such as "Page 33" or "Page 12, Table 2". Strategy-only and dead-end checkpoints may omit evidence when they make no factual claim.

BAD vs GOOD example:

// ❌ BAD — paraphrase, no page ref
{"type": "text", "source": "survey data", "content": "Most Republicans think the president can pardon himself"}
 
// ✅ GOOD — exact quote with page ref
{"type": "text", "source": "Page 7, Q24a", "content": "73% of Republicans and Republican-leaning independents say the president can legally pardon himself"}

If you cannot produce an exact quote, go back and Read the page first. Do not Note from memory.

This skill is half of the Note / Review pair that gives a multi-turn investigation a memory. It writes one structured entry per call — what you found, what you plan next, and the evidence backing the finding — into an append-only timeline stored on the session file.

Why bother writing notes?

Long-document work is read-heavy. By the time you're at turn 8 you've already seen pages 1, 3, 5, and 12; without notes you'd be lugging those chunks around in every prompt. Writing a note lets the next turn keep working from a short summary while the evidence stays addressable by note_id. Review can then cherry-pick the notes relevant to a final answer.

When to use

• You just confirmed something concrete on a page (a number, a claim, a

definition) — write a note with page-anchored evidence.

• You formed a hypothesis you might want to revisit later.
• You finished one sub-task of a multi-part question (e.g. "figured out

partisan split on ethics") and are about to pivot to the next sub-task.

• You hit a dead end and want to record that it was a dead end (so you

don't redo the same search).

When NOT to use

• For trivia already obvious from the user's message.
• When the finding is purely speculative with no supporting page. Go read

the page first, then note.

Arguments

All fields are per-call (the model decides them). Pass via JSON:

uv run --locked python docatlas/skills/note/scripts/run.py --json '{
  "found": "Page 7 shows Republicans +30pt more likely than Democrats to say president can legally pardon self.",
  "plan": "Next, check page 9 for the ethics-charges question.",
  "evidence": [
    {
      "type": "text",
      "source": "Page 7, partisan split table",
      "content": "55% of Republicans and 25% of Democrats say the president can legally pardon himself"
    }
  ]
}'

Provide at least one of found, plan, or evidence.

Evidence item shape

{
  "type": "text" | "table" | "image",
  "source": "Page N" | "Page N, <title>",
  "content": "quoted excerpt or caption",
  "filename": "for image items only"
}

Output

One JSON object on stdout:

{
  "text": "Saved as note #2 (step 5). Total analyses so far: 2.\n\n[Step 5] …rendered note…",
  "note_id": 2,
  "step": 5,
  "analysis_count": 2
}

Examples

Example — save an anchored finding

After Read returned page 7:

Direct calls require HARNESS_SESSION_FILE; see ../README.md.

uv run --locked python docatlas/skills/note/scripts/run.py --json '{
  "found": "73% of Republicans vs 31% of Democrats say the president has legal authority to pardon himself (Pew 2018).",
  "plan": "Cross-check with the partisan table on page 9.",
  "evidence": [
    {"type": "text", "source": "Page 7", "content": "73% of Republicans and 31% of Democrats say the president has legal authority to pardon himself"}
  ]
}'

Tips for combining with other skills

• Read → Note: after every meaningful page-read. Don't wait.
• Note → Review: when you're ready to synthesize — Review lets an

aux LLM select the notes relevant to a question.

• Automatic tree enrichment: once a note with page references is

saved, the harness automatically lifts its page-findings into the PageIndex tree, so a later Search sees them as hints. (No extra call — it happens on the post-Note hook when tree annotation is enabled.)