ingest-writing

name	ingest-writing
description	Paste any raw material and route extractions into the appropriate context pages (voice / about / vocabulary / preferences)
risk_class	soft

Ingest writing

The user pastes raw material — a draft, an email, an article they wrote, meeting notes, anything. This skill classifies the content, extracts what's useful, and routes the extractions into the right context pages. Append-only; never silently overwrite.

When to invoke

• The user explicitly says: "ingest this", "learn from this", "remember this about how I write", "this is what I sound like", or pastes a wall of text with a directive like "use this".
• /onboard hands off to this skill after its lightweight interview, if the user accepts the paste-raw invitation.
• /wrap-up invokes this skill when settings.ingest.auto_on_wrap_up is true and there is end-of-day material to ingest.

Gating

Before extracting, read the active settings:

rubber-ducky settings get ingest.kinds

The returned list bounds the extraction surface. If the user has restricted ingest.kinds to ["voice"], do not extract facts, vocabulary, or preferences from the pasted material — even if they're obviously present. Respect the user's stated scope.

Behavior

Step 1 — Classify the input

Read the pasted text. Tag what it contains:

Tag	Looks like	Routes to
voice	First-person prose, the user's own words at length.	wiki/voice.md
facts	Statements of who-does-what, role descriptions, project descriptions.	wiki/about.md
vocabulary	Acronyms, internal terms, codenames, jargon.	wiki/vocabulary.md
preferences	Statements about style, tone preferences, things to avoid.	wiki/preferences.md

A single paste can produce extractions in multiple tags. Don't force a single classification.

Step 2 — Read existing context pages

Before extracting, query each target page so duplicates aren't appended:

rubber-ducky context query <kind>

The CLI returns each ## section's body. Skim the existing entries — extractions that say substantively the same thing as an existing entry should be skipped, not duplicated.

Step 3 — Build the extraction summary

Don't write anything yet. Present the user with a structured summary:

EXTRACTION SUMMARY
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
voice (1 new sample)
  > "<first ~80 chars of the sample>…"
    Why kept: first-person, long-form — useful for tone calibration.

facts (2 new)
  - User leads the platform team at <employer>.
  - <Project codename> is the migration off the old auth stack.

vocabulary (1 new)
  - **HOOTL** — hands-off, out-of-the-loop.

preferences (skipped — no new patterns beyond what we already have)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Write all of these? (yes / no / pick)

If the user picks pick, walk them through each candidate one at a time (one question per turn) and write only what they greenlight.

Step 4 — Append to the pages

Hold to two invariants on every write:

1. Append-only — new content is added at the end of the matching section, never replacing or rewording existing lines.
2. Source-attributed — each new entry begins with a ### Source — YYYY-MM-DD heading (date in ISO form) plus a one-line note describing what the source was (e.g., "Pasted draft of OKR update").

For voice.md, each appended sample lands under ## Samples. For about.md, route into ## Identity, ## Projects, or ## People depending on the fact's nature. For vocabulary.md, route into ## Terms, ## Acronyms, or ## Internal system names. For preferences.md, route into ## Style, ## Pet peeves, or ## Decision-making.

If a heading the extraction needs is missing (e.g. an ## Acronyms section in a brand-new vault), create it; do not invent additional headings.

Step 5 — Audit log

Once writes complete:

rubber-ducky log append "[ingest-writing] +<N> voice / +<N> facts / +<N> vocab / +<N> prefs from pasted material"

Step 6 — Confirm and pivot

End with a one-line confirmation: "Added N entries across <pages>. Anything else?" If the user pasted before answering an active task or working flow, pivot back to that — don't strand them in the ingest path.

What this skill does NOT do

• No silent overwrites. Existing content is never rewritten — even when the new material seems to supersede the old. Append a new source-attributed entry; let the reader interpret the timeline.
• No extraction outside ingest.kinds. Respect the user's stated scope even when the source material is rich in other categories.
• No interpretation of what the user meant. Quote what the source said; don't paraphrase facts into your own voice. Preserving the source is the whole point of source-attribution.