add-context

name	add-context
description	Add context to the active project. Triages the input as a meeting transcript, document, personal note, or chat thread — then routes to the right writer. Replaces the four `/ctx-*` entry points.
user_invocable	true

/add-context

One entry point for adding context. Reads the input, classifies it (transcript / doc / note / chat), and routes to the existing writer skill.

Scope (v0.2): thin router only. The four /ctx-* writers stay as they are. Downstream UX alignment + frontmatter standardisation tracked separately (T2.5, T2.6).

---

Invocation shapes

/add-context <file-path>
/add-context <file-path> -- <hint>
/add-context                              # then prompt for content
/add-context --type <transcript|doc|note|chat> <file-path>

• <file-path> preferred over pasted content. Pasted content sits in the conversation window and is expensive for long files. If the user pastes a long block, suggest they save it to a file first.
• -- <hint> is anything after a literal -- separator. Fed into the classifier as a hint.
• --type X hard override. Skips classification, logs override=true in telemetry.

---

Step 1 — Resolve project

Read the active project:

1. If your Claude Code session ID is in context (injected by the SessionStart hook), read [workspace-root]/.sessions/<session-id>.
2. Fall back to [workspace-root]/.current-session if the per-session file doesn't exist.
3. If neither exists or both are (no project), stop and say: No active project. Run /os-start to pick one, then re-run /add-context.

Look up the project in projects.md to get PROJECT_ROOT. For portfolios, parse left of / as portfolio name and use the portfolio root.

---

Step 2 — Read the input header

If the input is a file path:

• Read the first 60 lines via the Read tool.
• Capture filename + extension.

If the input is pasted content:

• Take the first 60 lines of the paste as-is.
• No filename signal available.

If no input was given:

• Print: Paste content below, or give a file path:
• Wait for input. Treat the reply as either a path or paste.

---

Step 3 — Classify

Skip this step if --type X was passed.

Pre-check: is this global inspiration, not project context? If the content is an external article, talk, or post the user found inspirational (not tied to one specific project), suggest /ctx-inspiration instead and stop:

This looks like global inspiration, not project context. Run /ctx-inspiration
to add it to context-library/inspiration/ where it's available across all projects.

Continue with /add-context anyway? [y/N]

Signals it's inspiration: filename or paste contains substack.com, youtube.com, medium.com, blog, talk-transcript with external speaker, or the hint says "inspiration" / "for the library" / "found this". When unsure, ask. If the user confirms project-specific, continue to scoring.

Score the four types in this order. First match with high confidence wins. If nothing scores high, ask.

1. Chat (highest-confidence signature)

Indicators:

• 5+ lines matching a per-message timestamp pattern:
  • WhatsApp: [DD/MM/YYYY, HH:MM:SS] Name: … or [DD/MM/YYYY, HH:MM] Name: …
  • Slack: Name [HH:MM AM/PM] on one line, message on the next
  • iMessage export: Date Time header followed by Name: lines
• Short messages (most turns < 200 chars)
• Reaction markers: <image omitted>, emoji reactions (👍, ❤️ on own lines), :thumbsup: (Slack)
• Filename hints: whatsapp, slack, imessage, dm, chat, thread

→ high confidence: route to /ctx-chat

2. Transcript

Indicators:

• Speaker labels at line start without per-message timestamps: Speaker 1:, Name:, SIMON:, or bold patterns **Name**:
• Long turns (avg > 200 chars per turn)
• Headers like Recording started at …, Transcript - YYYY-MM-DD, Meeting:
• Total length usually > 500 lines
• Spoken-language disfluencies — these don't appear in chats, docs, or notes:
  • Bracketed transcript artifacts: [inaudible], [laughter], [crosstalk], [unclear], [pause], [music] — strong signal even at 1–2 occurrences
  • Filler words at conversational density: um, uh, er, ah, hmm, you know, I mean, sort of, kind of — count cumulative occurrences; 5+ in 60 lines is high signal
  • False starts and self-corrections: I — well, what I mean is…, repeated words at the start of phrases
• Filename hints: transcript, recording, meeting, call, zoom, granola

→ high confidence: route to /ctx-transcript

3. Note

Headline signal: the user is the author. A note is Simon's own observation, reflection, or voice memo. If a different author owns the content, it's a doc — not a note — regardless of length.

Supporting indicators:

• First-person prose, single author voice ("I think", "my read", "today I…")
• Short input (< 200 lines)
• No speaker labels
• Filename hints: note, thought, reflection, observation, journal
• Pasted content with no source attribution defaults here when short

→ medium confidence: route to /ctx-note

4. Doc (fallback)

Headline signal: external authorship. A doc is content authored by someone other than the user — articles, briefs, analyses, transcripts of external talks, Substack posts, PDFs. Preserve provenance (URL, author, captured-date) in frontmatter.

Supporting indicators:

• Long-form prose with structure (headings, sections)
• Third-person voice
• Filename hints: memo, report, analysis, log, critique, teardown, brief, prd, decision, .pdf source
• Anything that doesn't match the above three

→ route to /ctx-doc

Two-artifact case. When the user wants to capture both an external piece and their reaction to it, treat as two adds: one doc for the source, one note for the reflection (the note should link to the doc). Don't fold them into one — the source attribution is lost.

Tiebreakers

• Chat vs note: when ambiguous, bias toward chat. A note misclassified as chat is harmless. A chat misclassified as note locks the thread out of dedup on the next reply (Standing Decision: chats must use ctx-chat).
• Transcript vs chat: Zoom exports can look chat-like with Person [HH:MM:SS]: per-message timestamps. Disfluencies ([inaudible], filler words, false starts) break the tie toward transcript — chats don't have them. Falling back to length: turn averages > 200 chars + no reactions/short-message bursts = transcript.
• Note vs doc: first-person + short = note. Third-person or > 200 lines = doc.
• The hint wins close calls. If the user wrote -- this is my reflection on …, route to note even if the heuristics say doc.

Confidence threshold

If no type scores high after the heuristics + hint, surface:

I think this is a [best-guess], but I'm not sure. Type one to confirm:
  [1] transcript    [2] doc    [3] note    [4] chat

Wait for selection. Don't route confidently below threshold.

---

Step 4 — Log telemetry

Append one line to [workspace-root]/logs/triage.log:

<ISO8601-timestamp> | <input-shape> | <classified-as> | <confidence> | <hint-used> | <override>

Where:

• <input-shape> is file or paste or none
• <classified-as> is transcript / doc / note / chat
• <confidence> is high / medium / asked
• <hint-used> is yes / no
• <override> is yes (--type set) / no

Create logs/ and logs/triage.log if they don't exist. Append-only. Don't read it back to the user.

---

Step 5 — Route to writer

Print a one-line confirmation:

Type: <classified-as>. Routing to /ctx-<X>.

Then invoke the matching skill via the Skill tool:

Classified as	Skill to invoke
transcript	ctx-transcript
doc	ctx-doc
note	ctx-note
chat	ctx-chat

Pass the file path (or pasted content path) to the writer. From here the existing writer's flow takes over: it'll prompt for any missing context (attendees, person link, thread name, etc.) per its own SKILL.md.

Do not duplicate the writer's prompts. If the writer asks for the project again, that's fine — the user confirms.

---

Edge cases

• Empty file: stop. Say <path> is empty. Nothing to add.
• File doesn't exist: stop. Say Can't read <path>. Check the file path and re-run.
• Binary file: stop. Say <path> looks binary. /add-context expects text (markdown, txt, transcript export).
• Path contains a hint-like --: the separator only takes effect when surrounded by spaces. /add-context /Users/me/foo--bar.md reads the whole token as a path.
• --type X with malformed value: stop. Say Unknown type <X>. Use: transcript, doc, note, chat.
• Pasted content > 500 lines: suggest Long paste — save to a file and re-run with the path. That keeps the content out of the chat window (cheaper).

---

What this skill does NOT do

• Frontmatter standardisation (tracked: T2.6, this week).
• Writer UX alignment, including decommissioning /ctx-doc's legacy current-state.md sync (tracked: T2.5, this week).
• Synthesis. The writers do their own synthesis as today.
• Validation flag work (tracked: T3.3).