@forwardimpact/outpost 3.12.0 → 3.13.0

package/templates/.claude/skills/anarlog-follow/SKILL.md

@@ -1,7 +1,7 @@
 ---
 name: anarlog-follow
 description: >
-  Follow a live Anarlog session in real-time, coaching the user through a
+  Follow a live Anarlog session in real-time and coach the user through a
   meeting or interview. Understands context from the session title, knowledge
   base, and candidate pipeline. Provides talking points, flags gaps in
   coverage, and suggests follow-up questions as the conversation unfolds.
@@ -10,48 +10,60 @@ description: >
 
 # Anarlog Follow
 
-Follow a live Anarlog recording, read the transcript as it grows, and coach the
-user through the meeting in real time. Gather knowledge-base context once before
-the session, then poll the transcript and provide actionable nudges as new
-content appears.
+Write tier: none
+Frontmatter: none
+
+Follow a live Anarlog recording. Read the transcript as it grows. Coach the user
+through the meeting in real time. Gather knowledge-base context once before the
+session. Then poll the transcript. Provide actionable nudges as new content
+appears.
 
 ## Trigger
 
 - The user asks to follow, shadow, or coach them through a live meeting.
 - "Follow my meeting", "coach me", "shadow this call".
-- The user starts a Anarlog recording and wants real-time support.
+- The user starts an Anarlog recording and wants real-time support.
 
 ## Prerequisites
 
-- Anarlog installed; sessions at
-  `~/Library/Application Support/anarlog/sessions/`.
+- Anarlog installed. Meetings live in its local SQLite database (`app.db`).
+  Read them only through Anarlog's typed, read-only interfaces. Prefer the
+  **Anarlog MCP tools** (`list_meetings`, `get_meeting`,
+  `get_meeting_transcript`) when the MCP server is connected. Fall back to the
+  **`anarlog` CLI with `--json`** otherwise. Never `grep`, crawl `sessions/`,
+  or query SQLite directly. See Anarlog's own `AGENTS.md`
+  (`~/Library/Application Support/anarlog/AGENTS.md`).
 - An active or about-to-start session.
-- Knowledge base populated (attendee / candidate context).
+- A populated knowledge base (attendee / candidate context).
 
 ## Inputs
 
-- Live `transcript.json` (growing during the session).
-- `Knowledge/People/`, `Knowledge/Candidates/{Name}/{brief,screening,panel}.md`,
-  `Knowledge/Roles/`, `Knowledge/Organizations/`, `Knowledge/Projects/`.
+- The live meeting transcript, via `get_meeting_transcript` (MCP) or
+  `anarlog --json meetings transcript <id>` (CLI). See
+  [references/sessions.md](references/sessions.md) for both interfaces.
+- `3-Team/People/`,
+  `2-Confidential/Candidates/{Name}/{brief,screening,panel}.md`,
+  `2-Confidential/Roles/`, `3-Team/Organizations/`, `3-Team/Projects/`.
 - `~/.cache/fit/outpost/apple_calendar/*.json` for context.
 
 ## Outputs
 
-- Real-time coaching messages printed to the user.
-- **No files are written.** This skill is purely advisory.
+- This skill prints real-time coaching messages to the user.
+- **This skill writes no files.** It only advises.
 
 <do_confirm_checklist goal="Verify the follow session was useful and read-only">
 
-- [ ] Active session detected and confirmed with the user.
-- [ ] Meeting type classified from the title.
-- [ ] Knowledge-base context gathered for all attendees.
-- [ ] For interviews: standard expectations and screening focus areas loaded.
-- [ ] Coaching nudges were actionable and concise (1–3 lines each).
-- [ ] Coverage gaps tracked and surfaced before the meeting ended.
-- [ ] End-of-meeting detected; debrief provided.
-- [ ] Next steps offered (`req-assess` / `anarlog-process`); user decided
-      whether to run them.
-- [ ] No files were modified during the session.
+- [ ] Detect the active session and confirm it with the user.
+- [ ] Classify the meeting type from the title.
+- [ ] Gather knowledge-base context for all attendees.
+- [ ] For interviews, load the standard expectations and the focus areas from
+      screening.
+- [ ] Keep each coaching nudge actionable and concise (1–3 lines).
+- [ ] Track the coverage gaps. Surface them before the meeting ends.
+- [ ] Detect the end of the meeting. Debrief the user.
+- [ ] Offer the next steps (`req-assess` / `anarlog-process`). Let the user
+      decide whether to run them.
+- [ ] Modify no file during the session.
 
 </do_confirm_checklist>
 
@@ -61,112 +73,125 @@ content appears.
 
 #### 1. Find the active session
 
+No interface exposes a "recording now" signal. The only liveness signal is the
+most recently written `audio.mp3` on disk. This is the one deliberate
+filesystem read in this skill. All content reads go through Anarlog's
+interfaces. The session directory name **is** the meeting ID:
+
 ```bash
-node .claude/skills/anarlog-follow/scripts/follow.mjs --detect
+find "$HOME/Library/Application Support/anarlog/sessions" -maxdepth 2 \
+  -name "audio.mp3" -exec stat -f "%m %N" {} \; | sort -rn | head -5
 ```
 
-Returns the most recently modified session, whether it's live, and its title. If
-nothing has been modified in the last 5 minutes, warn the user and ask whether
-to follow a specific session. If multiple sessions could be active, confirm with
-the user.
+Take the top result's parent directory name as the candidate meeting ID. If
+its mtime is more than 5 minutes old, warn the user. Ask whether to follow a
+specific meeting. If more than one session could be active, confirm with the
+user.
+
+#### 2. Read meeting metadata
 
-#### 2. Read session metadata
+MCP: `get_meeting({ meeting_id })`. CLI fallback:
 
 ```bash
-node .claude/skills/anarlog-follow/scripts/follow.mjs <session-id> --meta
+anarlog --json meetings get <meeting-id>
 ```
 
-Capture **title**, **created_at**, **participants**.
+Capture **title**, **created_at**, **participants** from `data`.
 
 #### 3. Classify the meeting type
 
-Title pattern → type → coaching focus:
-[references/meeting-types.md](references/meeting-types.md). The type drives
-Phase 2 context loading and Phase 3 dimensions.
+[references/meeting-types.md](references/meeting-types.md) maps the title
+pattern → type → coaching focus. The type drives the context you load in
+Phase 2 and the dimensions you track in Phase 3.
 
 ### Phase 2 — Gather context (once)
 
 #### 4. Resolve attendees
 
-Extract names from title and participant list. For each:
+Extract the names from the title and the participant list. For each name:
 
 ```bash
-rg -l "{name}" Knowledge/People/
-rg -l "{name}" Knowledge/Candidates/
+rg -l "{name}" 3-Team/People/
+rg -l "{name}" 2-Confidential/Candidates/
 ```
 
-Read matching notes for role, organization, history, open items, prior
+Read each note you find for role, organization, history, open items, and prior
 interactions.
 
 #### 5. Load type-specific context
 
-**Interviews:** read `Knowledge/Candidates/{Name}/{brief,screening,panel}.md`,
-look up the `Req` field's matching `Knowledge/Roles/*.md` file (check the
-`**Status:**` field for context), and load standard expectations:
+**Interviews:** read
+`2-Confidential/Candidates/{Name}/{brief,screening,panel}.md`. Look up the `Req`
+field's matching `2-Confidential/Roles/*.md` file. Check its `**Status:**` field
+for context. Then load the standard expectations:
 
 ```bash
 bunx fit-pathway job {discipline} {level} --track={track}
 ```
 
 **General meetings:** read attendee People notes plus referenced
-Project/Organization notes. Check open tasks: `rg "{name}" Knowledge/Tasks/`.
+Project/Organization notes. Check open tasks: `rg "{name}" 3-Team/Tasks/`.
 
 #### 6. Build the coaching brief
 
-Synthesize gathered context into the pre-meeting brief format in
-[references/coaching.md](references/coaching.md#pre-meeting-brief-format) and
-print it to the user.
+Synthesize the gathered context into the pre-meeting brief format in
+[references/coaching.md](references/coaching.md#pre-meeting-brief-format). Print
+it to the user.
 
 ### Phase 3 — Follow loop
 
 #### 7. Read new transcript content
 
-First read (no `--after`):
+MCP: `get_meeting_transcript({ meeting_id, offset, limit: 500 })`. CLI
+fallback:
 
 ```bash
-node .claude/skills/anarlog-follow/scripts/follow.mjs <session-id>
+anarlog --json meetings transcript <meeting-id> --offset <offset> --limit 500
 ```
 
-Subsequent reads (pass the last word ID):
-
-```bash
-node .claude/skills/anarlog-follow/scripts/follow.mjs <session-id> --after <last-word-id>
-```
+First read: `offset` = 0. Each later read: `offset` = the `pagination.total`
+from the previous read. That fetches only the words added since then. When a
+page comes back full and `next_offset < pagination.total`, keep paging with
+`next_offset` until you catch up. Track `total` across reads. It is your next
+offset.
 
-Returns JSON with grouped text segments, channel labels, and the next last-word
-ID.
+`data.words` is a flat array of `{ channel, text, start_ms, end_ms }`. Group
+consecutive words into segments yourself. Start a new segment on a channel
+change or a >3s gap between one word's `end_ms` and the next word's
+`start_ms`. Channel `0` = user. Channel `1`+ = guest(s) (a heuristic, per
+Anarlog's own docs).
 
 #### 8. Analyze new content
 
-Track the dimensions in
-[references/coaching.md](references/coaching.md#dimensions-to-track) — interview
-vs. general — and apply the [principles](references/coaching.md#principles) and
+Track the interview dimensions or the general dimensions in
+[references/coaching.md](references/coaching.md#dimensions-to-track). Apply the
+[principles](references/coaching.md#principles) and the
 [constraints](references/coaching.md#constraints).
 
 #### 9. Provide coaching
 
-Output only when actionable. Use the
+Output only when the nudge is actionable. Use the
 [coaching output formats](references/coaching.md#coaching-output-formats). Keep
 each nudge 1–3 lines.
 
 #### 10. Detect meeting end
 
-Watch for: farewells ("bye", "take care", "have a good day"), wrap-up phrases
-("that's all", "let's wrap", "I'll let you go"), or final thank-yous with no
-substantive follow-up. When detected, move to Phase 4.
+Watch for farewells ("bye", "take care", "have a good day"). Watch for wrap-up
+phrases ("that's all", "let's wrap", "I'll let you go"). Watch for final
+thank-yous with no substantive follow-up. When you detect one, move to Phase 4.
 
 #### 11. Loop cadence
 
 **Do not use `sleep` or timed loops.** After each coaching output, read the next
-batch immediately — natural cadence emerges from read/analyze/output. On two
-consecutive empty reads, ask the user whether the meeting has ended.
+batch immediately. A natural cadence emerges from read/analyze/output. After two
+consecutive empty reads, ask the user whether the meeting ended.
 
 ### Phase 4 — Wrap-up
 
 #### 12–13. Debrief
 
-Use the appropriate template in [references/debrief.md](references/debrief.md) —
-interview vs. general.
+Pick the interview template or the general template in
+[references/debrief.md](references/debrief.md).
 
 #### 14. Offer next steps
 

package/templates/.claude/skills/anarlog-follow/references/coaching.md

@@ -45,19 +45,19 @@ Suggested topics:
 
 | Dimension               | What to watch for                                                                  |
 | ----------------------- | ---------------------------------------------------------------------------------- |
-| **Skills demonstrated** | Candidate shows (not just claims) standard skills                                  |
+| **Skills demonstrated** | Candidate shows standard skills. A claim alone does not count.                     |
 | **Behaviour signals**   | Own the Outcome, Think in Systems, Communicate with Precision, Polymath, Curiosity |
 | **Level signals**       | Autonomy, scope, complexity handling, mentoring                                    |
-| **Coverage gaps**       | Focus areas from screening that haven't been touched yet                           |
+| **Coverage gaps**       | Focus areas from screening that you did not touch yet                              |
 | **Red flags**           | Contradictions with CV, vague answers, deflection                                  |
-| **Time awareness**      | Are you running out of time with uncovered areas?                                  |
+| **Time awareness**      | Time runs out while areas stay uncovered                                           |
 
 ### General meetings
 
 | Dimension             | What to watch for                                    |
 | --------------------- | ---------------------------------------------------- |
-| **Open items**        | Were pending items raised and addressed?             |
-| **Decisions**         | What was decided? Who owns next steps?               |
+| **Open items**        | Someone raised and addressed the pending items       |
+| **Decisions**         | The decisions made, and who owns the next steps      |
 | **New commitments**   | Track who committed to what                          |
 | **Unasked questions** | Topics the user should raise before the meeting ends |
 | **Tone/dynamics**     | Shifts in energy, resistance, enthusiasm             |
@@ -81,30 +81,31 @@ Capture: decision made on {topic} — {who} will {what}
 
 - **Standard-grounded.** Every nudge ties to standard skills/behaviours. Don't
   coach on vibes.
-- **Evidence over claims.** Flag _claims_ vs. _demonstrations_; push for
+- **Evidence over claims.** Flag _claims_ vs. _demonstrations_. Push for
   specific examples.
-- **Coverage-aware.** Track which screening focus areas have been explored and
-  nudge before time runs out.
-- **Level-calibrated.** Know what the target level expects; flag
+- **Coverage-aware.** Track which screening focus areas you explored. Nudge
+  before time runs out.
+- **Level-calibrated.** Know what the target level expects. Flag
   below-/above-level answers.
-- **Don't lead.** Suggest open-ended probes, never loaded or leading questions.
+- **Don't lead.** Suggest open-ended probes. Never suggest a loaded or leading
+  question.
 
 ### General meetings
 
 - **Agenda-aware.** Track coverage of known topics or open items.
 - **Decision-focused.** Flag decisions and capture owners.
-- **Commitment-tracking.** Note who said they'd do what; surface unresolved
-  commitments from prior meetings.
-- **Time-conscious.** Nudge when the meeting is running long and key topics
-  aren't raised.
+- **Commitment-tracking.** Note who said they would do what. Surface unresolved
+  commitments from earlier meetings.
+- **Time-conscious.** Nudge when the meeting runs long and nobody raises the key
+  topics.
 
 ### Universal
 
-- **Be quiet when things are going well.** Speak only when actionable.
-- **Channel awareness.** Channel 0 = user; channel 1+ = guest(s). Reference what
-  the _guest_ said — the user already knows what they said.
+- **Be quiet when things go well.** Speak only when actionable.
+- **Channel awareness.** Channel 0 = user. Channel 1+ = guest(s). Reference what
+  the _guest_ said. The user already knows what they said.
 - **Never interrupt flow.** Hold nudges for the next natural pause.
-- **Respect the user.** Coach at the margin, don't patronise.
+- **Respect the user.** Coach at the margin. Do not patronise.
 
 ## Constraints
 
@@ -113,6 +114,6 @@ Capture: decision made on {topic} — {who} will {what}
   `req-assess`, or `anarlog-process`. Offer them after.
 - **Channel 0 = user.** Always.
 - **Transcription noise.** Live transcripts are messy. Don't coach from a single
-  ambiguous word — wait for context.
-- **Privacy.** Coaching may appear on a shared screen — keep nudges generic
-  enough that they don't reveal confidential assessment detail.
+  ambiguous word. Wait for context.
+- **Privacy.** Coaching may appear on a shared screen. Keep nudges generic. They
+  must not reveal confidential assessment detail.

package/templates/.claude/skills/anarlog-follow/references/sessions.md

@@ -0,0 +1,52 @@
+# Anarlog Meeting Data
+
+Reference for `anarlog-follow` Phases 1 and 3. Anarlog stores meetings in a
+local SQLite database (`~/Library/Application Support/anarlog/app.db`). Read
+them through Anarlog's typed, read-only interfaces. Its own `AGENTS.md`
+(`~/Library/Application Support/anarlog/AGENTS.md`) is explicit:
+
+> Use Anarlog's typed, read-only interfaces for meeting data. Do not use
+> `find`, `grep`, `rg`, filesystem crawling, or direct SQLite queries.
+
+**Prefer the MCP tools** when the `anarlog` MCP server is connected. **Fall
+back to the CLI** (`anarlog --json ...`) otherwise. Never guess a meeting ID.
+Phase 1 resolves it from the freshest `audio.mp3` on disk (the one signal
+neither interface exposes). Every other read goes through one of the two
+interfaces below.
+
+## `get_meeting` / `meetings get <id>`
+
+MCP: `get_meeting({ meeting_id })`. CLI:
+`anarlog --json meetings get <meeting-id>`.
+
+Returns `title`, `created_at`, `participants` (`{ human_id, display_name,
+email, role, job_title, organization_name }`, often sparse), `note`,
+`summaries`, `action_items`. For a live meeting, `note` and `summaries` are
+usually thin or empty. That is expected. They fill in once the meeting ends.
+Use this once per session, in Phase 1 Step 2, to capture the pre-meeting brief
+inputs.
+
+## `get_meeting_transcript` / `meetings transcript <id>`
+
+MCP: `get_meeting_transcript({ meeting_id, offset, limit })`. CLI:
+`anarlog --json meetings transcript <meeting-id> --offset <n> --limit <n>`
+(`limit` capped at 500 by both interfaces).
+
+```json
+{ "meeting_id": "…", "text": "Yeah, so the overall …",
+  "words": [ { "channel": 0, "text": " Yeah,", "start_ms": 0, "end_ms": 400 } ] }
+```
+
+`pagination.total` is the running word count. Use it as the next poll's
+`offset` so you fetch only the words added since the last read. Page with
+`next_offset` when a single poll returns a full page and more remain. Channel
+`0` = user, channel `1`+ = guest(s) — a heuristic, not guaranteed.
+
+## The one filesystem exception: liveness detection
+
+Neither interface reports which meeting is currently recording. The only
+signal for that lives on disk: the most recently written `audio.mp3` under
+`~/Library/Application Support/anarlog/sessions/{uuid}/`, where the directory
+name is the meeting ID. This skill touches the filesystem for exactly that one
+lookup (Phase 1 Step 1) and nothing else. Every read of meeting *content*
+goes through `get_meeting` / `get_meeting_transcript`.