mycelium-hq/ai-brain-starter/coaching

Coaching Session — Multi-Pass Panel + Accountability Tracking

A skill that turns a one-off hard moment into a tracked accountability arc. Runs panel passes that update with the user's corrections, files a synthesized record with a re-eval date, and updates the vault's rolling pattern tracker so growth can be measured over time.

Why this exists

Daily journals capture one moment per day. Panel reactions inside /journal are one-shot. Most real coaching, real therapy, real advisor relationships are NOT one-shot — they're multi-turn, they update when new evidence comes in, and they track whether the same blind spot keeps surfacing across months.

The vault has all the pieces (panel rules in ⚙️ Meta/rules/advisory-panel.md, daily journals, decision logs) but no skill that orchestrates the multi-pass coaching arc and files it for tracking. This skill fills that gap.

Three-tier architecture

This skill produces files at three different timescales:

1. Verbatim raw (immediate) — 📋 Strategy/Coaching Sessions/Processing Notes - YYYY-MM-DD - <topic>.md. The user's exact words during the session. Per the "save exact words" rule, no annotation, no synthesis. Available for re-read forever.

1. Synthesized accountability record (per session) — 🏠 Home/Coaching Sessions/YYYY-MM-DD - <topic>.md. What surfaced, commitments named, re-eval date one month out. This is the file /weekly and /monthly look at to ask "did the pattern repeat? did the commitments land?"

1. Rolling pattern aggregator (across sessions) — 🏠 Home/Panel Feedback Log.md. Patterns table at the top tracks mention counts. Single mention = watch. 2+ mentions across different contexts = promote to acute action item. The aggregator is what tells you "this is a real recurring pattern" vs "this was a one-off."

All three tiers honor `__SKIP`. Coaching sessions are exactly where someone says something out loud to think about it, not to record it. Before writing any of the three files, strip every line whose first token is __SKIP, then confirm one line per dropped item (Dropped __SKIP line N (token preview: ...)). Never paraphrase the dropped content into the synthesized record — tier 2 summarizing what tier 1 was told not to keep is the exact leak the convention exists to stop. Enforced by block-skip-prefix-in-vault-write.py; full rule at templates/rules/skip-prefix-convention.md.

When to use vs. /journal

Use `/coaching` when:

• The triggering event won't fit in a daily journal entry (a 2-hour call, a multi-day arc with one person, a decision that needs panel feedback over multiple iterations)
• The user wants the panel to UPDATE its takes when they provide corrections (not just hear them once)
• The user wants this tracked across future weeks/months, not just logged today
• Stakes are high enough that a re-eval one month out matters

Stay in `/journal` when:

• Daily check-in
• One-off reflection
• Floor work for the day
• Single-pass panel reaction is enough

How it runs

Step 1: Triggering event

Ask the user, in their language:

• What's the situation you want the panel to look at?
• When did it happen (date or rough range)?
• Who was involved?
• Is there a transcript / journal entry / message thread already in the vault, or does the user need to talk through it now?

If a transcript already exists in the vault, ask for the path. If not, the user talks through what happened. Either way, capture is the next step.

Step 2: Capture verbatim

Decide where the raw goes:

• Cofounder / business relational sync — 📋 Strategy/Co-founder Syncs/Processing Notes - YYYY-MM-DD - <topic>.md
• Decision review — 📋 Strategy/Decision Reviews/Processing Notes - YYYY-MM-DD - <topic>.md
• Personal relationship — 📋 Strategy/Personal Coaching/Processing Notes - YYYY-MM-DD - <topic>.md
• Generic — 📋 Strategy/Coaching Sessions/Processing Notes - YYYY-MM-DD - <topic>.md

If the user has a transcript already, link to it. If not, capture what they say in their exact words, organized by chronological order or topic. The verbatim file gets numbered sections (1, 2, 3...) with section headers naming the topic, and quoted bodies that are the user's exact words.

Critical rule (carry from the daily-journal skill): Verbatim section bodies are the user's voice only. Panel takes do NOT live in this file. Annotations, synthesis, and commentary live in the synthesized record (Step 6) or the aggregator (Step 7), never inline in the verbatim file.

Step 3: Initial panel pass

Read ⚙️ Meta/rules/advisory-panel.md. The Technology & AI, LatAm and Colombia sections only count for people they apply to. If About me/My panel.md exists, it decides: its ## Sections turn those sections on or off, the names under ## Removed are left out, and the names under ## Added are full panelists. Convene 3-5 voices most relevant to the triggering event. Apply the rules:

• 3-5 voices, each 1-2 sentences in their authentic voice
• At least one MUST dissent
• Cite sources for any factual claims
• Use only named panelists from the file, plus anyone under ## Added in About me/My panel.md (no archetypes, no invented experts)
• Subject-match weighting (rule 10): the panelist whose expertise matches the actual subject carries more weight than headcount
• Output-confidence filter (rule 11): flag any recommendation below 80% conviction as low-conviction (keep for texture, not for decisions)
• Peer-review pass (rule 9): after the 3-5 voices speak, run one chairman pass with the 5-bucket taxonomy — consensus (where they converged), contradictions (where they clashed), partial coverage (a dimension only one voice addressed), unique insights (a point only one voice raised), blind spots (what they all missed). The chairman analyzes, does not merge: surface the disagreement, never collapse it into a balanced-view paragraph
• Effort-aware sizing (rule 15): if ${CLAUDE_EFFORT} is set, scale the panel accordingly

Deliver the panel's read with a clear lead sentence answering the triggering question. Don't bury the answer.

Step 4: Wait for the user's response

The user will do one or more of:

• Confirm a take that landed
• Push back with a correction (new factual info, different context, a missed nuance)
• Add autobiographical context that recalibrates the panel's read
• Ask follow-up questions

This is the move that distinguishes coaching from journal. Whatever the user provides, the next pass UPDATES the takes transparently, not silently.

Step 5: Iterate panel passes (the corrections-update-takes loop)

For each new piece of info the user provides:

1. Acknowledge what changes from prior takes. Be explicit: "Pattern 3 (X) is now demoted because Y." Or: "Patricia Engel's earlier read assumed Z; the autobiographical context just shared inverts that."
2. Re-run the relevant subset of voices with the new info. Don't re-run the whole panel for small corrections; do re-run when the correction is substantive (changes the framing).
3. Note clashes between prior reads and current reads explicitly. The user should be able to see how their corrections moved the panel.

Keep iterating until the user signals they're ready to synthesize. Common signals: "OK that's enough," "let's wrap this up," "save this," or just shifting to operational mode ("let's track this").

Step 6: Synthesize and file the Coaching Session record

Create 🏠 Home/Coaching Sessions/YYYY-MM-DD - <topic>.md with this structure:

---
creationDate: YYYY-MM-DD
type: coaching-session
session_format: panel-with-claude
duration_approx: <one-pass | multi-pass | multi-pass-Nhr>
triggering_event: "<one-line description>"
themes: [theme1, theme2, ...]
panelists_seated:
  -<Name 1>
  -<Name 2>
  -...
related: ["[[verbatim file]]", "[[source transcript]]", "[[decision log]]"]   # quote EVERY wikilink individually
status: open
re_eval_date: YYYY-MM-DD (one month from session)
---
 
*Brief one-paragraph framing of what this session covered.*
 
## Triggering event
 
<2-3 sentences describing what triggered the session, with wikilinks to source material.>
 
## What surfaced (synthesized across all passes)
 
### Pattern 1: <Theme name>
 
**The observation**: <what the panel saw, with citations>
 
**User's correction (if any)**: <what they pushed back with, what shifted>
 
**Status**: <1 mention in tracker / demoted / promoted / contextual fact (not a watch pattern)>
 
### Pattern 2: <next theme>
...
 
## User's commitments coming out of the session
 
1.<Concrete behavior commitment, not vague>
2.<Another concrete behavior commitment>
...
 
## Codified outcomes (rules / files changed this session)
 
-<Any vault rules added>
-<Any files moved or scrubbed>
-<Any structural changes>
 
## Re-eval signals (check by <re_eval_date>)
 
-Did <Pattern 1> show up at least once in real time?
-Did the user act on <commitment 1>?
-Did <Pattern 2> surface in any other context, or stay <person/situation>-specific?
...
 
## Cross-references
 
-Verbatim source: <wikilink>
-Source transcripts (if any): <wikilinks>
-Decisions: <wikilinks>
-Pattern tracking: [[Panel Feedback Log]]

Step 7: Update the rolling Panel Feedback Log

Open 🏠 Home/Panel Feedback Log.md. Two updates:

1. Patterns table at the top — for each NEW pattern this session, add a row with:

• Pattern name
• Mention count (1, since this is the first)
• Action: "Watch for repeat. Re-eval YYYY-MM-DD."

For patterns that surfaced this session AND already existed in the table, increment the mention count. If count hits 2+ across different contexts, change the Action to a concrete commitment (promote to acute).

1. Synthetic Panel Reactions section — append a new entry:

### YYYY-MM-DD — <topic> coaching session
 
⚠️ **Synthetic panel reactions across one Claude session, NOT real investor feedback. Real-human content lives in "By Meeting" section above.**
 
**Context:** <one-paragraph framing>
**Panelists:** <list>
**Pass 1:** <what surfaced>
**Pass 2 (if multi-pass):** <how takes updated with user corrections>
...
**Convergence (across all passes):** <high-confidence signals>
**Clash:** <where panelists disagreed>
**Partial coverage:** <a dimension only one voice addressed; others silent>
**Unique insight:** <a point only one voice raised>
**Subject-match weighting:** <which panelist's voice carries most weight>
**Collective blind spot (chairman):** <what all panelists missed>
**User's commitments:** <numbered list>
**Re-eval:** YYYY-MM-DD (one month). <Re-eval signals.>
**Entry:** <wikilinks to coaching session record + verbatim>

Step 8: Confirm with the user

Tell the user, in plain language:

• Verbatim saved at <path>
• Synthesized record saved at <path>
• Patterns added to tracker: <list>
• Re-eval date: <date>

Then close. Do not pile on additional panel takes. The session is logged. The system now exists to remember.

Integration with other skills

• `/journal` — one-shot daily check-in with inline panel reaction. If the user starts a journal entry that turns into a multi-pass conversation, suggest switching to /coaching so the work gets tracked.
• `/weekly` and `/monthly` — natural surface for re-eval. The weekly review skill should read open Coaching Sessions, surface ones whose re_eval_date has passed, and ask "did the pattern repeat? did the commitments land?"
• `/patterns` — pattern detection across MANY sessions / journals. The /patterns skill reads the Panel Feedback Log Patterns table to confirm whether a pattern hits 2+ mentions in different contexts (the promote threshold).
• `/deconstruct` — if a coaching session surfaces a stakes: high decision, auto-offer /deconstruct for first-principles analysis before the user commits.

Failure modes to avoid