Pre Session Portrait Agent Skill

Pre-Session Portrait

Turn "help me with X" into a decision-grade brief before the session starts. The instrument asks the client open, voice-note-friendly questions across seven fixed lenses; the consultant (or an LLM) compresses each answer to one line, yielding a portrait that is iterable, compressible, and easy to visualize.

Design principle: rich in, compressed out. The client talks freely; compression happens after, not in their head.

The seven lenses

| # | Lens | Elicits | Compresses to |

|---|------|---------|---------------|

| 0 | ANCHOR | the topic — what the call is for (referent for every later "this") | the topic in one line |

| 1 | WHERE | what's been tried, where it stalls | current state in one line |

| 2 | HOW | cognitive style — fast/slow, visual/verbal, systems/stories | how they think |

| 3 | WHAT | live preoccupations, open loops | current focus |

| 4 | PROBLEM | the problem under the problem | the core job |

| 5 | IDEAL | concrete "solved" state (day/feeling, not tool) | desired outcome |

| 6 | TENSION | what holds them back / worries them | dominant anxiety |

| 7 | JTBD | Push · Pull · Habit · Anxiety · Trigger | switching forces |

Output schema

portrait:
  where: ""
  how: ""
  what: ""
  problem: ""
  ideal: ""
  tension: ""
  jtbd:
    push: ""
    pull: ""
    habit: ""
    anxiety: ""
    trigger: ""

How it visualizes

• 7-spoke radial / hexad map — one label per lens, the capture line as the value.
• JTBD 2×2 — Push+Pull (energy toward change) vs Habit+Anxiety (energy against). The gap = leverage.
• Iterable — re-run any lens next session; watch the capture line drift over time.

Workflow

• Gather context. Client name, consultant name, session date, and (if known) the topic. Pull prior history from vault/email/Fathom if available so the consultant-only prep notes are grounded.
• Fill the template. Copy assets/interview-prompt.md and substitute {{CONSULTANT}} (and topic if narrowing lens 4). Leave the seven lenses intact.
• Pick a delivery (ask the user):
• Raw text — paste the substituted prompt into a message; client runs it in any clean Claude/ChatGPT.
• Secret gist — gh gist create --desc "Pre-session portrait interview (for <name>)" interview-prompt.md. Share the gist link. Use the unpinned raw URL (/raw/<filename>) so edits propagate.
• Codex one-liner — see assets/codex-bootstrap.txt; fetches the raw gist URL and runs the interview interactively.
• Optional preview. Before sending, generate a synthetic filled-in version (answers simulated from known context) so the consultant judges the deliverable's shape. Mark it clearly as synthetic.
• After the session. Fold the returned portrait: YAML into the client's People/Session note; diff against any prior portrait to show movement.

Delivery notes

• Secret gist ≠ auth-private: anyone with the link can read it. Fine for a benign intake; don't put client PII in the gist itself.
• Codex: run interactive codex (not codex exec), and include the "do not write code / touch files — this is a conversation" guard so it stays in interview mode.
• Framing line to prepend when sending: *"Paste this into a fresh Claude or ChatGPT chat — it'll ask you 7 quick questions and give you a block to send back to me before our call."*

Call cockpit (interactive HTML)

Once a portrait is back, generate an interactive prep cockpit the consultant runs live during the session. Start from assets/cockpit-template.html — a self-contained, theme-aware single file (no external deps).

Tabs: Setup (structured stack/facts fields + a paste box for the portrait: block) · Framework (six-station pipeline with per-station AUTO/ASSIST/HUMAN + quality-gate inputs) · Questions (per-section bank, each with an autosaved answer field; add-your-own) · Decisions & Actions (dynamic add/delete rows; actions carry an owner) + a build/demo box and show-don't-tell cues · Agenda (accordion of time-blocks that expand into checkable sub-steps + per-block notes; a live timer auto-opens the current block and fills a progress bar) · Notes.

Key properties:

• Autosaved to localStorage, namespaced by the Client-name field — so multiple cockpit files opened from the same folder (same file:// origin) never clobber each other's data.
• Filled instances: copy the template and inject a const SEED = {fields, decisions, actions} object just before // init; a one-time guard (prep::<ns>::__seeded) writes the seed into the client's namespace on first load, then the consultant's edits persist. Use this to pre-populate a cockpit from a known portrait + prior-session facts.

Also generate a client-facing recap after the session (same visual language): what we covered, current→target pipeline, decisions, what we built live, their next steps (autosaved checkboxes + fields), tech notes. Deliver as a file or publish as an Artifact URL to share a link.

Assets

• assets/interview-prompt.md — the self-contained interviewer prompt (template).
• assets/intake-form.md — human-readable version with per-lens capture: fields, if the consultant prefers to interview live.
• assets/codex-bootstrap.txt — the Codex CLI one-liner template.
• assets/cockpit-template.html — the interactive prep cockpit (blank, reusable; autosaved + client-namespaced).