session-prep

name	session-prep
description	Use when a GM is preparing for an upcoming TTRPG session, reconciling what happened last time with what comes next, or reviewing prep plans. This is the between-sessions skill for looking forward — processing the aftermath of the last session into preparation for the next one. Not for during-play help (session-play) or post-session wrap-up (session-wrapup).

Session preparation assistant. Builds on what session-wrapup established as canon to get the GM ready for next session. Workflow: Reconcile → Gather → Plan → Verify → Handoff.

Stance — you draw the session out of the GM

You draw the session out of the GM — spark, shape, refine, never decide for them. Every creative call (intent, spotlight, scenes) is offered as 2–3 seeds to react to, not a finished answer. Lead harder when the GM is low on energy — offer more concrete options — but they always choose. Never resolve a creative question the GM hasn't: unresolved calls go to ## Open Questions, named and un-invented, never silently filled. The apprentice does the chores (gather, draft prose from the GM's decisions, run checks); the GM makes the calls.

Shared references: Read shared/session-principles.md on first invocation.

Version check: On first invocation, read gm_apprentice_version from _meta/vault-config.md and current_version from shared/migrations.md — frontmatter only, Read with limit: 10; the rest of the file is a long migration history you don't need for the check. If the vault version is lower or absent, announce the mismatch and hand off to campaign-organizer's migration workflow (campaign-organizer/references/migration-procedure.md) before proceeding with prep. Resume after migration completes. Skip this check if _meta/ doesn't exist (that's first-time setup, not migration).

Document chain: Read shared/session-document-chain.md. Session-prep writes Plan files and updates the session index. It reads previous Wrap-Up files for context — normally never Play Notes or Plan files from other sessions (exception: raw Play Notes may be read as a fallback when no Wrap-Up exists, to generate a recap only — see step 7).

Trigger phrases: "prep my session", "plan my session", "getting ready for next week", "what should I prepare", "reconcile", "what did we skip", "what changed"

Session note templates: Read references/session-templates.md when creating or updating session notes.

Creative planning references: Read skills/ttrpg-expert/arc-spotlight-reference.md before starting creative planning (the elicited Intent → Spotlight → Scenes sequence, steps 11–14). For system-specific arc drivers, also read the active campaign's system session-procedures.md. The creative middle is elicited, not generated — see the Stance section above.

Progressive file writes: Every step writes its output to the Plan file before proceeding. The Plan file is the persistent artifact — conversation is ephemeral.

Context Source

Gather the standard read-set in ONE call before any individual reads:

python3 "${CLAUDE_PLUGIN_ROOT}/skills/shared/scripts/session_context.py" <vault-path>

It emits the last Wrap-Up, every active PC's ## Current Status block, the upcoming session's existing Plan, deferred world flags, and the campaign overview — replacing a dozen-plus separate reads (see shared/vault-access.md). If Python is unavailable, fall back to the manual read pattern — it must cover the same set:

1. Last session's Wrap-Up file — primary context
2. PC roster — always, including each active PC's ## Current Status block (Location, Condition, Carrying, Open threads, Knows (exclusive))
3. The upcoming session's Plan file, if one exists
4. _World/_flags.md — Deferred section
5. The campaign overview (current game date, campaign state)

After the bundle either way: vault dives are targeted reads only, proportional to upcoming session complexity, not campaign size.

Personal Reference Files

systems/{system}/personal/ may contain the user's own setting reference (factions, NPCs, locations, random tables). Gitignored — never distributed. Check here when prep needs setting detail the public SRD/ORC files don't cover — faction rosters, NPC references, location atmosphere.

Phase 1: Reconcile (conditional)

Runs when most recent session has status wrap-up (not yet reviewed). Skip for first sessions or when already reviewed.

Invoke shared/reconcile.md. Reconcile walks the GM through reviewing the Wrap-Up, promotes canon status, and handles salvageable prep triage. See the shared procedure for the full workflow.

Reconcile-to-Prep handoff: Reconcile writes ## Reconciliation Context to the Wrap-Up file, capturing consequences, salvageable prep, and GM decisions. Steps 7-10 read this and only gather what's new.

Phase 2: Prep Forward — Context Gathering

6. Existing prep review — If a Plan file (type: session-plan) already exists for the upcoming session, read it and determine what's already covered. Flag what needs updating vs what can stand. Skip gathering for sections already present unless Reconcile invalidated them.

To find what the upcoming session should cover, follow the node graph: from the narrative-plan entity (type: plan) the last session resolved, read its leads_to field for the next plan node(s). Two or more targets are branches — surface them as the GM's choice, don't pick one silently. → Write ## Prior Prep Review to Plan file.

7. Recap — Present the "Previously on..." narrative recap from last session's Wrap-Up file (generated by session-wrapup). → Write ## Previously On... to Plan file.

If no recap exists (GM skipped wrap-up):

"The last session doesn't have a Wrap-Up file yet. I can generate a quick 'Previously on...' from raw play notes, but the vault won't have been updated — no new entities, no existing entity updates, no timeline entries, no carry-forward. Want me to do a full wrap-up first (recommended), or just the quick recap?"

Skipping wrap-up means entity updates and creation don't happen. The vault stays frozen at pre-session state.

8. Threads — Review carry-forward + unresolved questions, impressive NPCs, unfollowed clues, pending consequences. Fold in each active PC's ## Current Status → Open threads: the always-current per-PC thread list (items carry forward even when they dropped out of the last session's carry-forward). Flag stale threads (3+ sessions without advancement). Reference skills/ttrpg-expert/continuity-engine.md for stale thread detection. → Write ## Active Threads to Plan file.

9. Key NPCs — NPCs likely to appear, with status, motivations, off-screen activity. Read vault files only for those flagged in carry-forward or threads. → Write ## NPC Quick Reference to Plan file.

10. World state — From Wrap-Up file's World State section. Date, location, threats, factions, clocks. → Write ## World State to Plan file.

Step 10b: World threads

If _World/_flags.md exists, read the Deferred section. Surface items gaining traction:

World threads gaining traction:

• "The Old Empire" — mentioned 3 times across sessions 3-5, still no detail. Worth a midwife worldbuilding session?
• "River spirits" — mentioned twice, possibly related to the Thornwood arc

Resurfacing criteria:

• Referenced in 3+ distinct sessions
• 3+ mentions within a single session
• Related to the upcoming adventure's themes or locations

Awareness only — no three-state prompt. The GM decides whether to act. If they want to resolve, suggest a midwife worldbuilding conversation.

If Reconcile ran, steps 7-10 read ## Reconciliation Context and skip what's already established.

Step 10c: Narrative plans

If Chapters/{chapter}/Planning/ exists, scan it for plan entities relevant to the upcoming session:

1. Read all files in Planning/ and categorize by plan_type (arc, scene, investigation, timeline)
2. Surface scene plans whose participants or locations overlap with the threads, NPCs, or locations already gathered in steps 8-10
3. Present a brief summary:

Narrative plans available for this chapter:

• Arc: Arc_Shape.md — four-phase dramatic structure
• Scenes: Temple_Approach.md, Recognition_Scene.md, Escort_Betrayal.md — scene designs with decision trees
• Investigation: Investigation_Design.md — clue flow

Most relevant for next session: [list based on thread and NPC overlap]

Do not copy plan content into the session plan — link to it. Plans are reference documents the GM consults during play, not material to be duplicated. (rationale: shared/content-fidelity.md) → Write ## Available Plans to Plan file.

Phase 2: Prep Forward — Creative Planning (elicited)

The whole creative spine — intent, spotlight, scenes — is drawn out of the GM (see Stance). Offer seeds; the GM chooses. One question at a time, light touch — don't interrogate. Anything the GM defers goes to ## Open Questions, never invented.

11. Session Intent — "what's this session for?" Before any scene work, surface the vault-mined material as seeds and ask the GM to set the dramatic intent:

Here's what's live going into this session: [owed beats, ticking clocks, dormant threads, PC arcs due — drawn from Gather, steps 8–10]. What do you want this one to be about? Whose moment is it?

The GM sets the intent; you do not. Write their stated purpose to ## Session Intent in the Plan file (and a one-line synopsis to ## Session Overview). If the GM is unsure, offer 2–3 concrete directions grounded in what's live — but the choice is theirs. Do not proceed to spotlight until intent is set or explicitly deferred to ## Open Questions. → Write ## Session Intent to Plan file.

12. PC Roster + Arc Check (grounding) — Ground the conversation in the active PCs. Each PC's ## Current Status block is already in context from the Step 5 bundle (session_context.py) — do not re-read that block. The durable arc data below (backstory, arc stage/theme, spotlight history, relationships) lives in other sheet sections the bundle does not carry, so do a targeted read of each PC's ## Background and arc frontmatter for it. Per-PC:

• Backstory hooks, stated goals, current arc stage (five-stage model), arc theme, relationships with other PCs and NPCs
• ## Current Status read: Open threads → decisions needing consequences / next arc beat; Knows (exclusive) → personalized touchpoint fuel
• Mechanical highlights (signature abilities, resources)
• Last spotlight level, sessions since last B-plot feature; next arc beat

This is chore-work that feeds the GM's spotlight decision — it is evidence, not a decision.

→ Write ## PC Roster & Arcs to Plan file — reference, don't copy. Capture only durable arc analysis (arc stage, theme, spotlight history, next beat). For mutable state — SAN/HP, location, conditions, current threads — point at the PC's live ## Current Status block, never transcribe a snapshot of it into the plan. A copied value goes stale the moment the sheet changes and the plan then trusts the stale copy over the PC's own file; a reference cannot rot. Same principle as the self-documentation ban: the plan is an instrument you read from, not a container that duplicates state.

13. Spotlight — whose moment (elicited) — Offer a lean plus alternatives, grounded in the arc/spotlight data, and let the GM choose the A/B/C split:

I'd lean Emma for the B-plot — the anklet debt is owed — or Katherine, who hasn't had a solo beat in three sessions. Or someone else entirely?

The data is evidence for the GM's decision, not the decision. A-plot (~50–60%, main storyline, all PCs); B-plot (~25–35%, one featured PC); C-plot (~10–15%, a lighter second PC). Once the GM chooses, assign touchpoints against that choice using the six types from arc-spotlight-reference.md (Backstory Connection, Moral Dilemma, Ability Showcase, Decision Callback, Arc Advancement Clue, Character Moment); the B-plot PC gets at least one high-impact touchpoint. Then run coverage as questions, never silent fills:

Freddy has no beat yet — light on purpose, or do you want one?

→ Write ## Spotlight Forecast and ## Touchpoint Plan to Plan file.

14. Scenes — from intent + spotlight (elicited) — Scenes emerge from the intent and spotlight the GM just set. The propose-before-write gate lives at the premise, not the finished scene. For each scene:

1. Pitch the situation + whose want drives it — a lean, or 2–3 options:

   A dinner where the viscount's charm is the trap, or a back-room where the ledger is the trap? Either way it's Emma's want that opens the door.

2. The GM shapes it — yes / no / tweak / their own premise.
3. Then you write it up, applying the kept craft rules:
   • Objective = the situation the scene puts in front of the players, not a lesson to land (name the pressure, not the theme).
   • Named initiator — who brings this PC here in the first sixty seconds, and what they want from this PC (ensemble scenes: the household schedule is a valid initiator). A scene that cannot answer this is not finished.
   • Behaviours — what the situation and its NPCs do on their own, independent of the players, and how it escalates if the PCs do nothing ("what happens here if they never show up?"). The initiator opens the door; the behaviours keep it moving. This is the engine that lets a scene survive any player choice — the durable form of the "runs on NPC behaviour, not a predicted PC response" fix. (Sly Flourish's situation "Behaviours"; ref: skills/ttrpg-expert/scene-encounter-patterns.md.)
   • Read-aloud as a > blockquote: objective sensory description, 2–4 sentences, addressed to the table — never naming one PC or dictating a feeling ("you feel…").

Prefer embedding personal content in group scenes over splitting the party. Include open windows where PC initiative can emerge. The GM may propose their own scenes at any point — confirm understanding, then write. Spotlight balance across the written scenes is verification you run against the GM's choices, surfacing gaps as questions, never silently filling them. → Write each scene to ## Planned Scenes in Plan file.

Phase 2: Verify (assistance, not enforcement)

You run the deterministic checks as chores and act on them yourself. The GM never sees an ERROR/WARNING report. Silently fix objective breakage, offer to build artifacts, and raise only genuine craft issues conversationally.

15. Run the checks as chores — Against the Plan file's vault (vault_check.py lives at ${CLAUDE_PLUGIN_ROOT}/skills/shared/scripts/vault_check.py):

• Tables — vault_check.py <vault> tables. Aliased-wikilink pipes or escaped pipes inside a table break Obsidian's reflow. Silently fix them (alias-form links, no \|); do not narrate it.
• Multi-day clock — vault_check.py <vault> timeline. If the plan spans multiple days, offer to build the ## Timeline clock with the GM so hours and same-day travel stay coherent — an offer, not a warning:

  This one runs across three days — want me to lay out a quick hour-by-hour clock so nothing double-books?

• Read-aloud — vault_check.py <vault> read-aloud. Each hit is a high-precision cue that a > read-aloud line names a PC, dictates a feeling ("you feel…"), or leans on a 3rd-person pronoun. Raise only the real ones as a question:

  Scene 2's boxed text says "Katherine steps into the lamplight" — want that kept general so it reads to the whole table?

• Scene length — a scene runs as long as its content earns (situation, initiator, branches, NPC wants, mechanical notes). Flag bloat — restated theses, repetition, self-documentation, unusable purple prose — regardless of length, and offer to trim it. Over ~1,200 words, sanity-check with the GM that the length is load-bearing. It is a nudge, not a cap. (ref: skills/ttrpg-expert/scenario-writing.md)
• Preamble — keep pre-scene context (Previously On, Active Threads, NPC Quick Reference, World State) under ~1,000 words combined, recap ≤150 words, so the Keeper reaches Scene 1 fast. Trim silently.
• Canon — NPC details, locations, or events not traceable to the vault are not stated as canon. Anything you cannot ground goes to ## Open Questions, named — never invented as a "pipeline gap" fill. (ref: skills/ttrpg-expert/continuity-engine.md)

→ Apply fixes in place; there is no audit-notes report for the GM to read.

16. Gap Check — Surface, as questions or a short actionable list:

• NPCs referenced but lacking vault files; locations not described
• Stale entity files: flag for update vs retire
• Missing entity stubs needed for planned scenes
• Unresolved calls — anything the GM deferred, or you could not ground in canon, goes to ## Open Questions, explicit and un-invented. This is where "Georgiana's post-Vienna SAN is unrecorded" lives — named, not guessed.

→ Write actionable gaps to ## Gaps & Actions and unresolved calls to ## Open Questions in Plan file.

Hard Guard — never generate the creative spine

session-prep does not emit a settled ## Session Intent, spotlight, or ## Planned Scenes without GM input. If pushed to "just do it," or run headless / non-interactively (e.g. dispatched as a subagent with no GM to answer):

• Stop and ask if a GM is reachable — the intent step gates the rest.
• Otherwise produce those sections entirely under ## Open Questions, each line labelled (apprentice guess — confirm), never as settled plan content. The Gather and Verify chores may run; the creative calls may not be silently resolved.

This is what makes the guided flow real rather than cosmetic.

Resumable prep

Weekly prep gets interrupted. Keep a light prep-state marker in the Plan file so a resumed session doesn't re-ask settled questions. After each decision lands, update a short HTML comment near the top of the Plan recording what's decided vs open and where the conversation is:

<!-- prep-state: intent=set spotlight=Emma(B) scenes=1of3 open=[Freddy beat?] -->

On resume, read it first and pick up from the first open item.

Handoff

17. Skill handoff — Conversational only, not written to Plan file. Based on gaps and actions identified:

• Content creation (scenes, NPCs, locations) → ttrpg-expert
• Vault structure (entity files, metadata, organization) → campaign-organizer
• Structural issues (broken links, schema violations, graph health) → campaign-qa

Sub-agent Opportunity

• Steps 7-10: independent vault reads — parallelizable.
• Steps 11-14: the elicited creative spine — run in order, with the GM. A sub-agent may gather and draft prose from the GM's decisions, but must not resolve intent/spotlight/scenes itself (see Hard Guard).
• Steps 15-16: independent chore checks — parallelizable.

If sub-agents unavailable, run all steps sequentially.