gmassist-precheck - SKILL.md Agent Skill

name: gmassist-precheck description: Pre-extraction pass — run enhance_summary then consistency_check against a gm-assist + VTT, before any per-scene VTT extraction. Use when the user invokes /gmassist-precheck [session-dir]. Catches canon and scope problems on the human-authored scene structure before spending tokens on scene_extract. tools: Bash, Read, Edit

gm-assist precheck

Run the cheap pre-extraction pass: take a session's gm-assist + VTT, produce the enriched session-summary.md via enhance_summary.py, then run check_consistency.py against it. The point is to surface canon contradictions (NPCs that don't exist, world-state conflicts, named beats that don't fit) while the artifact is still a small structural document — before the expensive per-scene VTT batch in scene_extract.py.

This is an experimental pass. The user is trying it a few times to decide if it's worth keeping. Keep the workflow tight and don't over-elaborate.

Workflow

1. Locate the session directory

If the user passed a path argument, use it. Otherwise:

Run pwd to confirm CWD is a campaign workspace (contains docs/, summaries/, config.yaml).
List recent session directories: ls -t summaries/ | head -10
Ask: "Which session — pass the path under summaries/ (e.g. summaries/20260512)?"

The session dir should contain gm-assist.md and a .vtt transcript. Confirm both exist before continuing.

2. Run enhance_summary (with built-in pre-flight)

python ~/src/CampaignGenerator-unified-pipeline/enhance_summary.py \
  <session-dir>/<vtt-file>.vtt \
  --gmassist <session-dir>/gm-assist.md \
  --output <session-dir>/session-summary.md \
  --party docs/party.md \
  --gm-player <gm-name> \
  --batch

Pass --party docs/party.md and --gm-player <name> whenever they're available — they enable the wrong-VTT pre-flight inside enhance_summary.py itself (aborts with exit 2 if no VTT line starts with any expected display name, before submitting). The GM display name is whatever they appear as in Zoom (e.g. Kostadis).

--batch is the default for cost reasons (50% off list price). Block-and-poll mode is fine — the call is one cached request, so it returns quickly. If the user explicitly wants the live streaming path, drop --batch.

If session-summary.md already exists, ask the user whether to overwrite or skip the enrichment step (and just re-run the consistency check on the existing file). Don't silently overwrite — the existing file may have human edits.

If the pre-flight aborts (exit 2 with speaker-mismatch), the VTT is almost certainly the wrong file. Stop and report it to the user — don't pass --allow-speaker-mismatch to "make it work."

3. Run the consistency check

Delegate to the existing /consistency-check skill rather than reimplementing. Tell the user:

Enrichment done. Running /consistency-check <session-dir>/session-summary.md next.

Then invoke the consistency-check skill workflow on that file. That skill handles the session-prep step (REQUIRED — do not skip), the --context docs/party.md wiring, and the report presentation. Don't shortcut its session-prep prompt; the value of this pass collapses without prep.

4. Present findings and pause

After consistency-check returns:

Surface the issue count and the report.
Ask: "Apply any of these fixes to session-summary.md before scene_extract?"
If yes, edit session-summary.md directly using the Edit tool, citing each Suggested fix from the report.

Do NOT auto-advance to scene_extract.py. The point of this pass is the human checkpoint between enrichment and extraction. End with:

Ready for scene_extract.py when you are.

Notes

This skill exists to evaluate whether the gm-assist → enhance_summary → consistency_check chain is a worthwhile pre-extraction pass. Each run is also a data point — note (briefly, in the chat) anything that felt high-value or wasted in the run, so the user can decide whether to formalise it.
The consistency check is operating on a Stage-1 artifact (enriched summary), not a Stage-2 artifact (per-scene extractions) or final narration. False positives are possible — flagged issues at this stage are scope/canon hints, not transcript errors. The skill should frame findings that way.
If enhance_summary.py is unavailable in the unified-pipeline tree, fall back to ~/src/CampaignGenerator/enhance_summary.py. Both checkouts are kept in sync; either is fine.
Both enhance_summary.py and scene_extract.py share the same wrong-VTT pre-flight (--party, --gm-player, --allow-speaker-mismatch). They share the same root motivation: don't spend tokens enriching/extracting against the wrong recording. Always pass --party and --gm-player when running this skill — that's where the protection lives.