Run any Skill in Manus with one click

playtest-triage

Stars0

Forks0

UpdatedJune 14, 2026 at 13:10

Use when the user pastes a live Discord session log, transcript, or annotated complaints from a play session and wants a fix round. Diagnoses the failure chain from the logs, extracts player improvement-wishes, applies deterministic fixes, runs the test suite, commits per round, and updates progress.md (and an ADR when a real trade-off was decided).

Installation

Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.

Run Skill in Manus

Source

Pr0degie

Pr0degie/dungeonmaster

View GitHub Repository View Creator Repositories

Download

Run Skill in Manus

SKILL.md

readonly

name	playtest-triage
description	Use when the user pastes a live Discord session log, transcript, or annotated complaints from a play session and wants a fix round. Diagnoses the failure chain from the logs, extracts player improvement-wishes, applies deterministic fixes, runs the test suite, commits per round, and updates progress.md (and an ADR when a real trade-off was decided).

Playtest triage — turn a live session into a fix round

This is the project's signature loop. A real round was played, something felt off, and the user wants it fixed. The job is diagnosis first, deterministic fix second — never "the model got worse".

Golden rules that bind this work

#2 — dice rolling and resolution are code (dmbot/rules/engine.py + the active profile), never the LLM. A fix that lets the model invent results is wrong.
#3 — hard world state (HP, inventory, flags) is advanced by code, never written from LLM free text.
#4 — Bot A's user-ID filter in the sink stays. Never weaken feedback protection "for debugging".
#8 — anything the DM says stays German; code/logs/commits stay English.

Procedure

Read the evidence, not your priors. Sources, in order:
- the pasted log / annotations (the user's own complaints carry the most signal)
- logs/debug.log (opt-in verbose), logs/transcript.log (STT → narration → TTS)
- the live state under data/sessions/<id>/state.json (wounds, NPCs, scene, recap)
Classify each complaint → root cause. Past diagnoses to pattern-match against (progress.md decision log):
- D43 / ADR 018 — wrong channel → silent example-party fallback → echoed player line read aloud (a failure chain, not regression).
- D57 / ADR 027 — silent num_ctx truncation (8192 hardcoded) dropped persona + adventure from the prompt mid-session → "ignores the story", over-long answers, puppeting.
- D59 / ADR 028 — OCR/statblock junk leaking into the per-turn RAG block. Suspect the pipeline (sample rate, context budget, dedupe, channel) before the model.
Fix deterministically. Prefer a code guard over a prompt nudge. Prompt-only is allowed only when the failure is genuinely stylistic — and then note that a code guard is the fallback if nemo doesn't adhere.
Run the suite: uv run --with pytest python -m pytest. Add fixed-seed tests for any new deterministic logic.
Commit per round with a scoped imperative message (e.g. dmbot(memory): cumulative auto-recap so wrap-up can't drop the session start).
Record the round (end-of-session ritual): update progress.md ## Last session and ## Current focus, and rotate the previous ## Last session entry to docs/progress-archive.md so the live file stays lean (per session-ritual); write the next-numbered ADR if a real trade-off was weighed. Mark the result live-unverified — the next real session is the gate. Name the concrete thing the next round should confirm.

Notes

Every log the user pastes is pre-change: your fixes are unproven until the next live run. Don't claim "fixed", claim "fix landed, live-unverified, gate = X".
Cross-reference the playtest-tuning-loop memory and the relevant ADR before touching code in the area a complaint points at.

More from this repository

same repository

improve-architecture

Pr0degie/dungeonmaster

Find deepening opportunities across the codebase — refactors that turn shallow pass-through modules into deep ones, for testability and AI-navigability. Informed by architecture.md and the ADRs in docs/decisions/; never re-litigates a decided ADR. Use when Tobi wants to improve architecture, find refactoring opportunities, consolidate tightly-coupled modules, or make a part of the bot more testable. Whole-codebase altitude — not /simplify or /code-review (those are diff-scoped).

2026-06-150

to-prd

Pr0degie/dungeonmaster

Turn the current conversation context — ideally after a /grill-me session — into a PRD written to the repo. Do NOT interview; synthesize what's already been decided. Use when Tobi wants to turn a grilled-out plan into a written spec, or says "to-prd" / "schreib den Plan".

2026-06-150

grill-me

Pr0degie/dungeonmaster

Interview Tobi relentlessly about a plan or design decision until reaching shared understanding, resolving each branch of the decision tree one fork at a time. Use when he wants to stress-test a plan before building, asks to "get grilled", or says "grill me". For design/architecture forks — not for mechanical edits.

2026-06-150

tdd

Pr0degie/dungeonmaster

Use when changing the deterministic core (dmbot/rules/, engine.py, marker.py, any new pure function with verifiable behaviour) and you want the change driven by a failing test first. Enforces red-green-refactor against the repo's fixed-seed pytest setup: write one failing test, confirm red, implement minimal green, refactor. Not for prose/persona/prompt edits or RAG ingestion.

2026-06-150

session-ritual

Pr0degie/dungeonmaster

Use at the start of a working session for the handshake (read CLAUDE.md → progress.md → the latest ADR, then state where we are and what's next), and at the end of a session or on "wrap up" / "update progress" to update progress.md and scaffold the next-numbered ADR. Keeps continuity across context clears and model switches.

2026-06-140

character-build

Pr0degie/dungeonmaster

Use when a player's Imperium Maledictum character JSON comes in and needs validating + deploying, when building a new IM character from scratch, or when backfilling backstory onto an existing sheet. Validates budgets/formulas/psyker powers/augmetics against the active profile, writes data/party/<player>.json, generates the PDF sheet, and proposes (confirm first) the merge into the session characters.json + aliases.

2026-06-130

name	playtest-triage
description	Use when the user pastes a live Discord session log, transcript, or annotated complaints from a play session and wants a fix round. Diagnoses the failure chain from the logs, extracts player improvement-wishes, applies deterministic fixes, runs the test suite, commits per round, and updates progress.md (and an ADR when a real trade-off was decided).

Playtest triage — turn a live session into a fix round

Golden rules that bind this work

#2 — dice rolling and resolution are code (dmbot/rules/engine.py + the active profile), never the LLM. A fix that lets the model invent results is wrong.
#3 — hard world state (HP, inventory, flags) is advanced by code, never written from LLM free text.
#4 — Bot A's user-ID filter in the sink stays. Never weaken feedback protection "for debugging".
#8 — anything the DM says stays German; code/logs/commits stay English.

Procedure

Read the evidence, not your priors. Sources, in order:
- the pasted log / annotations (the user's own complaints carry the most signal)
- logs/debug.log (opt-in verbose), logs/transcript.log (STT → narration → TTS)
- the live state under data/sessions/<id>/state.json (wounds, NPCs, scene, recap)
Classify each complaint → root cause. Past diagnoses to pattern-match against (progress.md decision log):
- D43 / ADR 018 — wrong channel → silent example-party fallback → echoed player line read aloud (a failure chain, not regression).
- D57 / ADR 027 — silent num_ctx truncation (8192 hardcoded) dropped persona + adventure from the prompt mid-session → "ignores the story", over-long answers, puppeting.
- D59 / ADR 028 — OCR/statblock junk leaking into the per-turn RAG block. Suspect the pipeline (sample rate, context budget, dedupe, channel) before the model.
Fix deterministically. Prefer a code guard over a prompt nudge. Prompt-only is allowed only when the failure is genuinely stylistic — and then note that a code guard is the fallback if nemo doesn't adhere.
Run the suite: uv run --with pytest python -m pytest. Add fixed-seed tests for any new deterministic logic.
Commit per round with a scoped imperative message (e.g. dmbot(memory): cumulative auto-recap so wrap-up can't drop the session start).
Record the round (end-of-session ritual): update progress.md ## Last session and ## Current focus, and rotate the previous ## Last session entry to docs/progress-archive.md so the live file stays lean (per session-ritual); write the next-numbered ADR if a real trade-off was weighed. Mark the result live-unverified — the next real session is the gate. Name the concrete thing the next round should confirm.

Notes

Every log the user pastes is pre-change: your fixes are unproven until the next live run. Don't claim "fixed", claim "fix landed, live-unverified, gate = X".
Cross-reference the playtest-tuning-loop memory and the relevant ADR before touching code in the area a complaint points at.