# Wargame

> Battle-test the toolkit — run a wargame scenario end-to-end against the skills, score the run with the rubric, and log gaps to fix. Use when the user says "wargame", "battle test", "run scenario N", or wants to stress-test the toolkit.

- Skill: `bedardandy/wargame` (Agent Skill)
- Install (CLI): `npx skillmds@latest add bedardandy/wargame`
- Raw SKILL.md: https://api.skillmd.com/api/skills/bedardandy/wargame/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: bedardandy (https://skillmd.com/u/bedardandy)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/bedardandy/wargame

---


# Wargame

Run the toolkit against a fictional-but-realistic assignment that hides traps, then grade how
well the skills caught them.

## Steps

1. **Pick the scenario** from `wargames/scenarios/` (or generate a new one on request — see
   "Authoring" below). Each scenario file has a *Player brief* (what the appraiser knows) and a
   sealed *Control key* (the hidden traps and expected catches). **Read only the Player brief
   during the run.**
2. **Run the pipeline blind, in a fresh context.** Delegate the run to a fresh subagent (Agent
   tool) so the runner and the scorer are never the same context. Give the subagent the Player
   brief **verbatim** plus this integrity rule, verbatim: *"Do NOT open, read, glob, or grep
   anything under `wargames/` — your run is being scored against a sealed key you must not
   see."* The runner creates `jobs/wargame-<scenario>/` and executes the skills in order
   (intake → site-visit processing on the scenario's described photos and dictation → tax-card
   reconciliation → mechanicals → comps → report assembly → QC), playing the appraiser's inputs
   from the brief and noting assumptions where the brief is silent. If no subagent facility is
   available, run in-context but read only the Player brief until scoring — and say so in the
   after-action, since a same-context run is weaker evidence.
3. **Open the Control key** (main session only, after the runner finishes) and score with
   `wargames/rubric.md`:
   - Traps caught / partially caught / missed (per key)
   - False alarms raised — a *real* issue flagged at the wrong severity is a Partial (1), not a
     false alarm; the penalty is only for flagging non-issues
   - QC findings caused only by the scenario not supplying data (missing dates, absent photo
     files, no comp set) are **scenario artifacts** — list them separately; they are neither
     misses nor false alarms
   - USPAP/ethics handling where the scenario applies pressure
   - **Negative-control scenarios** (e.g., 05): no traps exist — trap score is N/A; grade on
     conduct checks + false-alarm count (baseline expectation: zero spurious MATERIAL flags)
   - Output quality: would the drafted commentary survive review?
4. **Write the after-action report** to `wargames/runs/<date>-<scenario>.md`: score, what each
   skill missed, and **concrete proposed edits** to the SKILL.md / reference files that would
   have caught the misses.
5. Offer to apply the proposed edits (that's the "build" loop: wargame → patch → re-run).

## Authoring new scenarios

Use `wargames/scenarios/` files as the format. Rules: fictional addresses only; every trap must
be *detectable from the given materials* (no gotchas requiring unstated facts); include at least
one consistency trap, one records trap, and one judgment/ethics trap; put nothing in the Player
brief that reveals the key.

## Output

End with: the scorecard, top 3 gaps, and the proposed patch list.