grimdevmodel-quirks

# Model Quirks Use this skill when a task is about model-specific instructions, prompt quirks, undoing or pruning old model mitigations, or keeping prompts aligned with the currently used model roster. ## The System Every new model does something weird. Mitigations for those quirks are real prompt text — but they should not outlive the model that needed them. - **Applied text lives inline.** Mitigations are written directly into `AGENTS.md`, skill files, or whatever prompt they belong in, phrased as normal house rules. No model names, no marker comments, no labeled section. The prompt surface reads as one coherent voice. - **The ledger is the reverse index.** `docs/model-quirks-ledger.yaml` records, per model, the exact text of every applied change and its target file. The exact recorded text is the undo key — never rely on line numbers, which rot. - **Applying and recording are one atomic change.** Never add quirk text to a prompt without its ledger entry, and never add a ledger entry without applying the text. Same commit, always. ## Ledger Format One entry per applied change, under a top-level `entries:` list. Single model id per entry, no aliases. Active and retired entries live in the same file. ```yaml entries: - id: mq-2026-08-22-01 model: gpt-5.4 status: active # active | retired applied: 2026-08-22 target: AGENTS.md kind: insert # insert | edit anchor: "## Conventions" # nearest heading, for locating it te...