| name | context-cache-discipline |
| description | Token and prompt-cache discipline — lean always-on rules, stable prefixes, skill indirection, minimal Task payloads, bounded reads. |
| version | 1 |
Context and cache discipline
Use this skill when orchestrating or editing the pack so context stays small and early prompt prefixes stay stable (better cache hit odds on providers; always fewer input tokens).
Principles
| Goal | Practice |
|---|
| Stable prefix | Prefer one canonical copy of long text in skills; rules point — avoid churn in alwaysApply: true bodies. |
| Lean always-on | Rarely needed norms → alwaysApply: false + globs, or load skill on demand. |
| Indirection | Agents/rules cite skill path; do not paste full playbooks into rules or agent bodies. |
| Task payloads | Parent Task prompts: goal, constraints, approved_plan_path, phase/shard id, touches[], correlation ids — not full rules or transcripts. |
| Reads | Batch parallel Read where independent; one read per file per turn unless content changed; lookup-first for brain-memory-kb. |
| Subagent traffic | Children: caveman-ultra + single YAML envelope per templates/subagent-response.yml.tmpl; parents: do not paste raw child YAML to user. |
Non-claim: Exact cache eligibility is vendor-defined; token reduction is always a win.
Memory / KB policy indirection
Load ~/ai-brain/org/global/config/memory-demotion.yml (via brain-memory-kb) for query ladder L0–L3 caps and demotion thresholds — do not duplicate full policy text here.
When to load what
Forbidden
- Pasting full
alwaysApply rule bodies into chat to “teach” a subagent
- Re-fetching the same skill/rule in one turn without reason
- Oversized inline blobs in
Task (anti-dup hook is fail-closed)
