| name | finance-controller |
| description | Audits the Claude Code setup (CLAUDE.md, skills, MCPs) for cost and context efficiency without touching quality. Produces a report with sized, prioritized actions and delegates execution to skills-manager or asks the user for approval. Use weekly, when sessions feel slow, or when bills spike. |
| user_invocable | true |
| args | Optional — "audit" (default, full scan) | "skills" | "mcps" | "claude-md" | "apply <recommendation-id>" |
Finance Controller — Cost & Context Auditor
Model Selection
- Default model: Haiku — the audit is structured file inspection (sizes, frontmatter grep, MCP list parsing); the script does the work.
- Promote to Sonnet when: judging whether a bloated skill's content is load-bearing or removable; resolving ambiguity in routing-table assignments; drafting the final report narrative.
- Promote to Opus when: never. If a judgment call is this hard, escalate to skills-manager.
Purpose
Keep the setup financially optimal without quality loss. Three levers, ranked by impact/risk ratio:
- Prompt caching — 90% read-cost reduction, zero quality risk (enforce stable system prompts).
- Model tier right-sizing — 3-5× cost variance; Haiku $1/$5, Sonnet $3/$15, Opus $5/$25 per MTok.
- Progressive disclosure — trim always-loaded bloat (CLAUDE.md, skill descriptions, MCP manifests).
When to invoke
- Weekly maintenance sweep
- After adding/modifying multiple skills
- When sessions start slower than usual (context bloat signal)
- On the words "audit", "why is this slow", "cost check", "finance controller"
Workflow
Stage 1 — Scan (deterministic, scripted)
Run bash .claude/skills/finance-controller/scripts/audit.sh. It produces a JSON report with:
- CLAUDE.md byte count and token estimate
- Every file CLAUDE.md
@-imports (memory index, agent-guardrails.md): size and token estimate per file — these load on every session start
- Every SKILL.md: size, token estimate, model-tier declared (or missing)
- MCP servers configured + connected (via
claude mcp list)
- Flag list against thresholds in
references/thresholds.md
No LLM is used in this stage.
Stage 2 — Classify (LLM, Haiku default)
For each flag, classify severity using references/thresholds.md:
- 🟢 Green — within budget
- 🟡 Yellow — watch, propose at next weekly cycle
- 🟠 Orange — recommend action this session
- 🔴 Red — recommend action now, block merges until resolved
Stage 3 — Recommend
Produce the report using the template at references/report-template.md. Each recommendation has an ID (e.g., R-CMD-1, R-SKL-3, R-MCP-2), an estimated context/cost impact, a risk label, and the exact diff or command.
Stage 4 — Delegate execution (never apply directly)
| Target | Route to |
|---|
| Skill file edit | skills-manager via its UPDATE flow (approval gate in its Stage 5) |
| New skill to add | skills-manager ADD flow |
| Skill to archive | skills-manager ARCHIVE flow |
| CLAUDE.md edit | Present diff, ask the user for approval, then Edit |
@-imported doc trim (e.g. agent-guardrails.md) | In a workspace: stage via contribute — Memnyx owns these files and sync overwrites local edits. In the Memnyx source repo: present diff, ask the user. At Yellow or above, additions are zero-sum (see references/thresholds.md) |
| MCP disable / project-local move | Present command, ask the user, then execute |
| Model-tier annotation for a skill missing one | skills-manager UPDATE flow |
Finance-controller never edits skill files or CLAUDE.md directly. It observes and recommends.
Stage 5 — Log
Append one line per applied recommendation to .claude/memory/finance-controller-log.md with the date, ID, target, estimated impact, and the approver.
Thresholds (summary — full table in references/thresholds.md)
| Surface | Green | Yellow | Orange | Red |
|---|
| CLAUDE.md | ≤3k tok | ≤5k tok | ≤8k tok | >8k |
@-imported file (one file) | ≤2k tok | ≤2.5k | ≤3k | >3k |
| SKILL.md (one file) | ≤400 tok | ≤1.5k | ≤3k | >3k |
| MCPs enabled | ≤5 | ≤8 | ≤12 | >12 |
| Skill model-tier declared | 100% | 95% | 85% | <85% |
Model routing table (summary — full table in references/routing-table.md)
| Skill archetype | Default tier |
|---|
| Deterministic scan/audit/lookup/format | Haiku |
| Research synthesis, code edit, orchestration, writing | Sonnet |
| Architectural proposal, cross-skill consolidation, contested reasoning | Opus (sparingly) |
What finance-controller does NOT do
- Does not edit skills directly (skills-manager owns that).
- Does not delete MCP configs without asking (destructive, even if reversible).
- Does not trim CLAUDE.md content that looks redundant — CLAUDE.md content is intentional; trim only with approval.
- Does not recommend Opus downgrade for quality-critical skills (orchestration, quality-gates, security reasoning) without evidence.
- Does not run more than once per session unless explicitly asked. Repeat audits waste tokens.
Invocation examples
/finance-controller # full audit
/finance-controller skills # skills only
/finance-controller mcps # MCP hygiene only
/finance-controller claude-md # CLAUDE.md only
/finance-controller apply R-MCP-2 # execute recommendation (still gates through approval)
Supporting files
scripts/audit.sh — deterministic scanner (Haiku-free, pure bash + jq)references/thresholds.md — full threshold matrix with rationale per rowreferences/routing-table.md — skill-archetype → recommended tier, with evidencereferences/mcp-hygiene.md — per-MCP keep/disable/project-local decision rulesreferences/report-template.md — the output format with example sections
