C.A.S.H. (Constraints, Ambiguity, Sequencing, Heterogeneity) is an agent skill that evaluates each component of a proposed system and routes it to traditional code, an LLM call, or a hybrid of both. It is an Orchestration Router, not an AI gatekeeper: hard operational constraints (latency, reproducibility, explainability, compliance topology) gate first, then three dimensions vote — producing a Decision Matrix with a stated, arithmetic-backed cost justification for every AI or Hybrid recommendation.
Two failure modes recur in AI-assisted architecture:
| Failure mode | What it looks like | Cost |
|---|---|---|
| Over-applying AI | An LLM call on the hot path, or on deterministic plumbing that a function already handles perfectly | Latency blown past SLA, token spend that scales with traffic instead of staying flat, non-reproducible output where the requirement is bit-identical |
| Under-applying AI | Regex chains and hand-rolled rule engines standing in for tasks that are inherently ambiguous (intent, sentiment, root-cause synthesis) | Ever-growing edge-case handling, brittle heuristics, a quality ceiling templates can't clear |
C.A.S.H. addresses both in a single evaluation pass. The evaluation is deliberately neutral — the skill contains no "champion AI" instruction, because framing an evaluator measurably biases its verdicts (see Research grounding below).
C — Constraints is a gate, resolved before any scoring:
| Gate | Fires when | Routes to |
|---|---|---|
| Latency | Sub-100ms budget on the critical path (fastest mainstream LLM TTFT is ~400–600ms) | Code |
| Reproducibility | Bit-identical output required (API-served LLMs aren't bit-reproducible even at temperature 0) | Code |
| Explainability | A reviewer must know why this decision — not merely that it was logged. Traceability alone (logging, per EU AI Act Art. 12) does not fire the gate; nor does a genuinely advisory output — one an accountable human is equipped to verify and realistically positioned to override | Code |
| Privacy / topology | PII/PHI/regulated data and no compliant deployment path (BAA-covered, on-prem, or pre-call redaction). Unstated topology in a sensitive domain → the skill must ask, not assume | Code |
A / S / H are votes on whatever the gate doesn't resolve:
| Dimension | Ask | Code when | AI when |
|---|---|---|---|
| A — Ambiguity | Is the logic deterministic or ambiguous? | Explicit, rule-based (scoring, formatting) | Nuance, intent, tone, meaning — even in structurally uniform data |
| S — Sequencing | Is the sequence of steps fixed at design time or decided at runtime? | Steps and order known upfront, however much design-time branching | The plan is assembled at runtime from intermediate results |
| H — Heterogeneity | Combined by structure alone, or synthesized across disparate sources? | Joins, merges, aggregations — structure only | Correlating heterogeneous, unstructured signals into one conclusion |
Decision rule: three voters, no ties. 3-0 → that engine. 2-1 → Hybrid, and the minority dimension names the AI's pipeline position (minority A→AI: AI interprets before code; S→AI: AI plans, code executes; H→AI: AI synthesizes after code assembles).
Day-2 rule: the matrix is a design-time verdict, and its inputs drift. SKILL.md defines re-evaluation triggers — ruleset growth on a Code component, ~3× cost overrun or a failed cache assumption on an AI component, model deprecation, SLA or topology change — each prompting a component-level re-vote, not a system redesign.
Why these letters? Each dimension asks exactly what its name says, the three voting axes ask distinct questions (logic-ambiguity to A, workflow dynamism to S, data shape to H — distinct questions, though H-synthesis usually co-occurs with A-ambiguous; the 2-1 mechanics survive the correlation), and the strongest routing signal — hard constraints — sits first in the acronym as a gate. The full design reasoning is in
references/philosophy.md.
These are estimates, not measurements — each is derived in the linked example as tokens/call × calls/day × per-MTok price, with model tier and cache assumptions stated, against Anthropic list pricing (retrieved July 2026). Re-derive before budgeting. The code side is held to the same arithmetic standard: infrastructure common to both designs cancels, and only the incremental cost decides (worked in the e-commerce example).
| Scenario | Constraint | Estimate |
|---|---|---|
| Real-time fraud detection | 50ms SLA on the transaction path | Hot path routes to code — $0 tokens on millions of transactions/day; async AI pattern analysis + investigation reports ≈ tens of dollars/day (arithmetic in the example) |
| E-commerce search @ 10M queries/day | Per-query LLM cost at scale | A frontier LLM on every query ≈ $54K–$180K/day (1–2 calls/query, ~500–800 in + 100–200 out tokens, Sonnet/Opus-tier rates); small-model + cached + embeddings design ≈ $1.5K–$2.5K/day, lower with self-hosted small models |
| DevOps incident response | ~50 incidents/month | ≈ $15–$75/month in tokens vs ≈ $4,350/month of engineer time saved by the MTTR reduction |
| Clinical decision support | PHI / HIPAA topology | Compliance gating resolves deployment first (BAA-covered or on-prem), so AI-assisted diagnosis is deployable instead of blocked outright |
Every operative rule in the skill traces to a checkable source — the full annotated list is in references/research.md:
- Code-first premise — Zinkevich, Rules of ML, Rule #1; Anthropic, Building Effective Agents (also the source of the workflow-vs-agent split behind the S dimension).
- Neutrality (no advocacy instruction) — Sharma et al., Towards Understanding Sycophancy in Language Models (ICLR 2024); Turpin et al., Language Models Don't Always Say What They Think (NeurIPS 2023).
- Reproducibility gate — Thinking Machines, Defeating Nondeterminism in LLM Inference.
- Explainability vs traceability — EU AI Act, Article 12 (record-keeping).
- Latency gate — e.g. LLM latency benchmarks (fastest mainstream TTFT ≈ 400–600ms).
- Skill structure — Anthropic, skill authoring best practices.
Seven scenarios in skills/cash-framework/examples/, each ending in a full Decision Matrix and Cost Checkpoint:
- Fantasy Football Team Drafter — code for stats and rule enforcement, AI for drafting strategy (a genuine 3-0 AI verdict on the drafting component)
- Customer Support Router — code routes and delivers, AI classifies intent and reads tone
- CI/CD with AI Code Review — fail-fast-with-code, AI reviews
- Real-Time Fraud Detection — the C gate in action: 50ms SLA routes the hot path to code; AI finds patterns rules can't
- E-Commerce Search & Personalization — all three Hybrid pipeline positions at 10M queries/day, with the worked cost arithmetic
- Clinical Decision Support — compliance-topology gating: PHI-bound components to code, AI-assisted diagnosis where the deployment permits
- DevOps Incident Response — the S dimension in practice: AI plans remediation, code validates and executes
skills/cash-framework/
SKILL.md # operative instructions the agent follows
references/
philosophy.md # design rationale — why route to code frees AI budget; why these letters
research.md # annotated bibliography: every rule → its checkable source
tco_checklist.md # the full Cost Checkpoint checklist
examples/ # seven worked scenarios (table above)
package.json # agent plugin/package manifest
.claude-plugin/, .cursor-plugin/, .codex-plugin/, .kimi-plugin/, .opencode/, .pi/
# per-harness plugin manifests and stubs
Installation differs by harness. If you use more than one, install the skill separately for each one. The repository lives at github.com/pathak-prashant/cash-framework.
Plugin systems evolve quickly - only the Claude Code commands below have been verified against a current release. Treat the rest as illustrative and check each tool's plugin docs.
No plugin needed - copy skills/cash-framework/ into your project's .claude/skills/ (or ~/.claude/skills/ for every project).
Or install as a plugin, since this repo ships a marketplace.json:
/plugin marketplace add https://github.com/pathak-prashant/cash-framework
/plugin install cash-framework-skill@cash-framework-skill-marketplaceagy plugin install https://github.com/pathak-prashant/cash-frameworkReinstall with the same command to pick up changes.
Skills load via the pi.skills entry in package.json, which already points at ./skills.
pi install git:github.com/pathak-prashant/cash-frameworkFor local development, run Pi with this checkout loaded as a temporary package:
pi -e /path/to/cash-frameworkThe commands above are not verified here - check your Pi installation's documentation.
Open the plugin search interface and install from https://github.com/pathak-prashant/cash-framework if supported, or paste skills/cash-framework/SKILL.md into your custom instructions.
Copilot supports the open Agent Skills standard that SKILL.md already conforms to (name + description frontmatter) — see GitHub's agent skills documentation. Copy skills/cash-framework/ into one of the repo-level skill directories Copilot recognizes:
your-project/
├── .agents/skills/cash-framework/SKILL.md # cross-tool location
├── .github/skills/cash-framework/SKILL.md # Copilot-specific
└── .claude/skills/cash-framework/SKILL.md # already covered above
For every project on your machine, use the global equivalent instead: ~/.agents/skills/cash-framework/ or ~/.copilot/skills/cash-framework/.
If your Copilot version predates Agent Skills support, fall back to pasting SKILL.md into .github/copilot-instructions.md (repo-wide) or into a scoped .github/instructions/cash-framework.instructions.md file with the required frontmatter:
---
description: 'Apply the C.A.S.H. framework to architecture and build requests.'
applyTo: '**'
---Cline also reads the .agents/skills layout from the section above - copy skills/cash-framework/ into .agents/skills/cash-framework/ (workspace) or ~/.agents/skills/cash-framework/ (global, available to every project) and it shows up under Cline's Skills panel, toggleable per-skill.
Alternatively, paste the contents of skills/cash-framework/SKILL.md into a .clinerules file at your project root, or into .clinerules/cash-framework.md if you keep rules split across multiple files.
These typically fetch from their own marketplaces - publish following their plugin guidelines, or paste SKILL.md into custom instructions/rules directly.
droid plugin marketplace add https://github.com/pathak-prashant/cash-framework
droid plugin install cash-framework-skill@cash-frameworkA stub ships at .opencode/plugins/cash-framework-skill.js - it's a placeholder that does not yet load the skill. Wire it to OpenCode's plugin API, or paste SKILL.md into your OpenCode rules. The Pi extension stub at .pi/extensions/cash-framework-skill.ts is in the same state.