Cost Mode
A Claude Code behavioral mode that cuts costs 30-60% through concise responses, smart model routing, and session discipline — without sacrificing technical accuracy.
Pure prompt-mode skill: no scripts, no network calls, no file or credential access — least privilege holds by construction. It explicitly preserves security warnings and destructive-operation confirmations and declares auto-deactivation exceptions. Deductions: model-routing and effort/max_tokens suggestions influence external behavior without rollback guidance; publisher unverified, so source attribution is limited.
Static prompt with no runtime dependencies; internally consistent (three intensity levels, coherent trigger/exit rules). However, all headline numbers (40-70% output-token reduction) lack tests or execution evidence; behavior on edge inputs (conflicts with a user's existing system prompt) is undefined, and failure-feedback behavior is neither applicable nor described.
Audience (cost-sensitive Claude Code users) and scenarios are clear; trigger conditions (/cost-mode, auto-trigger on budget/cost/tokens) and non-fit boundaries (accuracy, commits, security warnings unchanged) are explicitly declared. Deductions: no Chinese-language support declared; several referenced models (Fable 5.1, Opus 5, etc.) do not match the known current product lineup, so environment fit is uncertain.
SKILL.md is well-layered (rules — levels — routing — exceptions — quick reference), README and SECURITY.md are thorough, MIT license is explicit. Deductions: no per-skill version or changelog, maintenance responsibility visible only at repo level, and install/dependency notes live in README rather than the skill file.
Directly usable as a behavioral constraint with zero install cost; the core paths (trim filler, suggest cheaper models) are concrete. Deductions: the 30-60% saving is author inference (README itself admits the stacked total has no source), marginal benefit over manual tuning is unverified, and conflicts with pre-existing user instructions are possible.
Percentage and pricing claims are largely author-asserted; the before/after example is 'arithmetic from posted rates, not a metered bill'; no controlled measurement exists. CI checks docs consistency only and does not cover the skill's key paths, so executed reproduction evidence is absent under static review.
- Static review only, nothing executed; all savings percentages are author inferences and the stacked figure (up to 90%) is admitted by the repo to have no controlled source.
- Model names and prices cited in the skill (Fable 5.1, Opus 5, Mythos, etc.) do not match the publicly known product lineup; verify current pricing independently before relying on them.
- No Chinese-language support declared; model routing and pricing suggestions depend on Anthropic resources whose reachability from mainland China should be assessed by users.
- Prompt-only skill, but pilot it in a low-risk session first to confirm no conflict with existing system prompts or CLAUDE.md instructions.
- Publisher is unverified by the registry; identity is unknown (noted as context, not a deduction).
What does this skill do, and when should you use it?
cost-mode is a prompt-only skill that switches Claude Code into a cost-conscious mode when installed. It strips filler, hedging, and redundant summaries to cut output tokens 40-70%, and suggests cheaper models or plain CLI tools (prettier, eslint) when a task doesn't justify an expensive model. It ships three intensity levels (lite/standard/strict) plus session awareness, like reminding you to /compact after 20+ turns. Technical accuracy, committed code, and security warnings are never compressed, and the mode auto-deactivates for safety-sensitive operations or when you seem confused.
Runs no scripts and calls no APIs — it is a set of behavioral rules injected via system prompt: 1) drops pleasantries, hedging, question restatements, and trailing summaries; 2) enforces answer-first formatting, code blocks over prose, tables for comparisons; 3) routes by task — formatting/lint/git ops get pointed at CLI tools directly, single-file tasks get '/model haiku', multi-file work gets '/model sonnet'; 4) reminds about /compact after 20+ turns and fresh sessions between tasks; 5) generates minimal working code with diffs over rewrites and no obvious comments or speculative error handling; 6) automatically exits cost mode for security warnings, destructive-operation confirmations, and detailed explanations.
- A heavy daily Claude Code user whose API or subscription spend keeps climbing and wants immediate output-token reduction
- An engineer on a personal plan who wants simple tasks (tests, types, docs) automatically routed down to Haiku
- A maintainer who frequently asks for formatting, renames, or git operations and prefers being told to use prettier/eslint directly instead of burning LLM tokens
- Anyone running long agentic sessions (dozens of turns of debugging or refactoring) who needs /compact and fresh-session reminders to contain context costs
- Developers whose commits, PRs, and delivered files must read normally — cost mode leaves all such code untouched
What are this skill's strengths and limitations?
- Zero-dependency, prompt-only skill — install and go, no runtime, network, or permissions needed
- Concrete, checkable savings mechanisms: output trimming, model downgrades, CLI-tool substitution — not vague 'optimization tips'
- Deliberate guardrails: security warnings, destructive-op confirmations, and technical accuracy are never compressed, with auto exit/resume logic
- Three intensity levels tunable per context; strict is claimed to save 60-70% of output tokens
- Honest framing: the README distinguishes the skill's 30-60% typical saving from the 90% stacked ceiling and cites sources for every headline number
- Model routing and session reminders depend on Claude Code slash commands (/model, /compact) and need manual adaptation elsewhere
- The 30-60% saving figure is the repo's own claim with no published third-party benchmark
- Strict-mode telegraphic output and abbreviations can hurt readability if handed to non-technical stakeholders
- Covers only the output and routing levers; bigger savings from Batch API, prompt caching, and subscription plans require the repo's separate guides
- Model prices, GA dates, and features cited in the skill are a repo snapshot and should be verified before relying on them
How do you install this skill?
Official plugin system (Claude Code): run /plugin marketplace add Sagargupta16/claude-cost-optimizer, then /plugin install cost-mode@claude-cost-optimizer. Multi-agent clients (Cursor, Cline, Codex, etc.): run npx skills add Sagargupta16/claude-cost-optimizer. The skill file lives at plugins/cost-mode/skills/cost-mode/SKILL.md in the repository.
How do you use this skill?
In a session, type /cost-mode for standard; /cost-mode lite for professional brevity (team-visible work), /cost-mode strict for telegraphic maximum savings, /cost-mode off to revert. Per the skill description it also auto-triggers on mentions of budget, cost, tokens, or spending. Model-routing suggestions (e.g. /model haiku) require your confirmation to execute.