Manus Credit Optimizer v5
Automatically classifies task complexity before execution and routes each task to the cheapest capable model tier, claiming 30-75% credit savings with no quality loss.
The skill is pure Markdown instructions: no arbitrary code execution, no secrets, no external side effects, transparent data flow; but it makes absolute marketing claims ('zero downsides', '100% savings', '0% quality loss') with no verifiable basis, defines no rollback or fail-safe mechanism, and gives no security scoping for the referenced MCP/PyPI components. Deducted for: exaggerated claims, missing rollback/boundary disclosure.
Internal logic is broadly self-consistent (classify → check → route → execute), but there are no tests, no abnormal-input handling, and no failure feedback; all savings figures (90-100%, 40-60%) are assertions, not reproducible results; version numbers disagree across SKILL.md (5.2.0), CHANGELOG (5.2.2) and server. (5.2.2). Static cap is 10. Deducted for: no tests, no error paths, version inconsistency.
Target scenario (reducing AI token/credit spend) and trigger conditions are reasonably clear, with declared platform compatibility; however capability boundaries and non-fit ranges are nearly undefined, and the core value depends entirely on Manus AI and other overseas billing models, with no mention of mainland-China reachability and zero Chinese support. Deducted for: thin boundary/trigger evidence, dependence on overseas services.
MIT license, versioning, CHANGELOG, manifest and pyproject metadata exist and docs are readable and layered; but documentation is heavily contaminated by marketing, changelog carries future dates (2026) and versions disagree across files, maintenance responsibility rests solely on an individual developer, and there is no FAQ or known-limitations disclosure. Deducted for: version inconsistency, marketing-heavy content, weak governance.
The skill's output is prompt-level routing guidance that could be applied directly in conversation and is plausibly better than unstructured usage; but every quantitative benefit (47%, 115x, payback period) lacks source data, the '53-scenario audit' has no reproducible material, actual effectiveness is unverifiable, and some benefits are tied to purchasing a paid bundle via marketing funnels. Static cap is 7; scored lower. Deducted for: unverifiable benefit claims, no representative output samples.
Core claims (53 scenarios, 0% quality loss, 30-75% savings) have no auditable raw data, no test suite, and no third-party reproduction; the repo's CI only checks marketing channel health, not skill functionality; 'As Seen On' external references cannot be cross-corroborated from the source. Deducted for: nearly all key claims being author marketing assertions.
- All benefit figures (47% savings, 115x speedup, 0% quality loss, 53-scenario audit) are the author's unilateral marketing claims with no reproducible test data or raw audit material in the repository.
- The absolute 'zero downsides / never reduces quality' promise should not be trusted; any cost-optimization strategy may affect output quality — evaluate independently.
- The skill's value depends heavily on the specific pricing model of Manus AI and similar overseas platforms; it breaks if pricing changes, and mainland-China network reachability is unassessed.
- Version numbers are inconsistent across files (SKILL.md 5.2.0 vs CHANGELOG/server. 5.2.2) and the CHANGELOG carries future 2026 dates, undermining metadata credibility.
- The publisher is an unverified individual developer with no guaranteed long-term maintenance; the README strongly funnels readers toward a $9-$12 paid bundle, indicating clear marketing motivation.
- The CI workflows only monitor marketing channel availability and send WhatsApp notifications via a third-party API (Evolution); they do not constitute functional validation.
What does this skill do, and when should you use it?
Credit Optimizer is an Agent Skill distributed as a SKILL.md containing a routing playbook: it classifies each incoming task by complexity, scores prompt clarity, then routes to free, standard, or premium model tiers — splitting mixed tasks into sub-tasks where cheaper. The author reports a self-run audit of 53 scenarios with 30-75% average savings, zero quality degradation, and under 3% false routing. It ships in two forms: a paid Manus Skill bundle (which includes a Fast Navigation speed module) and a free Python MCP server on PyPI. It fits heavy AI-agent users burning credits fast, but note that all performance figures are author-reported and independently unverified.
Reads the incoming prompt and runs a four-step pipeline: 1) classifies the task (simple Q&A, code generation, research, creative writing, data analysis, complex reasoning); 2) scores prompt clarity (asking for clarification before executing if below 5/10), detects whether the task can be split into cheaper sub-tasks and whether real-time data is needed; 3) routes to free/standard/premium tiers, steering simple Q&A to zero-credit Chat Mode; 4) applies context pruning, scoped output formats, caching checks, and batching during execution. The MCP variant installs via pip and runs as python -m mcp_credit_optimizer for any MCP-compatible client.
- Heavy Manus users watching their monthly credits evaporate, who want simple Q&A to stop burning Max-mode credits.
- Batch processors with mixed-complexity workloads (e.g., 10 web scrapes plus 5 reports) who want per-task optimal routing instead of uniformly maxing out.
- Claude Desktop or Cursor users who want the routing logic for free via the MCP server rather than paying for the Skill.
- Users whose vague prompts frequently produce off-target output — the skill requests clarification before spending credits.
- Research tasks hitting many URLs: the paid bundle's Fast Navigation claims to fetch 10 URLs in ~1.3 seconds versus 150+ seconds via browser tool calls.
What are this skill's strengths and limitations?
- The approach is sound: complexity classification plus tiered routing is a standard cost-optimization pattern, and zero-credit Chat Mode routing is directly valuable for Manus users.
- There is a free path (the PyPI MCP server), and the boundary between paid and free versions is clearly explained.
- The paid version advertises a hardcoded Quality Veto Rule: optimizations that would reduce quality are skipped entirely.
- MIT license, auditable code, and an explicit list of compatible clients (Claude Desktop, Cursor, Windsurf, etc.).
- All performance claims (47% average savings, 53 scenarios, 0% quality loss, 115x speedup) are self-reported with no independent verification or public test suite.
- Marketing figures are inconsistent across pages: the GitHub description claims payback in ~27 prompts while the README says ~36; SKILL.md lists the bundle at $29 while the README says $12.
- The free MCP server requires manual invocation, unlike the automatic Skill version.
- Fast Navigation is exclusive to the paid bundle — the free version has no speed component.
- Even at a claimed <3% false routing rate, complex tasks can occasionally be downgraded to cheaper models.
How do you install this skill?
Paid Manus Skill: buy from Gumroad ($9 individual / $12 bundle with Fast Navigation) or SkillFlow, then copy the skill files to ~/skills/credit-optimizer/ and ~/skills/fast-navigation/. Free MCP server: run pip install mcp-credit-optimizer, then add {"mcpServers":{"credit-optimizer":{"command":"python","args":["-m","mcp_credit_optimizer"]}}} to your MCP config. Note: no installation details beyond the file locations are documented.
How do you use this skill?
The Manus Skill activates automatically on every task — no invocation needed. The free MCP version only saves credits when you manually invoke it. No example trigger prompts are documented; routing is applied by the model following the SKILL.md classification, scoring, and routing instructions.