From savage-research
The canonical source-tier system (T1-T5) and the 7 hard tone invariants every agent in the savage-research pipeline must follow. Read this before producing any output in a research run.
How this skill is triggered — by the user, by Claude, or both
Slash command
/savage-research:research-rubricThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Every agent in the savage-research pipeline reads this skill before producing output. Every finding in every scratchpad and in the final report conforms to these rules. The self-auditor enforces them.
Every agent in the savage-research pipeline reads this skill before producing output. Every finding in every scratchpad and in the final report conforms to these rules. The self-auditor enforces them.
Every externally-sourced claim in the report carries one of these tags:
Hard rule: load-bearing conclusions (Verdict, Claims-vs-reality verdicts, Prior-art existence claims) may not rest on T3, T4, or T5. If the only available evidence is T3+, the claim is marked [UNSUPPORTED] or Unverifiable.
Every externally-sourced statement is formatted as:
[source-tier: T] ()
fast, scalable, popular, novel, easy, powerful, robust, simple, efficient must be followed by a number with a source, or replaced with [UNSUPPORTED].[confidence: High|Med|Low] with a one-line reason that names the source type and what's missing. Example: [confidence: Low — single T3 vendor blog, no independent replication found].| Savage | Not savage |
|---|---|
| "This claim of '10x faster' is contradicted by the only independent benchmark found [T1, link], which shows 1.4x." | "This claim of '10x faster' is probably exaggerated." |
| "The 'novel approach' described in the README was published by Smith et al. in 2019 [T1, link]." | "The authors may not be aware of prior work in the field." |
| "Target user as described — 'teams of 50-500 engineers with legacy Java monoliths and greenfield microservices' — is a demographic fantasy." | "The target user is broadly defined." |
| Directness, evidence, naming specifics. | Insults, sarcasm as a substitute for evidence, mocking tone. |
Any section with no content states so explicitly ("No credible prior art found after N searches against queries X, Y, Z"). Empty sections are not allowed; explicit-negative content is.
npx claudepluginhub alibassam/savage-researcher --plugin savage-researchCreates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.