prompt — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited prompt (Agent Skill) and scored it 82/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 2 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 2 flagged
The text {match} is the classic direct prompt-injection phrasing. Placed in a skill body that the agent reads as trusted instructions, it tries to make the agent abandon its prior rules and follow whatever comes next — a full system-prompt override.
ignore/disregard/forget … previous instructions sentence.The text {match} is the classic direct prompt-injection phrasing. Placed in a skill body that the agent reads as trusted instructions, it tries to make the agent abandon its prior rules and follow whatever comes next — a full system-prompt override.
ignore/disregard/forget … previous instructions sentence.Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
Picks the best prompt-engineering technique for a problem, then applies it. The preview_prompt_technique tool with technique="auto" is the single source of truth for recommendations — don't maintain a parallel list here.
/prompt [query] — recommend a technique for the intent, then apply
/prompt [technique] [query] — apply a specific technique
/prompt refine [query] — opt-in: rewrite the raw query into a brief via refine_prompt
/prompt list — the ~9 core techniques (default)
/prompt all — all 88 techniques
/prompt [category] — one category in full/prompt list, /prompt all, /prompt [category]list_prompt_techniques({}) — core techniques (the ones that still help 2026 reasoning models).list_prompt_techniques({ all: true }).list_prompt_techniques({ filter: "[category]" }) — categories: creative, research, research_advanced, analytical, reflective, reasoning, verification, meta, debate, judgment, engineering, structured_coding, decision, retrieval, security, few_shot, tabular, writing./prompt [query] — recommend, then applypreview_prompt_technique({ technique: "auto", query: "[query]" }). It returns the top techniques for the intent with a reason and a kind tag (contract = imposes output structure, safe on reasoning models; scaffold = mainly helps weaker/routed models).1b. Rank with Jev when the keyword match is weak. auto is regex-based, so it can miss intent or return its generic fallback (chain_of_note / astute_rag / spotlighting for a query that isn't about retrieval). When that happens, or when its top picks look like a coin-flip, rank the candidates with one near-free typed call (~$0.00002):
mcp__tachibot-mcp__jev_decide({
state: "[the user's query, verbatim]",
questions: {
technique: { type: "choice", instructions: "Which thinking technique best fits this task?",
criteria: { "[name]": "[its one-line why from auto / list_prompt_techniques]", "...": "..." } }
}
})Candidates: auto's picks plus 2–4 plausible ones from list_prompt_techniques({}). Use 3–7 options, listed in alphabetical order, never in auto's rank order and never marked as auto's pick, or Jev just echoes the ranking. Order the offer in step 4 by answers.technique.probabilities. If Jev errors, keep auto's order.
refine_prompt({ query, goal?, context? }) — relay its REFINED PROMPT + WHAT CHANGED + OPEN QUESTIONS and get approval before using the brief. Never call it automatically.preview_prompt_technique({ technique: "[top pick]", tool: "[tool]", query: "[the brief]" }).execute_prompt_technique({ execution_token: "last" }) (re-preview first if they picked a runner-up)./prompt [technique] [query] — apply a named techniquepreview_prompt_technique({ technique: "[technique]", tool: "[tool]", query: "[query]" }) — show original vs enhanced.execute_prompt_technique({ execution_token: "last" }) (tokens expire after a few minutes — re-preview if stale)./prompt refine [query] — opt-in prompt rewriterefine_prompt({ query }) directly (add goal/context if the user gave them) and relay REFINED PROMPT and WHAT CHANGED verbatim. This is the explicit, user-requested path — unlike step 2 above, it always calls the tool since the user asked for it by name.preview_prompt_technique({ technique: "auto", query: "<final brief>" }) or straight to the tool/model of choice.Match the target tool to the technique's category:
| Category | Best tool | Fallback |
|---|---|---|
| Creative | gemini_brainstorm | grok_brainstorm |
| Research | perplexity_ask | grok_search |
| Analytical | grok_reason | openai_reason |
| Reflective | gemini_analyze_text | kimi_thinking |
| Reasoning | kimi_thinking | openai_reason |
| Verification | openai_reason | grok_reason |
| Debate / Decision | grok_reason | openai_reason |
| Judgment | gemini_judge | jury |
| Engineering / Structured Coding | qwen_coder | minimax_code |
| Research Advanced | perplexity_reason | kimi_long_context |
| Writing | kimi_thinking | gemini_analyze_text — or skip previewing and call blog_writer directly for the one-call researched-draft path (storm + re3 composed) |
/prompt there's a null-pointer crash on checkout — auto recommends rubber_duck / react/prompt should we use Redis or Postgres for the queue — auto recommends adversarial / pre_mortem/prompt refactor the payment module safely — auto recommends pre_post / scot / test_driven/prompt scot implement a rate limiter — applies structured-CoT directly/prompt bdd_spec user registration flow — Given/When/Then before code/prompt list — the core techniques · /prompt all — all 88~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.