メニュー
Run a taste evaluation on a research proposal or technical direction. Determines whether to proceed, simplify, defer, or kill before committing research effort.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
A-Mem thought evolution — check which prior thoughts should be updated when a significant new insight arrives. Spawns a lightweight subagent to classify prior thoughts as UPDATE or NO_UPDATE, then applies revisions. Use during long reasoning sessions when you reach a synthesis, make a decision, or discover something that changes earlier assumptions. Triggers on "this changes what I thought earlier", "update prior reasoning", "evolve thoughts", or automatically on conclusion/synthesis thoughts in sessions >10 thoughts.
Cross-session knowledge retrieval from the Thoughtbox knowledge graph. Searches entities, traverses relations, retrieves observations, and synthesizes findings from past sessions. Use when you need to recall prior decisions, check what's already known about a topic, find related insights, or build on past work. Triggers on "what do we know about", "have we seen this before", "recall", "prior decisions about", "knowledge graph", or when starting work that might have prior context.
Capture significant learnings from the current work session. Structures insights for future sessions and updates agent memory.
Write or improve prompts for Codex using Anthropic's official best practices. Creates system prompts, agent prompts, tool descriptions, and MCP resource templates. Pass an existing prompt to improve it, or describe what you need to create one from scratch.
Generate architecture diagrams for a codebase subsystem or module. Explores source files and produces Mermaid diagrams in docs/.
Extract reusable principles and decision frameworks from accumulated experience. Use after significant work sessions, project milestones, or when you notice recurring patterns worth codifying.
| name | taste |
| description | Run a taste evaluation on a research proposal or technical direction. Determines whether to proceed, simplify, defer, or kill before committing research effort. |
| argument-hint | [proposal or question to evaluate] |
| user-invocable | true |
Evaluate this research direction: $ARGUMENTS
Based on the proposal, select the appropriate depth:
| Situation | Operations |
|---|---|
| Vague idea, needs sharpening | Compression test only |
| Specific technical proposal | Landscape + Dead-end estimation |
| Major resource commitment | Full pipeline (all 6 operations) |
| Stuck, need fresh angles | Cross-pollination only |
| Choosing between options | Prediction query only |
Compression Test — Force into: "We believe [X] because [Y], and if we're right, [Z] follows." If it can't survive compression, verdict is defer.
Landscape Assessment — Use WebSearch and Exa (web_search_exa) to find adjacent work, abandoned approaches, new capabilities. What exists? What changed recently?
Prediction Query — Simulate success and failure worlds. Is the direction informative under both outcomes?
Dead-End Estimation — What's the most likely failure mode? What's the time-to-signal (not time-to-completion)?
Simplicity Audit — Is there an 80%-value simpler version? Apply recursively.
Cross-Pollination — Check against other domains. Structural resonance = real signal.
If the verdict becomes clear at any step, stop. Don't run more operations just to fill out the report.
## Taste Evaluation: [Title]
**Verdict:** proceed | simplify | defer | kill
**Rationale:** [One sentence]
**Compression:** [Single-sentence formulation]
**Landscape:** [Position relative to existing work]
**Prediction:** Success -> [X] / Failure -> [Y]
**Time-to-signal:** [Estimate]
**Simplification:** [Cheaper version, if any]
**Cross-domain:** [Analogies, if checked]
**Next step:** [Concrete action]
Omit unchecked sections.