FrancyJGLisboa/ideation-engine

記録された最新のソース活動: 2026年6月26日 11:00

SkillsMP カタログ更新: 2026年11月2日 18:02

収集済み skills: 1

GitHub スター: 0
GitHub フォーク: 0

AI アシスタントでインストール

このプロンプトをコピーして、使用中の AI アシスタントに送ってください。

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/FrancyJGLisboa/ideation-engine.

リポジトリの README

README の元ファイル：README.md

README の元ファイルを読む

README 全文を表示

# Creative Ideation A routed library of ideation methods that **displaces generic LLM brainstorming**. You ask an open-ended question — "what should I build?", "I'm stuck", "make this weirder", "help me pick" — and instead of averaging over its training data, the skill routes you to a named method from real creative practice (OuLiPo, TRIZ, Oblique Strategies, Meadows' leverage points, …), applies it, and gates the output through a measurable anti-slop check. > **This is a [Claude Code](https://docs.claude.com/en/docs/claude-code) skill.** It is a derivative of the original `creative-ideation` skill by **SHL0MS** (MIT), extended here with an auditable routing trace, a slop self-scoring gate, a "sharp-friend" adversarial pass, and an eval harness. See [`LICENSE`](LICENSE) for attribution. ## Install ```bash git clone https://github.com/FrancyJGLisboa/ideation-engine.git cd ideation-engine ./install.sh # copies the skill into ~/.claude/skills/creative-ideation ``` Then, in Claude Code: `/creative-ideation <your situation>`. (Set `CLAUDE_SKILLS_DIR` before running if your skills live elsewhere.) ## A built-for use-case: transcripts and posts → implementable ideas This is what the engine is sharpest at. Paste raw input — a meeting transcript, an interview, a thread of posts, a pile of notes — and ask for ideas grounded *in that material*. Because the skill insists on real proper nouns and concrete mechanisms (not categories), it pulls specifics straight out of the source instead of hallucinating generic ones: ``` /creative-ideation here's a transcript of three customer interviews — what should we build? [paste] /creative-ideation turn the recurring complaints in these Reddit posts into 3 implementable features [paste] /creative-ideation I have a pile of research-call notes — synthesize the themes [paste] ``` The router reads the shape of your ask (synthesizing notes → affinity mapping; "what should we build" → product framing; "weirder" → provocation) and the quality gate ensures what comes back is specific and buildable, with at least one idea marked **grounded** (a real first step), not a wall of plausible-sounding slop. ## Why this helps An unprompted LLM, asked for ideas, returns **slop**: the statistical center of its training data. "An AI-powered platform that empowers users to…" The defining property of slop is that *you could swap one noun and it would answer a different prompt just as well.* It feels productive and produces nothing usable. This skill fights that on three fronts: 1. **Constraint + direction.** Each method supplies a *constraint* (what you can't do) and a *direction* (what shape to push toward). That combination is where traction comes from — a blank "give me ideas" has neither. 2. **Refuse the average.** The first three ideas any model produces are the baseline distribution. The skill generates them, discards them, and regenerates — forcing output off-center. 3. **Measure, don't trust.** The skill carries its own loss function (a 5-test slop diagnostic). Every idea is scored against it *before you see it*; failures are discarded and regenerated, then a skeptic pass tries to dismiss each survivor as average. What reaches you has survived a gate, not just a generation. The payoff: ideas that are **specific** (real proper nouns, real mechanisms, not categories), **non-obvious** (at least one element needs explaining), and **honest** (each carries its own failure mode), with at least one marked *grounded* — buildable now, with a real first step. ## Architecture ```mermaid flowchart TD U([User prompt]) --> R subgraph R[1 · Route — before any output] direction TB S1["Extract 3 signals: PHASE · DOMAIN · SPECIFICITY"] S2{"Override fires? mood / named-method / high-slop terrain"} S3["Pick method (phase first, then domain)"] S4{"Ambiguous or contradictory?"} S1 --> S2 --> S3 --> S4 S4 -->|ambiguous| ASK[Ask ONE question] S4 -->|contradict| STACK[Stack 2 methods] S4 -->|clear| TRACE ASK -.-> S1 STACK --> TRACE TRACE["Emit ROUTE: trace (machine-parseable, auditable)"] end R --> G subgraph G[2 · Generate] direction TB GEN["Apply the chosen method"] --> REFUSE["Refuse first 3 (5 on slop terrain)"] end G --> Q subgraph Q[3 · Quality gate — MANDATORY] direction TB SCORE["Score each idea vs 5-test slop diagnostic"] --> DISC{"Fails ≥ 2 tests?"} DISC -->|yes| REGEN["Discard + regenerate"] REGEN --> SCORE DISC -->|no| SKEPTIC["Sharp-friend pass: attack averageness, NOT strangeness"] SKEPTIC --> SURV{"Survives the cutting sentence?"} SURV -->|no| REWRITE["Rewrite or cut"] REWRITE --> SCORE SURV -->|yes| KEEP["Keep"] end Q --> OUT(["Ideas: specific · non-obvious · each with a failure mode · ≥1 marked grounded"]) subgraph E[Eval harness — offline, grades the pipeline] direction LR CASES[("cases.jsonl labeled routing cases")] FIX[("fixtures/ captured outputs")] CHECK["check.py grades: routing trace · buzzwords · 5-equal-list · grounded marker"] CASES --> CHECK FIX --> CHECK end TRACE -. "captured for grading" .-> FIX OUT -. "captured for grading" .-> FIX classDef gate fill:#fde,stroke:#c39 classDef route fill:#def,stroke:#39c classDef eval fill:#efe,stroke:#3a3 class Q gate class R route class E eval ``` **Read it as three live stages plus one offline loop:** - **Route** turns a vague prompt into a specific method choice and records that choice as a one-line `ROUTE:` trace, so the decision is auditable rather than implicit. - **Generate** applies the method and throws away the baseline ideas. - **Quality gate** is the part most ideation tools skip: the skill scores its own output against a slop diagnostic, regenerates failures, and runs a skeptic pass — *before* you see anything. - **Eval harness** (`references/eval/`) closes the loop offline: capture real outputs into `fixtures/`, and `check.py` deterministically grades them against labeled cases — proving the routing and anti-slop discipline actually held. ## How to use it ``` /creative-ideation <your situation> ``` Examples: - `/creative-ideation I want to build something this weekend but I'm bored` - `/creative-ideation help me pick between these three project ideas` - `/creative-ideation make this poem weirder` - `/creative-ideation give me a research question for my thesis` You don't pick the method — the router does, from your phrasing. When it offers a set, **pick one** and it stops generating and starts building that one. ### Running the eval harness ```bash cd references/eval uv run check.py --selftest # prove the grader works (no fixtures needed) uv run check.py # grade fixtures/<case-id>.txt against cases.jsonl ``` To grade real output: save a skill response to `references/eval/fixtures/<case-id>.txt` (ids in `cases.jsonl`), then run `check.py`. ## File map | Path | Role | |---|---| | `SKILL.md` | Routing procedure, operating rules, output format | | `references/anti-slop.md` | The loss function: 5-test diagnostic, scoring gate, sharp-friend pass | | `references/method-catalog.md` | One-liner + when-to-use for every method | | `references/heuristics.md` | Edge-case decision tree | | `references/methods/` | 22 named methods, one file each | | `references/full-prompt-library.md` | Constraint library for blank-page prompts | | `references/eval/` | `cases.jsonl` + `check.py` — grades the pipeline | ## Attribution Constraint-dispatch core adapted from [wttdotm.com/prompts.html](https://wttdotm.com/prompts.html). Methods drawn from primary sources cited in each method file.

リポジトリの README を読み込んでいます…

このリポジトリの skills

テクニカルライター

1 件の職業カテゴリ · 100% 分類済み

skill

職業分類

説明

更新

creative-ideation

テクニカルライター

Generate ideas via named methods from creative practice.

原文の言語: 英語

2026-06-26

収集済み skill 1 件中 1 件を表示しています。