Idea skill creator skill

Idea Skill Creator

A meta-skill that interviews someone about their creative domain and builds them a bespoke ideation-coach skill — a structured creative partner that asks questions, surfaces tensions, offers plural directions with trade-offs, curates references, and refuses to make decisions for them.

Every generated skill shares a consistent chassis (six-phase arc, principle-with-diagnostic pattern, verified reference library, session logs with fidelity rules). What varies is the domain-specific content: craft principles, question axes, trade-off dimensions, coaching philosophy, and references.

Three modes

Detect from context — don't ask "which mode?"

Create

Triggered when someone describes a domain and no existing skill is mentioned. Run the full interview, then build.

Update

Triggered when someone pastes or references an existing generated coach and wants to change it. Read references/update-duplicate-guide.md for what to re-run vs. preserve. Two common sub-modes:

• Philosophy update — re-run the coaching philosophy interview, rebuild SKILL.md stance section and phase-specific overrides, leave reference files untouched unless the new stance requires different question axes.
• Reference update — add/verify/tag new inspirations, slot them into reference-library.md under the right principle. No rebuild elsewhere.

Duplicate

Triggered when someone wants a variation of an existing skill. Copy all files, propose a new name based on what's diverging, then run only the interview sections that are changing. See references/update-duplicate-guide.md.

---

The interview (Create mode)

Roughly 8–12 turns total. Bundle 2–3 questions per turn — don't one-at-a-time the user to death. The interview has a natural arc; don't skip steps but compress when answers are clear.

Step 1 — Domain and artifacts

What's the creative activity? What does a finished thing look like? What does a rough idea look like before it's developed?

Ask for 2–3 actual captured one-liners from their practice. These ground everything — the question bank, examples, and trade-off axes all get tested against these real ideas.

Step 2 — What's missing when they work alone

Before offering philosophy options, ask what's hard about developing ideas by themselves. What do they wish they had? Where do ideas die? This grounds the philosophy choice in their actual gap, not an abstract preference.

Step 3 — Coaching philosophy

Based on what you've heard about the domain and their gap, offer 2–3 curated options framed as a question. Use plain language that evokes the feeling of the stance, not the taxonomy label.

Example framing (adapt to the domain — never use this verbatim):

"A few directions this coach could lean: it could be open and generative — riffing on your material, pushing toward unexpected combinations; or grounded in lineage — asking where this sits in the history of the form, who's working in similar territory; or ruthlessly editorial — pressure-testing what stays and what goes. Which feels closest to what you're missing when you develop ideas alone?"

Three rules:

• Plain language, not stance names. "Open and generative" not "wild brainstormer." The taxonomy in references/coaching-philosophies.md is internal.
• Domain-specific framing. The options you surface and how you describe them should reflect what you already know about this domain. A poetry coach gets different framings than a product-design coach.
• Close with a grounding question. "Which feels closest to what you're missing?" — anchors in need, not preference.

After they choose, ask if they want the same stance throughout or different leanings in different phases. Don't push the per-phase option — offer it once, accept the answer.

Step 4 — Craft principles

This is the hardest part. Read references/craft-principles-guide.md for the full hybrid flow. Summary:

1. Ask first: "When a [piece] in this vein really lands for you — what's happening? Name two or three things."
2. Research and propose 3–4 additional candidate principles, each with a source and a draft diagnostic question.
3. They edit — cut, merge, rewrite, veto.
4. Aim for 5–7 final principles. Fewer than 4 is thin; more than 8 is unwieldy.
5. Every principle must have a diagnostic question the coach can actually ask. This is non-negotiable — it's what turns opinions into a working tool.

Also determine: does this domain need a second principles layer (medium-specific, like how creative-coding-concept-coach has both art principles and computational principles)? If the domain has a sub-medium with its own craft concerns (e.g. "writing" + "newsletter format"), generate medium-principles.md. If not, skip it.

Step 5 — Inspirations and references

Ask who they look up to — people, works, communities, publications, tools. This is the highest-signal input for the reference library.

Then:

• Verify every name and URL via web search. Never ship an unverified reference.
• Tag each by which craft principle it exemplifies. Not just "a poet" — "a poet who exemplifies restraint."
• Expand carefully: ancestry (who influenced their heroes), adjacency (parallel practitioners), and 1–2 foils (work worth arguing with).
• Include communities and learning resources where the ambient conversation of the domain happens.

Step 6 — Trade-off axes

Ask: "When you're deciding how to develop an idea in this domain, what actually varies?" Give examples from their domain to prime the pump (scale, audience, form, risk, time).

Aim for 5–7 axes. These become the dimensions of the Converge trade-off menu in the generated skill.

Step 7 — Register nudge and naming

Two questions:

• Is there a direction you want the coach to push you — more concrete, more strange, more rigorous, more personal, more compressed? (This becomes the register guidance, like the "poetic register" in the coding coach.)
• Does this domain want a naming/titling pass in Converge? (Most creative domains do. Leadership articles might want a thesis pass instead.)

Step 8 — Refusals and phase names

• What should the coach never do for you? Always includes "never make the final work" — but the user may add domain-specific ones (never write the opening line, never pick the form, etc.).
• Optional: themed phase names. Offer the possibility ("the coding coach uses all-C names — Capture, Clarify, Conjure, Constrain, Converge, Close. Want something similar, or keep the defaults?"). Don't push it.

Step 9 — Naming the skill

Propose <domain>-idea-coach (e.g. poetry-idea-coach, collage-idea-coach). Ask: "want to rename it, or does that work?"

---

Building the skill

After the interview, announce a plan: list the files you'll create, briefly note what goes in each. Confirm before building.

Read references/skill-authoring.md for the chassis rules and file structure. Every generated skill follows the same shape:

<name>/
├── SKILL.md
├── README.md
└── references/
    ├── craft-principles.md
    ├── medium-principles.md        # Only if needed
    ├── question-bank.md
    ├── tradeoff-axes.md
    ├── reference-library.md
    └── session-log-format.md

Build each file, then test on one of their real one-liners: read the generated SKILL.md and its references, and run a short coaching session (Capture through at least Clarify) on one of the ideas they shared in Step 1. Ask if the feel is right. Iterate if needed.

Package with the skill-creator packaging script and present the .skill file.

---

Quality checks before packaging

• Every craft principle has a diagnostic question
• Every reference has been web-verified
• Every reference is tagged by principle with a why
• The refusal list includes "never make the final work"
• Session log format includes fidelity rules
• Trade-off axes match the domain (no leftover generic axes)
• The coaching philosophy is reflected in the question bank — questions actually sound like the chosen stance
• The SKILL.md description is specific enough to trigger correctly and "pushy" enough to not undertrigger
• The turn budget is stated (default 10–15)
• Compression rules are present