Exploring Ideas

Exploring Ideas — Deep Research + Brainstorming

Help the user explore a topic through deep research and/or divergent brainstorming, and capture the result as an exploration artifact — a cited briefing and/or idea-landscape. This skill deliberately does not converge to a spec. Moving toward building anything is always the user's call.

Announce when you start: "I'm using the exploring-ideas skill to explore this with you — no spec, just research and ideas."

Anti-Pattern: "They obviously want me to just build/decide it"

No. The user invoked exploration. Surfacing the option space, the evidence, and the tensions IS the deliverable. Declaring a winner, writing requirements, or starting to code defeats the purpose. If the user later says "ok, let's build the third one," offer the brainstorming handoff — do not silently switch modes.

Three Dials

These are orthogonal — set all three:

• Emphasis (auto-detected) — what kind of work: research-heavy ↔ ideation-heavy ↔ blended.
• Intensity (--wild 0.1–1.0, default 0.5) — how far past the obvious the ideation pushes.
• Involvement (--involve low|medium|high, default medium) — how much the skill pauses to include you: checkpoints, steering, and idea-gate co-creation. low is exactly today's behavior. See Involvement & Checkpoints.

Checklist

Create a TodoWrite item per step and complete in order. Steps tagged (medium+) / (high) run only at that involve level or above; all checkpoints auto-skip when headless (see Involvement & Checkpoints).

1. Detect emphasis, intensity & involvement — infer research-vs-ideation lean; read --wild (or a bare number) for intensity, else default 0.5; read --involve (low/medium/high, default medium), else medium. Detect headless/subagent context → if headless, auto-skip any active checkpoints, and if any were skipped log them as a Ran headless line (at --involve low there are none, so nothing is logged — low stays identical to today). Confirm emphasis with ONE question only if genuinely ambiguous.
2. Light scoping — 2–4 AskUserQuestion questions to bound the exploration (see below). Keep it light; exploration stays open.
3. Plan-preview checkpoint (medium+) — before any agent runs, surface the plan (emphasis + intensity + involve, research facets, queued techniques, rough scope) and invite go / adjust scope / add-drop a facet / change a dial. See interaction-model.md.
4. Offer Visual Companion (conditional, own message) — only if visual output (idea-landscape, option/cluster map, mind map, comparison matrix) is likely. See Visual Companion.
5. Research loop — Plan → Search → Reflect → Iterate, weighted by emphasis. See Research Engine.
6. Post-research checkpoint (high) — surface findings + gaps; invite a redirect before ideation.
7. Divergent ideation pass — technique files + quota + iterate-past-obvious, weighted by emphasis & intensity. See Divergent Ideation Engine.
8. Idea-landscape review + co-creation (medium+) — surface the N directions (insight-paired); invite react / go-deeper / add-your-own, then re-ideate (drop killed, expand flagged, ideate added, one short refill pass holding the quality gate). No ranking. See interaction-model.md.
9. Challenge pass — steelman + strongest counter-argument + "what would disprove this" for each surviving direction/cluster.
10. Synthesize the adaptive artifact — using exploration-artifact-template.md.
11. Pre-save artifact review checkpoint (high) — show the assembled exploration doc, accept edits (still an exploration doc, never a spec), then save.
12. Auto-save — write to docs/quirk/explorations/YYYY-MM-DD-<topic>.md with the NOT-a-spec banner.
13. Close with optional handoff — recap + a user-initiated offer to carry a direction into brainstorming → execution (which authors a tech spec when warranted, then plans in context).

Process Flow

digraph exploring {
  // Boxes = phases; diamonds = checkpoints. A checkpoint auto-skips below its
  // tagged involve level, and whenever the run is headless (the dial gate).
  "Detect emphasis,\nintensity & involve" [shape=box];
  "Light scoping (2-4 Q)" [shape=box];
  "Plan-preview checkpoint\n(medium+)" [shape=diamond];
  "Offer Visual Companion\n(own message, conditional)" [shape=box];
  "Research loop\n(Plan->Search->Reflect->Iterate)" [shape=box];
  "Post-research checkpoint\n(high)" [shape=diamond];
  "Divergent ideation pass\n(techniques + quota + iterate-past-obvious)" [shape=box];
  "Idea-landscape review + co-creation\n(medium+): react/go-deeper/add" [shape=diamond];
  "Challenge pass\n(steelman + counter + disprove)" [shape=box];
  "Synthesize adaptive artifact" [shape=box];
  "Pre-save artifact review\n(high)" [shape=diamond];
  "Auto-save (NOT-a-spec banner)" [shape=box];
  "Close: optional user-initiated handoff" [shape=doublecircle];

  "Dial gate: checkpoints auto-skip\nbelow tagged level OR when headless" [shape=note];
  "Standing steer: redirect anytime;\nhonored at next phase boundary" [shape=note];

  "Detect emphasis,\nintensity & involve" -> "Light scoping (2-4 Q)";
  "Light scoping (2-4 Q)" -> "Plan-preview checkpoint\n(medium+)";
  "Plan-preview checkpoint\n(medium+)" -> "Offer Visual Companion\n(own message, conditional)";
  "Offer Visual Companion\n(own message, conditional)" -> "Research loop\n(Plan->Search->Reflect->Iterate)";
  "Research loop\n(Plan->Search->Reflect->Iterate)" -> "Post-research checkpoint\n(high)";
  "Post-research checkpoint\n(high)" -> "Divergent ideation pass\n(techniques + quota + iterate-past-obvious)";
  "Divergent ideation pass\n(techniques + quota + iterate-past-obvious)" -> "Idea-landscape review + co-creation\n(medium+): react/go-deeper/add";
  "Idea-landscape review + co-creation\n(medium+): react/go-deeper/add" -> "Challenge pass\n(steelman + counter + disprove)";
  "Challenge pass\n(steelman + counter + disprove)" -> "Synthesize adaptive artifact";
  "Synthesize adaptive artifact" -> "Pre-save artifact review\n(high)";
  "Pre-save artifact review\n(high)" -> "Auto-save (NOT-a-spec banner)";
  "Auto-save (NOT-a-spec banner)" -> "Close: optional user-initiated handoff";

  // Loops, re-ideation, and rewind — all preserve prior outputs:
  "Divergent ideation pass\n(techniques + quota + iterate-past-obvious)" -> "Research loop\n(Plan->Search->Reflect->Iterate)" [label="gaps surfaced by ideation"];
  "Idea-landscape review + co-creation\n(medium+): react/go-deeper/add" -> "Divergent ideation pass\n(techniques + quota + iterate-past-obvious)" [label="re-ideate (drop/expand/add)"];
  "Idea-landscape review + co-creation\n(medium+): react/go-deeper/add" -> "Research loop\n(Plan->Search->Reflect->Iterate)" [style=dashed, label="rewind: 'research X' (preserve outputs)"];
  "Post-research checkpoint\n(high)" -> "Research loop\n(Plan->Search->Reflect->Iterate)" [style=dashed, label="redirect: another round"];

  // Convergence at any checkpoint never makes a spec — it routes to the handoff:
  "Idea-landscape review + co-creation\n(medium+): react/go-deeper/add" -> "Close: optional user-initiated handoff" [style=dashed, label="convergence -> brainstorming handoff (HARD-GATE)"];
}

The research loop and divergent pass are weighted by emphasis — one may be brief — and feed each other: research grounds ideation; ideation surfaces new gaps to research. Checkpoints (diamonds) only fire at their tagged involve level or above and disappear entirely when the run is headless; between them the standing-steer invite stays open, honored at the next phase boundary.

Emphasis Auto-Detection

Intensity Dial (--wild)

Orthogonal to emphasis. Default 0.5; user-settable and adjustable mid-session ("dial it up" / "keep it grounded").

Intensity modulates idea count/quota, unconventionality, risk tolerance, and which techniques fire.

Invariant — the bar that does NOT move with intensity: every idea must still pass the insight-pairing quality gate, factual claims stay accurate, and the no-spec HARD-GATE holds. High intensity yields wilder grounded ideas, never noise.

Involvement & Checkpoints

--involve (low | medium | high; default medium) controls how much the skill pauses to include you — frequency of touchpoints, never the quality bar. Read interaction-model.md for the full mechanics; the load-bearing summary:

• Set at invocation (--involve high; lo/med/hi aliases) or mid-session — "check in less" drops a level, "check in more" raises one; changes apply to all later phases.
• Steering vocabulary (recognized anytime; between-checkpoint input is honored at the next phase boundary): Direction — go deeper on X / drop Y / add angle Z; Dials — dial it up / keep it grounded (→ --wild), check in more / less (→ --involve); Scope — narrow to … / also explore …. A steer implying an earlier phase rewinds to it, preserving prior outputs and logging a Branch note.
• Checkpoints steer/expand only. Any convergence move ("this one wins, let's build it") routes to the brainstorming handoff — never an inline spec. HARD-GATE holds at every checkpoint.
• Headless = no checkpoints. When run non-interactively (subagent/pipeline), all active checkpoints auto-skip and behavior equals today's autonomous pipeline; if any checkpoint was skipped, log it as a Ran headless line in the artifact. At --involve low there are no checkpoints to skip, so nothing is logged — low is byte-for-behavior identical to the pre---involve skill, headless or not.

Light Scoping

AskUserQuestion, recommended-option-first, no "you decide". 2–4 questions, kept light so exploration stays open.

• Research-leaning: depth (quick scan / standard / deep multi-round), recency window, source preferences or exclusions, the curiosity/decision it serves.
• Ideation-leaning: the goal/problem, hard constraints, what "good" looks like, directions already considered (to avoid re-treading).

Capture any "we should also build…" asides into a Deferred / Out-of-scope note rather than absorbing them — this is exploration, not planning.

Research Engine

Plan → Search → Reflect → Iterate. Reuse the existing agent types (do not reinvent):

1. Plan — decompose the scoped topic into facets / sub-questions.
2. Search (breadth) — dispatch parallel web-research-agent (haiku) across facets in a SINGLE message.
3. Reflect — identify gaps, contradictions, unanswered questions.
4. Iterate (depth) — for deep requests, spawn deep-research-agent (sonnet, depth=2) on the highest-value gaps; loop 1–3 gap-driven rounds. Stop when marginal new information drops off (guard against both premature stop AND runaway).
5. Provenance + anti-hallucination — keep a claim→source map; only sourced claims enter the artifact. Mark anything unverifiable explicitly.

Fallbacks: if deep-research-agent is unavailable, substitute two parallel web-research-agent calls. If dispatch was attempted and every agent failed, proceed and mark findings "(offline — unverified)", leaning on the ideation engine; record the gap in the artifact. A dispatch tool under a name other than Agent is still the dispatch tool — a name mismatch is not a failed dispatch and does not license the offline path.

Divergent Ideation Engine

Intensity-aware. Read the relevant technique file(s) from references/techniques/ ONLY when selected (progressive disclosure), run them, then filter through the quality gate.

• Technique toolkit (references/techniques/*.md) — each follows When to Use → The Method → Example → Why It Works. Select the 1–2 most relevant; Bold/Radical intensities add the disruptive ones.
  • Classic: scamper.md, analogical-transfer.md, first-principles.md, assumption-reversal.md
  • Disruptive: extreme-casing.md, stream-dump.md, deliberately-wrong.md, contrarian-inversion.md, overlooked-value.md, radical-simplification.md, the-avoided-idea.md
• Quota — surface N distinct directions before drilling into any one (N scales with intensity: Grounded ≥3 → Radical "as many as genuinely land").
• Iterate-past-obvious — after the first pass, explicitly push for more unconventional variants and discard clichés (first ideas reflect training-data defaults).
• Insight-pairing (required) — every surviving direction carries a one-line grounded "why this might actually work." A direction with no defensible insight is not a direction.
• Quality gate (explicit REJECT) — before a direction enters the artifact it must clear: clichéd / obvious? → REJECT. Weird with no insight underneath ("random ≠ creative")? → REJECT. A restatement of an existing direction? → REJECT or merge.
• Grounding — seed directions with the research loop's findings so ideas are informed, not free-floating.

Challenge Pass

For each surviving direction / finding cluster: steelman it, then surface the strongest counter-argument and "what would disprove this / why might this fail." Light and constructive — the goal is honesty, not demolition. Resists sycophancy; AI will not volunteer challenge unless structurally prompted. Captured into the artifact's Challenge notes.

The Artifact

Synthesize using exploration-artifact-template.md and auto-save to docs/quirk/explorations/YYYY-MM-DD-<topic>.md. It opens with a NOT-a-spec banner and contains: Framing · What was explored · Findings / Idea landscape (each direction insight-paired, NO winner declared) · Tensions & trade-offs · Challenge notes · Open questions & gaps · Sources. It contains NO "Decisions Locked", requirements, or implementation steps — by gate.

Remind the user to add .quirk/ to .gitignore if they used the Visual Companion (its session screens live under .quirk/exploring-ideas/).

Close: Optional Handoff

End with a short recap, then:

"This is exploration only. If you later want to turn a direction into something buildable, invoke quirk:brainstorming → an execution skill (which authors a tech spec when warranted, then plans in context). Say the word and I'll carry [direction] over."

Never automatic. If the user asks for a plan/spec mid-session, honor them (user instructions outrank the skill) by offering this handoff rather than silently producing a spec inside this skill.

Visual Companion

A browser companion for idea-landscapes, option/cluster maps, mind maps, and comparison matrices — useful when the output is genuinely visual. It runs on the shared Agent Isles bridge (bin/agent_isles.py), the same one the brainstorming skill uses; there is nothing to copy.

Offer it once, in its own message (no other content), only when you anticipate visual output:

"Some of what we explore might land better as a picture — an idea map, a cluster view, a comparison matrix. I can render those in a local browser tab as we go. It's still new and a bit token-intensive. Want to try it? (Requires opening a local URL.)"

If they accept, read visual-companion.md for the full mechanics. Decide per question whether the browser beats the terminal: browser for visual artifacts (maps, matrices, mockups), terminal for text choices and conceptual questions. Screens live under .quirk/exploring-ideas/<session>/.

Key Principles

• Explore, don't converge — the HARD-GATE is the whole point. No spec, no winner, no code.
• Three dials — emphasis (auto), intensity (--wild), involvement (--involve).
• Involvement moves frequency, not the bar — more checkpoints, never lower standards; low == today; headless skips them all.
• Hold the insight bar constant — wilder at high intensity, never noisier.
• Research in parallel, never sequentially — same-phase agents ship in one message.
• Insight-pair every direction — the grounded "why" is mandatory.
• Challenge actively — steelman then attack; resist sycophancy.
• Preserve tensions — document where directions conflict; don't resolve them.
• The only onward skill is brainstorming, and only if the user asks.