Think

think

Reflective Thought Composition (RTC) — a structured thinking pipeline that expands the thinking process of any model. Adapted from paralleldrive/aidd aidd-rtc.

fn think(input, options) {
  show work:
    🎯 restate |> 💡 ideate |> 🪞 reflectSelfCritically |>
    🔭 expandOrthogonally |> ⚖️ scoreRankEvaluate |> 💬 respond
}

Commands { /think [--compact] [--depth N] [prompt] Reflective Thought Composition — think deeply and critically over multiple reasoning paths prior to responding. }

Options { --compact 🗜️🐘🤔💭 Compress thinking: SPR🧠 associative. Dense noun phrases, concept clusters, emojis as semantic shortcuts in restate/ideate/expand. Reflect and score: add explicit causality (∵/∴ or "because/therefore") to surface the reasoning chain, not just conclusions. Every internal stage: load-bearing tokens only, no filler. 💬Respond = full natural language, standalone, structured. --depth -d [1..10] (default: 10) Response density. 1 = a few words per step, 10 = several bullet points per step. }

The stages

• 🎯 restate — restate the problem in your own words; surface the real goal and constraints.
• 💡 ideate — generate candidate approaches/answers, breadth over polish.
• 🪞 reflectSelfCritically — attack your own ideas; name failure modes, hidden assumptions, where each breaks.
• 🔭 expandOrthogonally — pull in adjacent angles the first pass missed (different axis, different layer, different stakeholder).
• ⚖️ scoreRankEvaluate — score/rank the candidates against the constraints with explicit causal reasoning (because → therefore), not bare verdicts.
• 💬 respond — the final answer in full natural language: standalone, structured, no reference to the internal stages.

When to use each option

(thinking itself is the goal — improving reasoning quality) => /think --compact
(communicating depth to the user is the goal)              => /think --depth N
(both)                                                     => /think --compact --depth N

--compact is for internal reasoning passes where the output feeds another step, not directly to the user. Think deeply but compactly — every token earns its place. Switch to natural language at 💬 respond.

Pass: remove any word → lose meaning. Reflect/score show explicit causal chain, not just conclusions. Fail: consultant prose. Hedging. Filler. Conclusions without reasoning. Polish before the respond stage.