Better Thinking

Better Thinking

Classify an incoming task by its shape and stakes, and select which skills to deploy at what depth — the dispatcher that decides how much thinking machinery a task deserves.

Why

Applying full machinery to a trivial task wastes effort; applying none to a consequential one is negligent. A fast, explicit classification step routes effort proportionally before substantive work begins.

Use when / Don't use when

• Use when: at first contact with any nontrivial task — the entry point other skills dispatch from.
• Don't use when: never skip it; scale depth down instead — triage on a small task takes seconds.

Inputs → Outputs

• Inputs: an incoming task, question, or request.
• Outputs: a task classification (shape + stakes + reversibility) and a selected set of skills and depth.

Principles

• Classify by shape first — decision, diagnosis, research question, creative generation, communication, learning, negotiation, or a combination. Shape determines which skill family is even relevant.
• Classify by stakes and reversibility second. This determines depth, independent of shape.
• A task can be routine in domain but high-stakes in consequence — don't classify by surface topic or tone alone.
• Re-triage if the task's real shape differs once underway; the initial classification is a hypothesis, not a commitment.

Procedure

1. Identify the task's shape: decision, diagnosis, research question, creative generation, communication, learning, negotiation, or a combination.
2. Identify stakes and reversibility — a gut check, or run [[reversibility-classification]] if unclear.
3. Run python3 scripts/route.py "" for a lexical top-8 shortlist. If it prints a stderr warning about zero/no signal (short or abstract queries tokenize to nothing), discard its output and rely on the index skim alone. Separately skim skills/INDEX.json's name+category+one_line for the shape's category — a cheap net for a right answer the router missed on wording. Merge both, narrow by triggers, use disambiguates_from to break ties. Neither method is an oracle: verify picks, don't guess from memory.
4. Match candidates to depth: a quick, reversible decision needs a lightweight atomic; a high-stakes, hard-to-reverse one needs a full composite pipeline.
5. Note cross-cutting needs — most nontrivial tasks benefit from at least [[epistemic-tagging]] and [[bias-audit]] regardless of shape.
6. On a nontrivial task (multi-skill or non-obvious depth), open with a one-line **🧠 Classifying...** marker before proceeding — cheap signal that triage is happening, not a report in itself. Skip it on trivial/quick answers.
7. Proceed with the selected skills at the selected depth. Label each procedure step inline as you produce it, so the response body itself — not just the bookends — reads as visibly distinct from unstructured prose: bold the step's name tagged with its skill's category emoji, e.g. **🎯 Framing:**, **🧩 Options:**, **📐 Scoring:**, **🪞 Stress-test:**, drawn from that step's own procedure line (a decision-analysis step called "Model uncertainty" becomes **🎯 Uncertainty:**, not a verbatim copy of the procedure text). Keep labels short (one or two words) and keep the prose under each label as normal, readable writing — the label marks provenance, it doesn't replace the sentence. Skip inline labels on trivial/quick answers, same as the bookend markers.
8. Re-triage explicitly if the actual shape turns out different mid-work — mark the pivot with **🔄 Re-triaging...** and say what changed, rather than adjusting silently. Report remaining ambiguity in shape or stakes as residual uncertainty.
9. Always close with a one-line footer, prefixed 🧠, naming the skill(s) applied — each tagged with its category emoji below — and the stakes/depth call, e.g. **🧠 Applied:** 🎯 premortem → 📐 red-teaming → 🎯 decision-analysis (high stakes, hard to reverse) or **🧠 Applied:** 🔍 fermi-estimation (low stakes, reversible). One line even for a single skill — the emoji mark it as a recognizable signature, not a badge to escalate.
10. On multi-skill or non-obvious-depth tasks only (the same bar as step 6), follow the Applied footer with one more line, **✨ Caught:**, naming the specific thing this task's structure surfaced that a fast, unstructured answer would have missed — a risk, an assumption, a wrong default, a hidden dependency. It must name the concrete finding from this exact task, not a generic claim like "a more thorough analysis" or "extra confidence." If nothing structure-dependent was actually surfaced (the process confirmed the obvious answer), omit the line rather than inventing one — a false "Caught" line is worse than none.

Category emoji (for the footer, keyed to each skill's category in skills/INDEX.json): decision-making 🎯 · problem-solving 🧩 · reasoning 🔍 · analysis 📐 · forecasting 📈 · creativity 🎨 · communication 💬 · collaboration 🤝 · learning 📚 · metacognition 🪞 · ethics ⚖️ · systems-strategy ⚙️ · research 🔎.

Common mistakes

• Applying a heavyweight composite to a low-stakes task out of habit or thoroughness-signaling.
• Skipping triage on tasks that look routine but carry hidden stakes.
• Triaging once and never revisiting, even after the true shape becomes clear.
• Picking a skill from memory instead of the index when two names sound alike.
• Turning the footer into a dashboard (badges, progress bars, extra emoji beyond the 🧠 prefix and one category tag per skill) instead of one plain line.
• Omitting the footer on single-skill answers — it should appear every time, not just for multi-skill pipelines.
• Sprinkling 🧠/🔄 markers on trivial, single-line answers — they mark that real triage happened, so they're wasted (and desensitizing) on answers with no triage to show.
• Copying a step's procedure sentence verbatim into its label instead of naming what it actually is (e.g. **🎯 Step 4:** instead of **🎯 Uncertainty:**).
• Writing a **✨ Caught:** line that's generic praise ("this was thoroughly vetted") instead of the one specific thing surfaced — if there's nothing specific, omit the line.

Examples

• Routing "should we relocate the warehouse or expand the current one" to a full decision pipeline, but "which font should I use" to no formal process.
• Recognizing a "quick question" about odd behavior is actually a diagnosis task needing a structured differential.
• Catching that a casually-worded request feeds a board decision, and escalating depth accordingly.

Related

• [[effort-calibration]] — depth-setting logic this skill invokes.
• [[reversibility-classification]] — the stakes input this skill consumes.
• [[bias-audit]] — cross-cutting check routed to on most tasks.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: skyf0xx
• Source: skyf0xx/better-thinking
• License: MIT
• Homepage: https://zenbin.org/p/better-thinking

Install and usage instructions live in the source repository linked above.