⚡ god-tier-mini-prompts

170 one-line power prompts for coding agents. Each sentence so dense every word does work.

Density beats length. One loaded sentence outperforms a page of instructions.

Why this exists

A single sentence of vectors so juicy each verb, noun, and adjective acts like a flavor bomb. A simple prompt that ends up being VERY impactful. Short prompts also get a fresh context window's full attention, so delegate the real hunting to subagents and keep the main thread narrow.

This repo is a living collection of those sentences: harvested reply-by-reply from the state-machine thread (136 replies, 193 quotes, plus the replies on the quotes that invited them), the threads it spawned, Prompt Snacks, Notchbird's Banger Prompts, Anthropic's official patterns, and the best engineering threads on X.

The skill stays lean on purpose: SKILL.md is prompts only, no credits, so agent context isn't bloated. Every author credit lives in this README.

🗂 The collection

Category	Count	What it steers
🧠 Meta	1 + rule	Let the agent pick its own prompts
🏛 Architecture & abstractions	18	Abstraction-hunting, system framing
🔍 State-machine audit kit	36 + full texts	The complete implicit-state-machine playbook
🐛 Debugging & root cause	29	Symptoms vs roots, failure modes, hunts
📐 Planning & alignment	30	Specs, grilling, handoffs, autonomy
✨ Code quality & refactor	19	Smaller diffs, TDD, elegant rewrites
✅ Verification & proof	8	Evidence, holdouts, ship decisions
👀 Review & PR	2	Two-axis reviews, PR shape
🧹 Agent hygiene	3	Transcripts, AGENTS.md, usage audits
📚 Learning & understanding	14	Mental models, diagrams, ELI5
🎛 Steering, autonomy & status	16	Morale, BOSS coda, updates, skills

Total: 170 prompts, plus 3 paste-ready expanded audits and one full phased-audit text.

🪄 The meta-prompt

Don't browse, delegate. Paste this with your task, idea, or code:

Consider all the mini-prompts in this skill, pick the ones that would
improve my idea/input/software the most, and apply them.

Narrow it per phase: "…pick the 3 best for planning / review / debugging this."

📦 Install

Drop SKILL.md into your agent's skills folder. That's the whole install.

Claude Code: ~/.claude/skills/god-tier-mini-prompts/SKILL.md

mkdir -p ~/.claude/skills/god-tier-mini-prompts
curl -sL https://raw.githubusercontent.com/Drix10/god-tier-mini-prompts/main/SKILL.md \
  -o ~/.claude/skills/god-tier-mini-prompts/SKILL.md

OpenAI Codex: ~/.codex/skills/god-tier-mini-prompts/SKILL.md

mkdir -p ~/.codex/skills/god-tier-mini-prompts
curl -sL https://raw.githubusercontent.com/Drix10/god-tier-mini-prompts/main/SKILL.md \
  -o ~/.codex/skills/god-tier-mini-prompts/SKILL.md

Cursor: reference it from .cursor/rules/ or paste prompts straight into Agent mode.

Then invoke by matching a task to a prompt, or run the meta-prompt and let the agent pick.

🙏 Sources & credits

Every prompt in the skill came from a human. Full credit table:

The state-machine thread (@steveruizok, Oct 2026; seed: "find all of the implicit state machines in my app")

Author	Contribution
@trevin	The expanded one-file audit (states, invariants, transition tables, retries + concurrency, prioritize, group, fix, keep updated)
@Nikzaaaaad	The 9 follow-ups (false screens, UI business logic, drifted copies, unenforced DB rules, races, swallowed errors, magic numbers, loose types, effects-as-events)
@flowforth_76	The full phased audit (discovery → verification → plan → implement)
@carpenterdroid	The refactor-directive mega-prompt (eliminate state, derive values, explicit machines, regression + integration checks)
@asmirkn	Transition tables; impossible-but-reachable transitions
@aloobhujiyan	Where two machines touch the same data (the race lives there); model-checker shortest-path
@Outgathr	"Which of these explain the bugs I've been calling flaky"
@liuzhao_666	Concrete event sequences; permitted vs reachable states
@eclarkhalid	Cite handlers + conditions; mark proposed changes separately
@kr4chinin	Visualize the state flow
@sjmurdoch	Verify correctness in Quint
@garbnzgh	Subagents + propose building it into reviews/workflows
@badbirb	Frame out stateful elements and plan before coding
@EternitiesAI	Failing regression test = boundary of suspected vs fixed
@Pps831	Global-rules prevention; no-implicit-state-machines rule
@feddiwhip	Unenforced invariants
@bonsaielectric	Implicit invariants
@peytoncasper	Non-terminal states variant
@jussisaur	The explicit inversion (turso)
@twid	"Find all tautologies"
@wahome_thuku	The legal port: implicit obligations in contracts
@jimlux	"Find the dumb requirements" (MCAS debate)
@cocobiskit	Adversarial red-teaming
@joshest3p	"Pushing determinism down"
@krimuthu	Autonomous fix loop ("keep me out of the loop")
@_benwick	"Just find the bugs"
@orcsrise	"What pattern allowed this bug to exist"
@udaysy	"What breaks if we revert this in 6 months"
@Shannoncode	Root-cause tokens; SDK/backend/schema context assembly
@menhguin	Scientific method; decision frameworks (Chesterton, OODA, autonomy levels)
@matt_dot_sol	Targeted "are you sure" double-checks
@PreciousBa82157	"If it's not in the prompt, it doesn't exist"
@bigdatachads	Research-first, then root-cause
@asavs_	Roleplay user / agent / system internals
@adityavg13	Internalize metaphors (allegory of the cave)
@impervium888	Explain it as a story
@codeby_ABI	Recursive prompt booster

The god-tier thread (@KingBootoshi, Nov 2026; the collection call)

Author	Contribution
@Wyrmfalls	The 5 abstractions (abstract thinking, irreducible constraints, principal architect, day-one test, higher abstraction)
@merterdir	Rabbithole check: is there a native way?
@Gtwy	The plan-reorganization spec-guard
@See_Elegance	Decisionmaker archetypes
@jp_aidev	Domain experts; current vs ideal; better version
@nemous02	The four learning lenses
@roanoke_gal	Perspective-shift trio
@veigapunk	Godspeed subagent block
@taskinfatih	Feedbacks-into-vectors
@PyrosAI	The meta-pick prompt
@dkubb	Illegal states unrepresentable; "call them spells"
@jmKoudsi	Kaggle-grandmaster lens
@CrudeAmericana	The "Robust" adjective
@antipdoom	"Make no mistakes" trio
@Abigail91187506	"Wow this codebase is shit"
@DrishtantGhosh	One-prompt code-rabbit critique
@TrillinDaGame	Complete working program, no fragments
@ChuhaiDev	Local minima vs greedy-path
@ybsgige	Time-discipline; "what's next after the list"
@observatii	"3 critical bugs" hunt
@nicholasdimoro	"Autosimplify"
@tunginton	"One-shot this"
@MattmreM	Senior-architect roadmap review
@ruyterfontes	Single sentence > skill pack
@alex98075wa	The repo-that-should-exist (this one)

Agent-operator threads (X)

Author	Contribution
@doodlestein	Agent-ergonomic planning pass; BOSS autonomy coda; oversights pass
@ManacasterBen	"Rocket-science-level approach"
@OnFinality	Failure modes per component
@janodev	Critique-until-converge
@itechnologynet	Human-in-the-loop + inspectable
@GreyforgeLabs	A/B experiments with benchmarks
@mikejt4	Requirements → architecture → improvements → draw
@0xDataWolf	"Bugs love unnamed modes"
@nickbaumann_	Tight updates + phone rule
@Graham_Tonkin	Progress / Problems / Plans updates

Curated collections & docs

Source	Contribution
Prompt Snacks	26 power phrases (poka-yoke, holdout, Tufte, half-the-code, tradeoffs, …): @nateberkopec, @poteto, @simonw, @thorstenball, @danluu, @juliknl, @petekp, @dimitribosch, @bhumi1102, @puntpoints, @netnostalgic, @_quinnwins, @Atin, @JoshDefilippi, @ade_oshineye, @decruz, @UmamiByTheSea, @markahesketh, @VesterDe, @SteveGBanton, @schwad_rb, @skatkov
Notchbird Banger Prompts	Codebase map (read-only), transcripts prune, AGENTS.md restructure, usage-habit audit
Anthropic prompt patterns	Outcome-only tasks, self-checking tasks, reference-file matching, measurable targets, artifact-giving, format-specifying
BestPromptFinder guide	Smallest-diff discipline, fix-cause-not-test, one-behavior-per-test, evidence rules, repeated-correction-to-rule
@mattpocockuk skills	4-sentence grilling, tracer-bullet tickets, plan-space beads, wait-what re-pitch, teach-me, questionnaire, small interfaces, arch surveys, two-axis review, to-spec, feedback-loop-red, TDD loop
@ykdojo tips	HANDOFF docs, fresh-context discipline
Boris Cherny (via team tips)	Plan-mode-first, grill-before-PR, scrap-for-elegant, CLAUDE.md self-update, subagents for compute

🤝 Contributing

One line per prompt. Keep every verb, noun, and adjective working.

Add the prompt under the right heading in SKILL.md. No credits in the skill. Context stays lean.
Add your credit row to the matching table in this README.
Open a PR. If it doesn't bite on a frontier model, it doesn't merge.

📄 License

MIT: see LICENSE. Steal these prompts; they're meant to be pasted everywhere.

⭐ Star this repo if a one-liner saved your codebase. ⭐