A curated collection of AI interaction patterns, system prompts, and tooling built from hands-on experience working with large language models in real workflows.
Skill Claude CodeCodex
Falsification engine that won't stop until your position is genuinely sound or you've explicitly acknowledged its flaws. Use when you have a thesis, plan, or belief you want to stress-test. Does not concede to eloquence, does not stop at an arbitrary round count, and requires you to demonstrate — not just assert …
Skill Claude CodeCodex
Surfaces implicit assumptions, evaluates them against available evidence, and refuses to let you proceed on unexamined foundations. Use when you suspect you're trapped in a frame you can't see. Doesn't just make assumptions visible — classifies them as supported, unsupported, or contradicted, and requires you to deal…
Skill Claude CodeCodex
Enforces calibrated reasoning by challenging unjustified priors, blocking asymmetric evidence processing, and refusing to let motivated reasoning pass uncorrected. Use when reasoning about an uncertain question. This is not a passive ledger — it intervenes when your reasoning is corrupted and won't proceed until the…
Skill Claude CodeCodex
Scientific debugging that goes beyond the immediate bug to the process failure that allowed it to exist. Generates ranked hypotheses with falsification experiments, confirms root cause, then demands the root-cause-of-the-root-cause — what systemic gap let this happen, and what prevents the entire class of bug, not…
Skill Claude CodeCodex
Finds deep structural patterns across ideas, but refuses to let synthesis remain abstract. Every synthesis must produce falsifiable predictions and connect to specific action — or be downgraded to aesthetic association. Use when you have concepts that feel connected and want to discover if the link is real, testable…
Skill Claude CodeCodex
Evaluates decisions through structured pressure, then refuses to let you proceed without addressing critical failures found. Maps the system, applies pressure from five angles, traces consequences — and holds you accountable for what the pressure test revealed. Does not end with 'here are the risks' — ends with 'what…
Skill Claude CodeCodex
Rigorous editorial feedback that holds writing to evidence standards regardless of venue. Will tell you a piece is not ready for publication when it isn't. Enforces claim-evidence alignment, flags unsupported assertions as failures not style choices, and refuses to lower standards because the intent is 'casual.' Never…
Skill Claude CodeCodex
Strips solutions to fundamental constraints and rebuilds from atoms — but demands that stated constraints are verified, not assumed. Blocks 'it depends' without data-gathering commitment. Challenges every constraint claim with 'prove it' before accepting it into the derivation. Garbage in, beautiful reasoning, garbage…
Skill Claude CodeCodex
Guided learning that verifies understanding rather than assuming it. Calibrates to your level, teaches through progressive depth, and requires you to PROVE comprehension — not just nod along. Uses retrieval practice, explanation demands, and application challenges because feeling smarter is not being smarter.
Skill Claude CodeCodex
Maps the failure landscape of a plan and issues a go/no-go verdict based on what it finds. Use before committing to a decision or shipping a project. Assumes failure, works backward to find causes — and if the failure landscape is severe enough, will tell you not to proceed. Includes follow-up protocol to check…
Skill Claude CodeCodex
Rebuilds a developer's complete, accurate ownership of work they directed but did not fully author. Especially effective on large or multi-file AI-generated changes where whole regions — and whole code paths — were never in the developer's head. Enumerates every unit of change into a coverage ledger, classifies by…
Skill Claude CodeCodex
Day 1 practice — ask so answers are checkable (FACT/GUESS, sources, UNKNOWN), then verify before you ship. Paste into any chat when a factual answer might fool you by tone alone.
Skill Claude CodeCodex
Day 2 practice — force inherited assumptions onto a labeled list (ASSUMED/UNKNOWN, GIVEN/FILLED-IN), then accept/replace/reject before you ship a plan. Paste into any chat when a brief might be incomplete.
Skill Claude CodeCodex
Day 3 practice — keep hard rules alive in long threads by re-sending a CONSTRAINTS block with every risky ask, forcing conflicts into the open, and checking the output against each rule (ideally with a test outside the chat). Paste into any chat when a session-wide rule must not fade.
Skill Claude CodeCodex
Day 4 practice — write a hard rule that refuses even you. When a later soft ask lifts it in words ("use requests for this demo; we'll remove it before merge"), the model should refuse and name the rule instead of complying. Test with three asks (weak rule first, then the hard-rule template) and a grep/CI check.
Skill Claude CodeCodex
Day 5 practice — stop the model from agreeing with you against the evidence. State your belief as a hypothesis, make it list contradicting data first with exact quotes, force a SUPPORTED / CONTRADICTED / UNKNOWN verdict, then test the verdict with a belief swap, one pushback, a quote grep, and a script where the claim…
Skill Claude CodeCodex
Day 6 practice — stop trusting the model's summary of a tool result. Make it show the exact command, exit code, and output before any interpretation, stop on error or empty, report only calls it really made, and end with PASSED / FAILED / STOPPED / NOT RUN. Then grep the quotes against the real output, gate the run…
Skill Claude CodeCodex
Day 7 practice — stop trusting the model's memory about anything that changes. Give it today's date and the lockfile, make every changeable claim carry "as of " and a source, label memory-only claims FROM MEMORY, stop when there is no live source, and add dependencies through the package manager. Then grep for undated…
Skill Claude CodeCodex
Day 8 practice — self-graded is not passed. When the same agent writes the code and the tests, it grades its own work: a wrong reading can't fail. Before any code, have the agent restate each requirement as concrete examples (input -> expected result), flag lines with two readings, and stop for your approval. Approved…
Skill Claude CodeCodex
Day 9 practice — worked around is not solved. A blocked agent works around the blocker (a default for a missing secret, a skipped test, a mock) because nothing said stopping was the job. Before it works, name the conditions where it must stop, the workaround shapes that count as tripping them, and a fixed BLOCKED…
Skill Claude CodeCodex
Day 10 practice — parsed is not valid. When code reads model output, lenient parsing turns a wrong reply into a wrong value with no error (a decimal comma shrinks a refund 1,000 times, a missing amount becomes zero). Define the exact shape, make missing a legal null, tie values to their source, and let a validator…
Skill Claude CodeCodex
Day 11 practice — flagged is not gated. In a multi-task plan, a premise flagged at the top ("settlement timestamps are UTC") holds no task back, and later tasks and sessions inherit the guess. Each task names the premises it needs; a premise is KNOWN only with a source you can open; the check runs when a task starts…
Skill Claude CodeCodex
Day 12 practice — copied is not corroborated. For a premise that moves money or data, more sources usually means more copies of one page, so agreement proves nothing. Trace each source to its origin (the page or system its words or value first came from), count origins not links, and confirm a critical premise only…
Skill Claude CodeCodex
Day 13 practice — approved is not reviewed. An agent's approval request usually shows what it did and which checks passed, so nothing on it could make you refuse and the yes is a reflex. Ask only before actions that can't be undone or that move money, data or messages outside the repo, and ask with a card - exact…
At most 3 mods per repository are shown here, and a mod shipped inside a plugin is left to that plugin's page — the rest are on their repository pages: