All authors
Claude Skills by Kohulan
github.com/Kohulan22 skills0 installs22 views
- Bounded GoalsWrites the success criteria of an optimization or evaluation loop as one maximized metric, relaxable bounds on every other metric, and a written trade licence. Use when writing or revising goals.json, acceptance thresholds, benchmark targets, or agent objectives, especially with more than one metric, and when an iteration loop has stalled because every proposed change is rejected for regressing something. Not for running the iterations themselves (use eval-loop).Votes: 0GitHub stars: 2
- Change Asserted ValueRequires three artifacts before a committed expected value moves (a primary-source quote, an independent check that does not use the code under test, and a mutation test) and records anything weaker as unverified. Use when a change would alter a golden file, snapshot, asserted label, reference output, or benchmark expectation, including when the new value looks obviously right or the old one looks obviously wrong. Triggers on updating a failing assertion to match current output, 'the test exp...Votes: 0GitHub stars: 2
- Check TargetTests whether a failure pattern named as "the thing to fix" is one real defect class, by recomputing its signature over all rows (passing ones included), reporting hits, failures, precision and counterexamples, and splitting the hits by cause. Use when a cluster, tell, signature, or failure pattern has been named as the fix target — before proposing, scoping, or starting any fix. Triggers on "the biggest failure class is X", on a fix target inherited from a handoff note, roadmap, or review, a...Votes: 0GitHub stars: 2
- Cluster FailuresRanks the failures of an eval or benchmark run by a structural feature of the input (ring system, charge state, stereocenter count, size, functional group) to set the build order. Use when an eval or benchmark run has finished, when the question is "what should I work on", "where are we losing", or "what is the biggest failure class", and before scoping any accuracy work. Triggers on failures bucketed by what the output looked like rather than by a structural feature of the input molecule. An...Votes: 0GitHub stars: 2
- CouncilAnswers a strategy or judgment-call question by re-deriving its load-bearing numbers first, then weighing the options through three or more distinct expert voices, then making one call with the condition that would flip it. Use when the user asks a strategy, prioritization, or judgment-call question — which option, what to work on next, is X worth it, should we do A or B, your call, show me the options, what's the best move, is this the right approach — or any decision that rests on a number,...Votes: 0GitHub stars: 2
- Enumerate FirstReplaces a reachability or likelihood argument with a seeded, generated sample of the input space and a failure count. Use when about to argue whether a case is reachable, likely, rare, or worth fixing — triggers on 'probably unreachable', 'edge case', 'in practice this can't happen', 'unlikely to matter', or deciding whether a reported defect is real. Also use before optimizing or tuning any decision point, to confirm it executes at all. Not for cases where a reproduction already exists.Votes: 0GitHub stars: 2
- Eval LoopRuns a bounded improvement loop on a chemistry model or pipeline (measure, cluster failures, fix one cluster at the root, re-measure, gate, log), maximising one metric while every other stays a bound. Use when the user says "improve accuracy", "work the loop", "run the eval loop", "what should we fix next", or wants measurable progress across iterations. Also use when an accuracy push has no stated objective and no bounds, or when an iteration moved the headline number by zero and the next st...Votes: 0GitHub stars: 2
- Fable ReviewGets one adversarial review of a load-bearing plan, diagnosis, or finding from a reviewer on a different model family (fable by default), told to refute its premise. Use when a plan, diagnosis, or finding is about to ship and its premise, if wrong, would waste the next iteration. Triggers on "fable-review", "review the plan with fable", "get a second opinion", a diagnosis about to size or scope a build, and any claim repeated across sessions without being re-derived. Not for ordinary line-by-...Votes: 0GitHub stars: 2
- Gauntlet LoopWrites one short, paste-ready prompt that makes a fresh agent loop a builder against a separate harsh critic until the output beats a named, fetchable reference, and can run it on request. Use when the user wants an agent to grind on a goal until it beats a named reference — triggers on "/gauntlet-loop", "gauntlet loop", "gauntlet this", "make a gauntlet prompt", "loop until it beats X", or "make it as good as [named thing]". Applies to builds, writing, code, research, and design goals alike....Votes: 0GitHub stars: 2
- HandoffEnds a chemistry-dev work session by verifying that memory/notes are up to date and writing a handoff note, so the next session needs no hand-written prompt. Use when the user wants to wrap up, clear, or "start a new session", or asks "is everything stored in memory?" / "give me a prompt for the next session". Not for resuming at the start of a session (use kickoff).Votes: 0GitHub stars: 2
- KickoffStarts or resumes a chemistry-dev work session by self-priming from the project's handoff note, durable notes and git, with a drift check against the roadmap before the next task begins, in place of a hand-written resume prompt. Use at session start when no task has been given yet. Triggers on "continue", "resume", "kickoff", "where were we", or "pick up from the handoff note". Not for wrapping up a session (use handoff).Votes: 0GitHub stars: 2
- Promo VideoMakes promotional, launch and explainer videos as code (motion graphics, kinetic type, real product screen footage, 3D with three.js, fluid ink and a composed or supplied soundtrack), rendered frame-exact to MP4 for LinkedIn, X, YouTube, Instagram, talks or a website. Starts with a short interview, storyboards stills before anything moves, and ships a review page for comments. Use when the user wants any promo video, teaser, showreel, launch or announcement video, product demo video, animated...Votes: 0GitHub stars: 2
- Prove InvariantDerives a property the output must satisfy by necessity (symmetry, invariance, counting, an algebraic identity), gates the output on it, and mutation-tests the gate. Use when adding or reviewing tests for correctness-critical output, when a suite passes but the values may still be wrong, or when stored expectations (golden files, snapshots, asserted labels) are the only oracle. Triggers on 'all tests pass but', 'is this actually correct', regression suites for scientific or algorithmic output...Votes: 0GitHub stars: 2
- Refusal CensusAttributes each abstention or failure in a run to the internal code site that refused it, from per-input refusal logs, and ranks those sites by sole-blocker count (the inputs that only one site blocks) to set the build order. Use when the question is "why did we abstain/fail on these", "which site is blocking the most inputs", or "what should we build to raise coverage" — and before scoping any coverage work. Triggers on a run whose failures all carry one undifferentiated reason code, and on ...Votes: 0GitHub stars: 2
- Reuse Before RerunSearches the results ledger, the repo, scratch directories and notes in one sweep for an existing result, then gives a three-line REUSE / EXTEND / RERUN verdict before anything expensive runs. Use when about to launch any measurement over a minute long — benchmark, failure census, spy over many rows, head-to-head against another engine, determinism eval, shard generator, full gate, large-corpus run — and when the user asks for a number, a table, or the run from a given date. Also use on the t...Votes: 0GitHub stars: 2
- Run EvalRuns the project's committed eval harness on a named split and reads a headline correctness rate together with emit/coverage and precision-on-emitted, citing the run id. Use when the question is "did that change help", "what is our accuracy now", or "run the eval/benchmark" — and before and after any fix to a chemistry model or pipeline. Triggers on quoting an accuracy number with no run id, on a hand-rolled measurement loop in a scratchpad, on a single headline rate quoted alone, and on any ...Votes: 0GitHub stars: 2
- Run GateRuns a project's regression gate (a protected regression set plus a determinism check, or whatever the gate checks) by its full recipe — fast pre-gate while iterating, background launch, waiting on a sentinel verdict file, and reconciling the verdict's fields against a current baseline. Use when a change needs gating, before shipping a phase, or when the user says "run the gate". Not for measuring accuracy on an eval split (use run-eval) or for keeping watch over some other long job (use watc...Votes: 0GitHub stars: 2
- Spy SiteProves that a code site named as the place to fix actually runs for the target inputs, by counting its calls with a spy validated on at least two known positives, before any edit lands there. Use when a fix names a function, method, or module as "the place to change", and when a plan, roadmap, audit doc, or review points at a central decision point. Triggers on "the fix goes in X", "X is where this is decided", and any code site inherited from a document rather than from a measurement. Not fo...Votes: 0GitHub stars: 2
- Test GateChecks that a test fails on some plausible wrong code, such as a guard that always allows or always refuses, an empty result, an identity transform or a missing lock, and gates each new test on four answers: behaviour, regression, owner and seam. Use when writing, changing or reviewing tests: a regression test for a fix, a PR or diff that adds tests, a coverage push, or a green suite that may not catch a bug. Triggers on 'add a test for', 'review these tests', 'is this test enough', fail-clos...Votes: 0GitHub stars: 2
- Understand Before MergeDelivers code with the why of each non-obvious decision, a failure-mode table of what the code does when reality misbehaves, and open questions only the human can answer, so the reviewer does real review before merging. Use when writing, refactoring, or reviewing code the user intends to ship, merge, or run against real data — bug fixes, API endpoints, data pipelines, migrations, background jobs, and any script touching a network, database, filesystem, or user input. Triggers even on a bare "...Votes: 0GitHub stars: 2
- Verify SourceChecks what a specification, standard, RFC, paper, or datasheet actually requires against the primary source, and states the rule with a verbatim quote and its section number. Use when about to state what such a source requires — including when a project doc, research note, prior session, or another agent already asserts the rule. Triggers on 'the spec says', 'per IUPAC/RFC/ISO', 'the standard requires', citing a section number, or resolving a disagreement between an implementation and a clai...Votes: 0GitHub stars: 2
- Watching Background JobsKeeps a long-running background job watched from the main agent — a baseline and checkpoint, scheduled wake-ups that tell a stall from a quiet compute phase, a one-line progress report per wake-up, and a verdict-first alert on completion. Use when a long-running job (benchmark, regression gate, eval, test suite, build, batch run) is or will be running in the background and someone must notice completion, stalls, or crashes — triggers include "keep an eye on it", "ping me when it's done", "kee...Votes: 0GitHub stars: 2