All authors
Claude Skills by whilehq
github.com/whilehq12 skills0 installs17 views
- CharacterTrain how a model talks from a constitution (a model spec) and prove the trait moved without costing anything else. Use when asked to give an agent a character, persona or voice that survives without a system prompt, to turn a model spec or list of principles into training data, to build DPO pairs or SFT rows for style, or to measure a trait before and after training. The judge is checked against the spec's own labeled replies before any pass rate is read; every trait is a Behavior on the pla...Votes: 0GitHub stars: 18
- Dpo PairsTurn graded rollouts of an agent into chosen/rejected preference pairs for DPO, check that the pairs teach the behavior and not a tic, write the JSONL a DPO trainer loads, and prove the result on a frozen held-out test reported to the While platform. Use when a coding agent is asked to build DPO or preference data from an agent's own rollouts, to fix a tool-use or policy slip with preference training, or to show that a trained version beats the served one.Votes: 0GitHub stars: 18
- Grpo VerifierTrain an agent with GRPO when the reward is a program, and prove the gain. Use when a coding agent is asked to do RL on a rule code can check (a tool called before another, an argument in range, a phrase in the reply), to write or test a verifier reward, to export a verifiers environment for a GRPO trainer, or to report a trained version against a frozen held-out test with its training curve. Not for rewards that need a model judge.Votes: 0GitHub stars: 18
- Harness SearchHow a coding agent runs the Meta-Harness loop as the proposer on an agent's own production traffic: point the recipe at yesterday's traces, read out/proposal.md, write the next candidates/<n>_<name>.py, run the recipe, read the ledger, stop when the gate passes or after N rounds, report Changed / Moved / Why / Learned / Reproduce, post every candidate as a harness version, and score the next day's traffic on the pick. Use when the thing to improve is the harness (the prompt, the tools, the tu...Votes: 0GitHub stars: 18
- Manage ExperimentsHow a coding agent tells the person what it did, why the number moved and how to run it again, so the platform page reads in one look. Use before posting a second version, a sweep, a replicate or a training run to while.ai/platform: the question first, then for every run five lines (Changed, Moved, Why, Learned, Reproduce), one picture, the rows that failed, points not fractions, and readback(tracked) to read the account the way a teammate will. No GPU, no key until the report.Votes: 0GitHub stars: 18
- Sft From TracesTurn an agent's own production traces into supervised fine-tuning rows and prove the result on a frozen held-out test. Use when a coding agent has a JSONL of traces (prompt, steps, final_text, a 0/1 label) from a deployed agent, wants to fix the failures they show with SFT rather than RL, and needs the before/after delta, the noise floor and the contamination count reported to the While platform so a person can decide whether to promote.Votes: 0GitHub stars: 18
- Strengthen Your EvalsTurn the tests an agent already has into an eval that can fail, measure it the way a paper would, and report it so the platform can tell a fix from a fluctuation. Use when a coding agent is asked to build or improve evals for an agent that runs on a frontier model (Claude, GPT, Gemini) or on its own weights: find what the suite never reaches, build a frozen held-out set, check the judge against people, read pass@1 with an interval and the failure-capable count, measure the noise floor, size t...Votes: 0GitHub stars: 18
- Tool Call EfficiencyMeasure and train down the number of tool calls an agent spends per solved task, with no judge. Use when a coding agent is asked to make an agent cheaper or faster without making it worse: redundant re-fetches, tools called that the task did not need, lookups repeated across turns. The score is a program (checks pass AND at most `budget` calls), so it is a verifiable reward for GRPO, and correctness is tracked beside it so cutting calls never cuts solves.Votes: 0GitHub stars: 18
- WatchKeep the platform's Live tile honest with one command a day, without the platform ever ingesting a trace. Use when an agent is served and someone asks "is it still fine in production", when the Live tile is empty or stale, when a served version needs a flagged rate that means the same thing as its held-out score, or when deciding whether a rise in failures is real enough to open a new run. Covers reading yesterday's traces, a date-seeded sample scored by the held-out judge, posting one LiveDa...Votes: 0GitHub stars: 18
- Whileai SimulationsUse the While simulations SDK to generate agent trajectories from a policy, tools, an agent harness, seed tasks, or production traces; focus generation on failures, grade with the developer's own authority, inspect quality, and export usable datasets. Use when a coding agent is asked to simulate an agent, expand coverage, turn traces into new scenarios, create eval or post-training data, or run the simulate-grade-select loop.Votes: 0GitHub stars: 18
- Audit Your JudgeCheck whether a judge can be trusted before its scores reach Select or Train: label a sample blind, grade it with every candidate judge, and read agreement, kappa and the false-pass rate against floors. Use before trusting any judge, rubric or grader score, before choosing a judge model, before calling compare_judges, judge_trust or attach_labels, and before Select curates rows by a judge's score. Triggers on judge, rubric, LLM-as-judge, grader, judge_trust, compare_judges, kappa, false-pass ...Votes: 0GitHub stars: 18
- Pick A MethodPick the training method from the graded rows before spending a GPU, out of everything the SDK ships: the hosted sft/grpo/dpo/rm, and on your own GPUs OPSD, OPD, GroupwiseGrading (GRS/GAR), Async, FlashReinforce, SAO, BPCO, ReinforceAda and prime-rl's grpo/max_rl/rae. Use when a coding agent has scored a model and must choose a method or its knobs, when a base scores zero everywhere, when every task already passes, when rows come one per prompt, when rows came from another model, or when some...Votes: 0GitHub stars: 18