← ClaudeAtlas

full-powerlisted

Load when the result must be the best the agent can produce and a mistake is expensive: a client deliverable, a production change, a decision others will act on, a hard bug, or when the user says "full power", "best possible", "no errors", "be perfect", "max effort", "potencia máxima", "lo mejor de lo mejor", "sin errores", "que sea perfecto". It does not make the model smarter; it spends the model's capacity where it changes the outcome: it sorts the task into steps by risk, sends only the hard step to the strongest model at high effort, runs independent attempts plus a judge when one answer is not enough, has a fresh-context verifier attack the result before "done", checks what the model cannot know or compute and uses tools instead of guessing, and writes every failure into a log the next run reads. Routine steps stay cheap. Not for quick questions or one-line edits.
RabbitDesigns22/claude-skills · ★ 0 · AI & Automation · score 75
Install: claude install-skill RabbitDesigns22/claude-skills
# Full power Most bad results from a strong model are not a lack of intelligence. They come from the same few places: the hard step got the same quick pass as the easy ones, one first draft was shipped as the answer, the agent graded its own work, it guessed a fact it could have looked up, and the same mistake happened again next week. Each has a fix that costs tokens only where it matters. One rule: **spend effort in proportion to what a mistake costs, and never let the author be the only judge.** ## 1. Sort the task by risk (1 minute, before any work) Write the steps, then mark each one: | Mark | Meaning | Examples | |---|---|---| | R (routine) | wrong is cheap and obvious | reading files, listing, formatting, renaming | | H (hard) | needs judgment; wrong is plausible-looking | the plan, the root cause, the design, the copy a customer reads | | X (irreversible) | acts on the world | sending, paying, deploying, publishing, deleting | Write the acceptance list for the whole task: 3-7 checks a stranger could verify ("the price matches the price list", "tests pass", "the page shows X live"). ## 2. Right model and effort per step - **R steps:** the current model at low/medium effort, or a small-model subagent for bulk reading that returns only what matters. - **H steps:** the strongest available model at high effort, in a subagent with a short brief that carries only what that step needs (goal, constraints, the acceptance checks, what was already ruled out). In Claude Co