SSkills line

Skills

A few skills I kept wishing I had while building with Claude Code and Codex. Each is one plain SKILL.md with instructions and no code; the agent does the work with the tools it already has.

Change here from the purple line. Works in Claude Code, Codex, Gemini CLI and claude.ai.

S1

silent-decisions

Makes the agent ask before choices that come back to bite you.

Agents make many small choices. A few matter later: an in-memory store that falls over on real data, a migration that drops a column, an endpoint that went public without a login. The agent stops on scale, data loss, security, cost and anything hard to undo, gives two or three options with its pick, and writes what you chose to decisions.md so the next session knows.

/silent-decisionswrites decisions.md
S2

fix-without-overfitting

Fixes the cause of a failure, not just the failing cases.

No hard-coded expected values, no if for one input, no bumped timeouts, no edited tests. Before the fix it sets up a check on cases the fix hasn't seen. After, it searches its own diff for values copied from the failures and shows before and after on the failures, the unseen cases and what already worked.

testsLLM pipelines
S3

abstract-dont-copy

Fixes what let a bug happen, everywhere it happened.

A report names one hole: Bob can read Alice's order. The same cause usually left others nearby. The agent names the cause in one sentence, finds every other place it shows up, sorts each as a hole, safe or intended, fixes at the smallest level that closes them all, and adds a test shown failing before the fix.

pairs with S2
S4

priced-in

Checks whether a Polymarket market moved before the news did.

Pins down when the news went public, looks for earlier leaks, pulls prices and trades around that time and explains each move. You get a verdict of priced in, partly priced in, or a surprise, with a source for every time it quotes. It can also score your own forecasts against the market.

Polymarketnews timing
S5

reply-review

Reviews a plan by quoting the lines you disagree with.

Long plans get skimmed. The agent writes plans as short numbered blocks; you reply by quoting lines and writing under them. Every answer is changed, explained, disagree or still open, and nothing changes until you say so.

plansreviews

Install

Claude Code

/plugin marketplace add sk1122/skills /plugin install silent-decisions@sk1122

Codex

codex plugin marketplace add sk1122/skills codex plugin add silent-decisions@sk1122

Gemini CLI

gemini extensions install https://github.com/sk1122/skills

Claude desktop and claude.ai

Download the zips from the latest release and upload each under Settings → Capabilities → Skills.

github.com/sk1122/skills Apache-2.0 ← satyam.fyi