skills

Stellarlink Skills

Agent skills for work that needs more control than a single prompt: long-running execution, high-recall code review, and measurable self-improvement loops.

These are not vibe-coding macros. They are small operating protocols for agents: explicit state, hard gates, deterministic rollback, and proofs before claims. Use them directly, fork them, or steal the patterns.

Quickstart

Install the collection with the skills.sh installer:

npx skills@latest add stellarlinkco/skills

Then call the skill you need from your agent:

/harness implement prd.md; loop verify, fix, retest

For manual installs, copy the skill directory you need from skills/ into your agent's skills directory. harness also needs hook registration; see skills/harness/README.md.

Why These Skills Exist

AI agents fail in boring, repeatable ways. They stop too early. They review too narrowly. They make a prompt or skill better once, then lose the path that made it better. This repo turns those failure modes into protocols.

#1: The Agent Stops Before the Work Is Done

The problem: long tasks die at session boundaries. Context windows reset, partial state disappears, and the agent starts summarizing instead of finishing.

The fix: harness gives the agent a durable task ledger, append-only progress log, hook-driven stop blocking, dependency checks, leases, and recovery rules. Progress files become the context.

Use it when a task has many subtasks, must survive sleep/resume cycles, or needs automatic recovery after a failed attempt.

#2: The Review Misses the Bug That Matters

The problem: most agent reviews optimize for precision too early. They produce a tidy list, but miss the dangerous bug hiding in a changed contract, deleted branch, or wrapper boundary.

The fix: code-review uses a max-recall pipeline: gather the real diff, generate candidates from independent angles, verify them, run a final gap sweep, then return a capped JSON findings list.

Use it for PRs, branch diffs, local working-tree diffs, security-sensitive changes, or any review where a missed P1 costs more than an extra candidate.

#3: Improvement Without an Oracle Turns Into Drift

The problem: “make this better” is not a loop. Without a measurable oracle, each mutation is just taste with confidence.

The fix: self-evolution turns prompts, skills, documents, configs, code, and experiments into evaluate-gate loops. It supports GT case suites and scalar scoreboard metrics, keeps a ledger, and reverts mutations that do not pass.

Use it when the artifact can be measured and you want repeated improvement without guessing.

Reference

Execution

• harness — Long-running agent framework for multi-session task execution, progress persistence, stop blocking, dependency handling, rollback, and recovery.
• harness install guide — Hook setup, state files, activation marker, concurrent mode, and failure-recovery behavior.

Review

• code-review — High-recall code review protocol for PRs, branch diffs, and working-tree diffs. Focuses on correctness, security, contracts, concurrency, performance, reuse, simplification, altitude, and convention failures introduced by the change.

Evolution

• self-evolution — Autonomous mutation-evaluate-gate loop for measurable artifacts. Supports GT Suite Mode, Scoreboard Mode, Hybrid Mode, layered mutations, trace-driven diagnosis, and deterministic keep/discard decisions.
• self-evolution references — Ground-truth format, artifact guide, evaluation layers, mutation strategy, and gate rules.
• self-evolution scripts — Assertion evaluation, structural checks, and results tracking helpers.

Design Notes

• One skill directory is one product surface.
• SKILL.md is the agent-facing protocol.
• Extra docs belong beside the skill only when setup or operation needs more detail.
• Scripts are support tools, not hidden behavior. The skill must explain when and why to run them.
• Verification is part of the product. A skill that cannot tell the agent how to prove success is unfinished.