agentsclimarketplace
Sign in

Writing skills

Skill the-inconvenience-store/superbeads/skills/writing-skills

Use when creating, editing, pruning, or behaviorally testing an agent skill.From its SKILL.md

Install
npx -y skills add the-inconvenience-store/superbeads --skill writing-skills

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

5.3 KB, as published. Nobody here has run it

Writing Skills

Skill authoring is test-driven process design: make agent behavior fail, change the smallest instruction that addresses the failure, then prove the behavior converges.

Required background: use superbeads:test-driven-development for the RED-GREEN-REFACTOR discipline.

Trigger / Non-trigger

Use this skill for new skills, behavioral changes to existing skills, description/trigger changes, progressive disclosure, or pruning that could alter agent behavior.

Put one-project conventions in project instructions, enforce mechanical constraints in code or guards, and keep one-off solutions out of the skill catalogue.

Iron law: NO SKILL WITHOUT A FAILING TEST FIRST. For an edit, the failing test demonstrates the desired behavior is absent or the current structure violates a measurable contract.

Inputs

Establish:

  • target skill and actor;
  • invocation conditions and non-triggers;
  • observed baseline failure, pressure, or structural violation;
  • required behavior and checkable completion criterion;
  • host/tool constraints; and
  • affected tests, references, and shared-policy owners.

Steps

  1. RED — expose the failure. Add the smallest realistic scenario or deterministic contract that fails for the intended reason. Include a no-guidance control for wording experiments and use fresh context for each behavioral sample. Complete when the failure and its evidence are reproducible.
  2. Classify the failure. Distinguish invocation failure, omitted structure, wrong output shape, rule-breaking under pressure, missing reference, and stale/duplicated prose. Choose the matching intervention: trigger, required field, positive recipe, hard guardrail, context pointer, or deletion. Complete when one failure class and one target behavior are named.
  3. Design the information hierarchy. Read information-architecture.md now. Keep ordered actions and must-have gates in SKILL.md; move branch-only reference behind a precise conditional pointer; assign every meaning one owner. Complete when every retained section has a rung, branch, and owner.
  4. GREEN — write the minimum binding change. Model-invoked descriptions state invocation conditions and begin Use when or Use before; user-invoked descriptions are short, human-facing, and trigger-free. The body owns execution. Prefer a positive target and checkable shape. Keep prohibitions for demonstrated discipline failures or unavoidable safety boundaries, paired with the desired action. Complete when the RED test passes without unrelated guidance.
  5. REFACTOR — close observed variance. Run multiple fresh samples for behavioral wording, inspect every result, and tighten the form before adding prose. Delete duplication, sediment, and no-ops; preserve quality-bearing caveats by co-locating or disclosing them. Complete when results converge and existing contracts remain green.
  6. Verify deployment. Run the skill's named tests, frontmatter/description lint, repository guards, and relevant host fixtures. When a change affects a common invoked path, snapshot the candidate and run python3 scripts/workflow-metrics.py compare --baseline tests/fixtures/workflow-metrics/baseline.json --candidate <snapshot>; growth needs behavioral evidence or pruning/progressive disclosure. For design, discipline, or judgment skills, weave the Production-Grade Doctrine and security floor into domain language. Complete when current evidence covers triggers, execution, completion, and non-regression.

Completion

A skill change is complete only when:

  • RED evidence predates the behavior change;
  • GREEN and regression tests pass in fresh context;
  • a model-invoked description contains triggers rather than workflow instructions, or a user-invoked description is a trigger-free label;
  • every step has a checkable completion criterion;
  • conditional reference is reachable through a condition-bearing pointer;
  • each meaning has one owner and moved content has no orphaned requirement;
  • host/tool syntax is platform-correct; and
  • verification evidence is current.

Routing

  • Skill behavior failure → create or extend a deterministic contract/microtest first.
  • Pressure/rationalization testing → read testing-skills-with-subagents.md.
  • Platform packaging or schema uncertainty → read anthropic-best-practices.md and the target host's official specification.
  • Unexpected test behavior → superbeads:systematic-debugging.
  • Completion claim → superbeads:verification-before-completion.

Conditional References

  • Read information-architecture.md before restructuring, splitting, pruning, or changing descriptions.
  • Read testing-skills-with-subagents.md only when designing behavioral or pressure fixtures.
  • Read persuasion-principles.md only when an observed discipline failure needs a rationalization counter.
  • Read graphviz-conventions.dot only when a non-obvious decision or loop materially benefits from a diagram.
  • Read anthropic-best-practices.md only for platform-specific authoring or packaging questions.

When this workflow closes after producing durable insights, read Durable Memory and apply it; otherwise do not load it.

Keep looking

Skills are one crate of 627,801. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.