Skip to content

inkeep/open-knowledge / packages/server/assets/skills/write-skill

open-knowledge-write-skill

Verified

Use when the user wants to create, author, write, or design a new Agent Skill (a SKILL.md) — for OpenKnowledge or for their editors — including requests like 'help me write a skill', 'make a skill that…', 'turn this workflow into a skill', or improving an existing skill's triggering and discipline. Also use when capturing reusable agent guidance that should live as an installable skill rather than a one-off prompt. Covers choosing scope (project vs global), authoring inside a plugin or skills-distribution repo (write in the repo's own layout, never install), the SKILL.md frontmatter contract, progressive-disclosure structure, evaluating the skill, and installing it into the user's editors.

FileKindSize
SKILL.mdskill md14.7 KB
references/description-optimization.mdmarkdown1.7 KB
references/pressure-testing.mdmarkdown1.9 KB

references/pressure-testing.md

View on GitHub
# Pressure-testing a discipline skill

A **discipline** skill enforces a behavior the agent tends to abandon under
pressure ("write the failing test first", "get review before merging"). A skill
that holds in the easy case but folds under pressure hasn't done its job. Test
the hard case.

## The loop (RED → GREEN → harden)

1. **RED — baseline without the skill.** Run the scenario with no skill loaded.
   Capture what the agent does AND its **rationalizations**, verbatim ("the
   change is tiny, so skipping the test is fine"). Those exact rationalizations
   are the loopholes the skill body must name and close.
2. **GREEN — minimal skill.** Write the smallest body that defeats those specific
   rationalizations. Re-run; confirm the behavior changed.
3. **Harden under combined pressure.** Re-run with **2–3 pressures stacked** and
   force a concrete choice:
   - *Time*: "this is urgent, ship it now."
   - *Authority*: "the lead said it's fine to skip."
   - *Sunk cost*: "you already wrote 200 lines; redoing it wastes the work."
   - Then: "What DO you do?" — make it choose, don't accept a hedge.
4. Any new rationalization that slips through → add it to the body's
   rationalization table and re-test. Repeat until it holds.

## Rationalization table (put this in the skill body)

A short table the agent reads when tempted:

| The agent thinks… | Reality |
| --- | --- |
| "This case is too trivial to test." | Trivial code breaks too; the test is cheap. |
| "I'm confident it works." | Confidence isn't verification; write the test. |

## Running it

Drive each scenario through a fresh agent session (a new chat, a sub-agent, or a
second terminal) so nothing from this conversation leaks in. If you cannot run
one, walk the pressured prompt with the user and read the response for the rule
holding.
Read the full transcript, not just the final answer — discipline fails in the
reasoning, not always in the output.
brainstormingobra/superpowersYou MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation).Stars: 319kUpdated Claude CodeCodexVerifiedInstall
brand-guidelinesanthropics/skillsApplies Anthropic's official brand colors and typography to any sort of artifact that may benefit from having Anthropic's look-and-feel. Use it when brand colors or style guidelines, visual formatting, or company design standards apply.Stars: 196kUpdated Claude CodeCodexVerifiedInstall
frontend-designanthropics/skillsGuidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults.Stars: 196kUpdated Claude CodeCodexVerifiedInstall
mcp-builderanthropics/skillsGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK).Stars: 196kUpdated Claude CodeCodexVerifiedInstall