Imported from ArttuAn/agent-harness-skills (
AGENTS.md). Install upstream withnpx skills add ArttuAn/agent-harness-skills. Copyright stays with the author.
AGENTS.md
Instructions for an agent contributing to this repository. Read this before editing anything. It is the whole contract — there is no implicit convention you are expected to infer.
What this repo is
Seven skills that teach an agentic IDE to build an agent harness (ReAct,
plan-and-execute, reflexion, multi-agent, code agent, RAG/memory,
self-evolving). Each skill is one skills/<name>/SKILL.md plus one matching
commands/<name>.md. Shared prose lives in references/.
The bar, before you write anything
If a frontier model already gets this right from a bare spec, the pattern does not need a skill.
This repo exists for the things a good model gets wrong unprompted: the evaluator that rubber-stamps its own output, the lesson that poisons every later run, the tool result that breaks the next request rather than the one that caused it. Content that restates what the model already does is rejected, however well written.
A contribution is worth making if you can name a specific failure you have seen, and the guard that prevents it.
Validate
One command. No network, no API key, under a second:
python3 tools/check_skills.pyIt must print OK before you open a pull request. It enforces:
| Rule | Detail |
|---|---|
| Location | skills/<name>/SKILL.md |
| Frontmatter | name: harness-<name> (must match the directory), plus description |
| Description length | 220 characters maximum |
| Required sections | ## Use this when, ## The decisions that matter, ## Failure modes, ## Required tests, ## Verify |
| Python blocks | every ```python block must parse; <placeholders> are fine |
| Matching command | commands/<name>.md must exist |
Add a skill
- Copy
docs/TEMPLATE-SKILL.mdtoskills/<name>/SKILL.md. - Copy
docs/TEMPLATE-COMMAND.mdtocommands/<name>.md. - Fill them in. Cite
references/rather than restating it — the LLM-seam and verification rules are shared and must not be duplicated. - Run
python3 tools/check_skills.py. - Open a PR. Say in one line which failure the skill prevents.
Add a failure mode
The most valuable contribution, and the smallest. Append to
references/failure-modes.md: what you saw, which endpoint or model, and the
guard that fixed it. It improves all seven skills at once.
Rules
- One skill per pull request. Additive changes only; do not restructure existing skills in the same PR.
- Do not edit
tools/check_skills.pyto make your contribution pass. If the validator is genuinely wrong, say so in an issue first. - Do not add dependencies. The validator runs on a bare Python 3 with no packages installed, and it stays that way.
- Write from experience, not from the model's priors. Concrete numbers, real error strings, named endpoints. If the text would be equally true of any framework, it is too vague to ship.
- Python blocks are illustrative, not runnable programs; they must parse.
Do not
- Touch
LICENSE, or the CI workflow, without saying why in the PR. - Add a skill for a pattern already covered — extend the existing one.
- Open a PR that only fixes typography, reflows prose, or "improves clarity" without changing meaning.