AGENT SKILL

hand-test

Hand the user ONE chat-native test after a completed build/swing — run-card's rigor delivered in the chat (nothing to open), 3–5 DO/SEE-checkbox steps they tick as they flit. Use when laying a test/verification on the user: "hand-test", "lay a test on me", "how do I check/verify this", "give me steps to confirm it works", or after building something they'll verify by hand. NOT run-card (a paste-ready FILE for touching foreign systems), NOT rolls (status reports), NOT a CI runner.

License 许可
GPL-3.0
所属仓库
inputloop/garden
最近更新
2026-10-14
Stars
★ 0
Forks
0

01 前言 Front Matter

name
hand-test
description
Hand the user ONE chat-native test after a completed build/swing — run-card's rigor delivered in the chat (nothing to open), 3–5 DO/SEE-checkbox steps they tick as they flit. Use when laying a test/verification on the user: "hand-test", "lay a test on me", "how do I check/verify this", "give me steps to confirm it works", or after building something they'll verify by hand. NOT run-card (a paste-ready FILE for touching foreign systems), NOT rolls (status reports), NOT a CI runner.

02 资源文件 13

03 SKILL.md 正文

hand-test — a chat-native test-program, handed to the keeper

A hand-test gives the user one small test to run after a completed unit of work, delivered in the chat so they never navigate away, never open a file, never parse code-vs-markdown in a pager. It carries run-card's spine (ordered steps · exact expected results · a clean/pass check) but is sized for a person who is jumping around and must not lose the chat.

When to use

  • After you finish a build the user will verify by hand — a feature, a fix, a cross-fabric change.
  • Ration it: one hand-test per completed unit ("swing"). Not after every edit; not a wall of checks.
  • Do not use it for running against a foreign/production system you didn't build — that is run-card (a paste-ready, auditable file). Do not use it for status reporting — that is rolls.

The form (keep it exactly this shape)

A titled card, 3–5 steps, closing in one line. No preamble, no "why" inline (offer it on ask — extra prose is noise that hides signal).

HAND-TEST — <the claim this proves> (<what unit/swing>)
Run at <the working directory>.

① <what this step proves>
   <ONE fenced, copy-paste command>
   - [ ] SEE `<the exact string/shape to look for>`  → <pass condition, ≤5 words>

② …
③ …

all ✓ ⇒ <the claim> holds.
  • DO = one fenced command. The user's only terminal act is paste + enter. No cd-then-open-then-read.
  • - [ ] SEE … is a checkbox, not prose. It carries the exact expected output next to the command so verification is a glance-compare, and the user ticks it inline as they go. It does not report back to you — it is their progress marker while they flit. (Mirror in a widget only if you also keep the plain-chat version.)
  • Close with one line: all ✓ ⇒ <claim> holds.

Non-negotiables

  1. Prove every DO yourself first. Run each command in your sandbox before handing it. A hand-test that fails on the user's machine is worse than none — it spends their focus and their trust.
  2. Prereqs are proven, not assumed. Name the real interpreter invocation and pin the working directory. Watch the user's actual toolchain — e.g. an interpreter not on PATH, a path with spaces that bites in bash, a relative require/import that needs the right CWD, a stricter compiler on their host than in your sandbox. If in doubt, lead with a one-time alias/PATH line.
  3. Chat-resident. Never "go read some.md". The steps and expected outputs live in the chat message so they inherit real rendering (headings, fences) — code and prose never blur.
  4. Self-locating. Because each step is self-contained (do / see / pass), flitting away and back costs nothing: the chat still says what to paste, what to see, and how to know.
  5. Un-pedantic. 3–5 steps, five-word pass conditions, no ceremony. Over-instruction is the failure mode.

Why this shape (the reading model it serves)

People who move fast between the chat and a terminal read differently: the chat is the reading surface and every file-open is a flit-away cost plus a parse burden. They copy-paste rather than navigate, they verify by expected-beside-actual, they filter hard so rationed + high-signal wins, and many are spatial thinkers for whom an aligned DO / SEE / ✓ layout reads faster than a paragraph. Design to that.

Interlocks

  • run-card — the ancestor; use it (not this) for careful, auditable runs against systems you didn't build.
  • rolls — status/close reports; a hand-test proves, a roll summarizes.
  • takushi — the last-mile delivery; a hand-test is one clean thing that must land where the user can see it (the chat) and be acted on without you in the room.

How this skill can fail (falsifiers)

  • An unproven DO. fails if a step's command was not run in the sandbox before handing; check: each fenced command appears in an earlier tool call of the session, with the output its SEE line quotes.
  • SEE does not match. fails if the keeper's output differs from the SEE string; check: put their output beside the SEE line. The first suspects are the toolchain prereqs (interpreter, PATH, working directory).
  • Too big. fails if the card has more than 5 steps or asks the keeper to open a file; check: count the steps and look for "open" or "read <file>".

<!-- footer: minted by bin/ft1mt from conventions/house.env and conventions/FOOTERS.org; edit those, not this --> Version 1.1.0: edition revised from 1.0.1 on 2026-10-10; spine: 2026-09-11.

Copyright 2026 Corwin Brust corwin@bru.st and Contributors; CC-BY-SA4: https://creativecommons.org/licenses/by-sa/4.0/ (scripts GPL-3.0-or-later); Mentat (for Corwin; claude-opus-5-5) <!-- /footer -->

04 同仓库更多 Skills