install-prompt-authoring
Create install with a prompt: generate a ready-to-paste install prompt for the README plus AI-executor install docs (`docs/INSTALL-PROMPT.md`) in any project, any ecosystem (pi, npm, Homebrew, pip, cargo, ...), after a mandatory gate that checks whether the owner already has an `INSTALL.md`. Use when the user says "create install with a prompt", "bikin install dengan prompt", or asks for a preset prompt their users can paste into an AI to install the project, together with docs that AI executes (dependency check, ask-first, enable config, verify, refusal paths). Not for writing human-facing install docs or the project's own INSTALL.md, not for running the install itself, and not for general README work unrelated to an install prompt.
- License 许可
- MIT
- 最近更新
- 2026-10-15
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- install-prompt-authoring
- license
- MIT
- description
- Create install with a prompt: generate a ready-to-paste install prompt for the README plus AI-executor install docs (`docs/INSTALL-PROMPT.md`) in any project, any ecosystem (pi, npm, Homebrew, pip, cargo, ...), after a mandatory gate that checks whether the owner already has an `INSTALL.md`. Use when the user says "create install with a prompt", "bikin install dengan prompt", or asks for a preset prompt their users can paste into an AI to install the project, together with docs that AI executes (dependency check, ask-first, enable config, verify, refusal paths). Not for writing human-facing install docs or the project's own INSTALL.md, not for running the install itself, and not for general README work unrelated to an install prompt.
02 资源文件 10
- 参考文档eval/fixtures/project-a/package.json172 B
- 参考文档eval/fixtures/project-a/README.md194 B
- 参考文档eval/fixtures/project-b/INSTALL.md240 B
- 参考文档eval/fixtures/project-b/package.json182 B
- 参考文档eval/fixtures/project-b/README.md156 B
- 参考文档eval/results/queries-summary.md1.1 KB
- 参考文档eval/results/scenarios.md438 B
- 脚本eval/run-eval.mjs12.4 KB
- 参考文档eval/scenarios.md2.6 KB
- 参考文档eval/trigger-queries.json2.3 KB
03 SKILL.md 正文
Install prompt authoring
Create an "Install with a prompt" README section plus AI-executor install docs for a target project. The prompt is paste-ready and points an AI at the docs; the docs are written for the AI executor, not for a human running a shell script.
Two artifacts, always understood as a pair:
- README section
### Install with a prompt— a short blockquote prompt the user pastes into their AI. docs/INSTALL-PROMPT.md— the instructions that AI executor follows: dependency check, ask-first before installing, enable config, verify, refusal paths, security note.
Workflow
Follow in order. Each stage ends on its completion criterion.
1. Gather target context
Read the target project to learn its ecosystem before writing anything: package manifest (package.json, pyproject.toml, Cargo.toml, go.mod, ...), existing README, existing install docs, config files.
Derive from context, never from memory:
-
the detection command (
pi listvsnpm ls -gvsbrew listvspip list, ...) -
config paths and scopes the project actually uses
-
reload/restart facts (new session,
source, service restart, ...) -
what "installed and working" means for this project
-
Done when: you can name the detection command, the config location, and the verification step for this project.
2. Gate: check for the owner's INSTALL.md (mandatory — never skip)
Check whether the repo already has an INSTALL.md (repo root or docs/).
If it does NOT exist: generate docs/INSTALL-PROMPT.md from scratch (Stage 3). No question needed — there is nothing to reconcile with.
If it DOES exist: you must ask the user which of these three options to take. Do not pick for them, do not skim past it, do not generate first and ask later:
- Pakai yang ada (use the existing one) — no new file. The README prompt points at the owner's
INSTALL.md. No redundancy;docs/INSTALL-PROMPT.mdis not created. - Ada tambahan (there are additions) — generate
docs/INSTALL-PROMPT.mdcontaining only the additional instructions the user supplies, written in AI-executor format and explicitly referring to the owner'sINSTALL.mdas the base. - Dibedakan (differentiate them) — ask back: "mau dibedakan bagian apanya?" (which parts should differ?). If the answer is the cliché "INSTALL-PROMPT.md is for the AI, INSTALL.md is for regular users", accept it and build
docs/INSTALL-PROMPT.mdas a separate document for its own audience, referencingINSTALL.mdfor the rest. If the answer is something more specific, dig into what they actually mean before writing.
- Done when: existing
INSTALL.mdchecked, and (if found) the user answered — recorded as the chosen option.
3. Generate the artifacts
README section — append ### Install with a prompt:
- If the README exists but has no such section, append only this section — do not restructure the rest.
- If the README does not exist, create it containing only this section (no other sections invented).
- Shape (adapt to project tone):
### Install with a prompt
Prefer to let your AI do it? Paste this prompt into `<harness>`:
> Install `<project>` for me. First read [docs/INSTALL-PROMPT.md](docs/INSTALL-PROMPT.md)
> from the repository and follow it exactly. Check my environment with `<detection command>`,
> ask me before installing anything I do not have, complete the config described there,
> and verify at the end. Do not loosen any guardrail mentioned in those docs.
When the gate chose option 1, the prompt links the owner's INSTALL.md instead and no docs file is created.
docs/INSTALL-PROMPT.md — create docs/ if missing. Written for the AI executor, in imperative second person. Required parts:
- What this is / who reads it — the README points here; the executor reads this file first and runs it step by step.
- Dependency check — table or list of what is required, the detection command for this ecosystem, and the rule: read the output, do not guess.
- Ask-first before install — anything missing gets asked about, never silently installed.
- Config — paths, scopes, the exact value that activates the feature; merge, never clobber.
- Verify — concrete end-state checks, plus the reload/restart fact.
- Refusal paths — situations where the right move is to stop and report, not to "fix" the failure by loosening a guardrail or config. Never weaken guardrails to make an install succeed.
- Security note — what gets loaded with what permissions; point the user at the source before installing anything privileged.
Option 2 output instead contains only the user's additions, in the same AI-executor format, referring to the owner's INSTALL.md for everything else.
- Done when: both artifacts (or just the README, for option 1) exist and contain their required parts.
4. Verify (barrier before finishing)
Check every line before reporting done; a failed line means fix it first:
-
Gate actually ran; if
INSTALL.mdexisted, the chosen option and the user's answer are reflected in the output. -
README section exists and links the correct doc (
docs/INSTALL-PROMPT.mdor the owner'sINSTALL.mdper the gate choice). -
When the README was created new, it contains only the install-prompt section.
-
docs/INSTALL-PROMPT.mdincludes ask-first, config, verify, refusal paths, and the security note. -
Every command and path comes from the target project's actual ecosystem — no placeholder like "run the appropriate command".
-
Nothing was installed, enabled, or modified in the target project while authoring (unless the user separately asked for that).
-
Done when: every bar passes.
Length
Write concise, complete instructions — not for a fixed character budget. No minimal-word quotas, no padding to hit a length: judge by completeness of the required parts, not by size.
Language
Discuss and write in the language the user uses; the generated prompt and docs follow the target project's own README language.
04 同仓库更多 Skills
skill-audit-test
A skill for testing stuff. Helps with test files.
foo-skill
Use this skill whenever a Markdown table needs to be made clean and readable: align columns, pad cells, and normalize spacing and pipe indentation so every column lines up consistently. Use it even when the user never says "format": they paste a table with ragged, uneven, or broken-looking columns; ask to fix, tidy, prettify, rapiin, or align it; write in Indonesian ("tabel ini berantakan, rapiin dong", "rapikan tabel ini"); or report that their Markdown table renders badly. Not for generating table content, editing or sorting data, explaining Markdown syntax in general, or formatting tables in non-Markdown formats (CSV, HTML, AsciiDoc).
skill-authoring
Create, audit, fix, and evaluate Agent Skills — the agentskills.io format (a folder with SKILL.md: frontmatter name/description, instructions, optional references/scripts) — so they meet the official specification and actually work. Use when making an Agent Skill from an expertise, procedure, or workflow ("buat skill untuk X", "turn our release process into an Agent Skill", "jadikan ini skill"); auditing or fixing an existing skill against the official spec (frontmatter, naming, length, references, instruction quality); or testing whether a skill activates when it should and behaves as intended (trigger queries, behavior scenarios, with/without baseline). Also use when a skill is being written inside the knowledge-to-agent-instructions flow: own the skill artifact per this spec, do not re-decide component choice. Not for deciding which component knowledge becomes (AGENTS.md, supervisor rule, subagent) — that is knowledge-to-agent-instructions; not for generic documentation.
knowledge-to-agent-instructions
Turn a user's knowledge — best practices, standards, conventions, hard-won experience — into instructions AI agents actually follow: AGENTS.md, pi-tool-supervisor rules, Agent Skills, and pi-subagents. Use when a discussion or implementation reaches a conclusion worth capturing ("so the way to do X is Y — that should be the rule", "berarti best practice-nya begini, jadikan aturan"), when the user asks to formalize a standard, convention, or lesson into agent instructions, or when they want knowledge applied to future agent work. Also use when asked to adopt or enforce a specific standard or practice whose purpose is not stated — clarify the purpose before designing. Not for general documentation, code review, or anything outside those four components. When triggered by a newly reached conclusion, offer once to capture it as durable knowledge and wait for the answer — never execute immediately, and never re-offer the same conclusion after it was declined or already captured.