--- name: agent-prompt-kit description: "From finished work to a specialist environment. On /agent-prompt-kit the skill reads the current session and the runtime's related sessions, writes a Task Card and a Domain Card (modules M1, M2), creates a new workspace or saved-agent directory with AGENTS.md rules, a specialization skill, worked examples, an acceptance criterion and a store of prior work, then names the registration step and hands over a starting prompt. Load when a class of work repeats and the next round should start from a prepared environment: 'make a specialist for this topic', 'create a workspace for these tasks', 'we did this before — prepare the working environment', 'analyse the session and build the environment'." license: MIT version: 1.2.2 upstream: https://github.com/paulmann/agent-prompt-kit --- # `/agent-prompt-kit` — from a session to a specialist environment One command, one result: **a directory in which similar work is done well from the first step** — because the rules, the specialization, the worked examples, the acceptance criterion and the prior experience are already inside it. This is **not** "give the current session a different personality". Changing the role inside the current window is a side mode (see §7). The main action is to create a working environment, register it, and hand the task over to it. ## 1. Material: this session and the related ones Sources in order of trust: | # | Source | What it gives | |---|---|---| | 1 | The current session's own context | what was asked, done, verified, and where it broke | | 2 | The runtime's index over all sessions, when it has one | the same across past sessions and other environments | | 3 | Session transcripts and files | detail the index truncated | | 4 | The project tree and `memory/` | rules that apply now, and existing experience records | Rules: - Quote a session, not an impression: "in session `` we tried X and got Y". - **Filter index noise.** An index contains injected context, tool transcripts and environment dumps. "We discussed this" is concluded only from the user's own messages. - A past session is evidence of what happened, not of what is true now. Re-verify before reuse. ## 2. Cards: what this class of work is ```text TASK CARD Subject · Deliverable · Audience · Success looks like · Constraints · Unknown · Out of scope DOMAIN CARD Practitioner · Vocabulary (5-9 terms) · Skill set (3-6) · Must verify (3-5 facts) · Sources of truth · Typical failure · Definition of done · Anti-patterns ``` What the material does not support, write as `unknown`. A guess instead of a fact is a defect: the next agent will read it as a rule. ## 3. Environment: what exactly to create ```text / AGENTS.md rules for this class of work (short: invariants plus domain bans) DOMAIN-CARD.md the domain card; inside it, an acceptance criterion a reviewer can test .dsh/skills//SKILL.md the specialization: visible in the palette only inside this environment examples/ two or three worked cases: one routine, one failure memory/ the experience store (M13): recipes/, failures/, tasks/, artifacts/ README.md what it is, how to use it, version and date ``` Content requirements: | File | Must contain | |---|---| | `AGENTS.md` | safety rules, no secrets, a verification step for every piece of work, domain boundaries | | `DOMAIN-CARD.md` | the acceptance criterion in one sentence, testable without asking the author | | `SKILL.md` | when to load it, what to do, how to verify; the domain's anti-patterns | | `examples/` | at least one failure with its cause — otherwise the examples teach only the happy path | | `memory/` | at least one recipe or failure record that passes `experience.py check` | ## 4. Registration: honest per runtime An environment works only when the runtime knows about it. This is the one step a model often cannot do, and it may never be reported as done when it was not. | Runtime | The model does | The human does | How to check | |---|---|---|---| | **HARNESS** | creates the whole directory with every file | adds the directory as a workspace (left panel → add workspace) | the directory appears in the workspace list | | **AutoClaw2** | writes the skill into `\.pi\agent\skills\` and the command file into `...\prompts\` | reloads the app so the resource snapshot is rebuilt | the command appears in the `/` palette, the skill in the list | | **SagaAI / other** | depends on whether it exposes a registration API | as the runtime requires | name the step instead of assuming it | How to report it: "the directory exists at X; one step remains that I cannot do: ". ## 5. Handoff: the starting prompt The last output is ready-to-paste text for the new window, not "now go and work": ```text Environment: Rules: this environment's AGENTS.md — read it first. Task: Acceptance: Start with: python tools/experience.py search "" --root . ``` ## 6. Verifying the result - Every file exists at the named path; no existing file was overwritten. - `python tools/experience.py check --root ` → `PASS`. - The acceptance criterion in `DOMAIN-CARD.md` is testable without asking the author. - The registration step is named and its effect is observable (an entry in the workspace list, a command in the palette). - The report ends with a gate line: `gate 12/12`, or `gate N/12 — failed: `. ## 7. Side mode: specializing the current session When no environment is needed — the task is one-off — the skill works as a session prompt: Task Card and Domain Card before acting, `[fact]`/`[inference]`/`[assumption]` labels, the twelve-item gate as one line. Same contract, without steps 3–5. Do not mix the modes: having promised an environment, deliver a directory, not a mood. ## 8. Safety 1. Credentials, tokens, keys, cookies and session data never enter the environment or `memory/`. Use ``, ``, ``, ``, ``, ``. 2. Personal data and customer content are not stored; the shape of a solution is. 3. Experience records are deprecated, never deleted. 4. Do not overwrite existing files: on a path collision, keep a backup beside it and say so in the report. ## 9. Kit contents | File | What it is | |---|---| | `PROMPT.md` / `PROMPT.ru.md` | the framework: 14 modules, 14 `Check` lines, 10 invariants, a 12-item gate | | `docs/INSTALL.md` | installation into AutoClaw2, HARNESS, SagaAI, with verification | | `docs/ADAPTATION.md` | how to fill the domain card | | `skills/experience-reuse/` | M13: record templates, examples, policy | | `tools/experience.py` | the store: `search`, `add`, `check`, `deprecate`, `selftest` | | `tools/install.py` | installs the skill and the command into a runtime, with `--check` | Version 1.2.0 · MIT · https://github.com/paulmann/agent-prompt-kit