PromptKit
English · 简体中文
Repository version: 0.2.1 (local release preparation; publishing has not been performed). Upgrade history and migration notes (Chinese) · Technical changelog
Turn a rough draft into a structured, executable prompt — in one click, inside DeepSeek Harness.
dsh plugin --profile web add dsh-promptkit
Write a draft, hit ✦ Enhance, and receive a structured prompt in the preview panel, then in the composer without sending it. PromptKit needs no separate API key: semantic enhancement reuses a model already configured in DSH. If the session has no model route, it reports that requirement. Local lightweight enhancement needs no model.
Why PromptKit
✦ One-click enhancement with diagnosis. Before rewriting, the plugin examines your draft on five dimensions (clarity · hidden premises · falsifiability · actionability · context fit) and shows the verdict. The rewrite is driven by the diagnosis — undefined terms get defined, assumptions get flagged, vague requirements become testable ones.
🎯 21 thinking methods, matched automatically. A built-in library of complete Markdown methods (Socratic questioning, first principles, steel-man, minimal experiments…). The enhancer picks one based on your draft's signal words — or browse the full library in the Method Studio.
📚 A vault that closes the loop. Diagnosis findings (hidden premises, unfalsifiable requirements) can be saved as "to-verify" assumption cards. Verify them later; checked-in cards feed future enhancements as context. Your prompt quality compounds.
🔌 Zero telemetry, with a zero-token option. Local lightweight enhancement works offline. Semantic enhancement sends the draft and selected context to the host's configured model service. Optional project-memory searches make requests when that feature is used; @ file references come from the native DSH mention menu and are not reimplemented by the plugin.
Compatibility: verified by a real DSH 0.1.2-alpha.2 profile boot after installing the local bundle; CI runs the same alpha-channel smoke test. Running the DSH plugin requires Node >=22.6. The older 0.1.0-rc slot adapter is covered by simulated contract tests; this does not establish real-instance compatibility for every older release.
More capabilities (click to expand)
- PromptStudio — the advanced workspace: browse 21 methods, fill in facts/constraints, compose and preview a structured prompt before sending.
- PromptKit Vault — local library for drafts and finished prompts: search, favorites, project grouping, derivation with version diff, JSON backup/restore.
- Streaming output — enhancement results stream segment-by-segment into a preview panel with elapsed-time badge and cancel button.
- Strength levels — low (polish) / mid (standard) / high (expand ~3x) length budgets.
- Auto-enhance before send — available when a custom host provides
onSubmitDraft and composer.isInputTarget(). Intercepts only the composer's plain Enter; enhancement failure may send the original once, but send failure is never retried. Cancellation or a changed draft prevents sending. The standalone DSH plugin does not wire this send hook by default.
- Skill-mention preservation — lost
/tdd-style tokens are restored automatically. Dismissing the notice does not rewrite the draft again.
@file completion — searches the current session's workspace by file name without reading content. Entry, depth, and time limits bound indexing; incomplete results are labeled./pk quick insert — type /pk keywords for a compact vault candidate menu (arrow keys + Enter), never touching DSH-native commands.- Private methods — paste Obsidian-style Markdown prompt cards; stored locally only, exportable as JSON.
- Template variables —
{{name}} placeholders in vault items prompt a fill-in panel before insertion.
The enhancement loop
write a rough draft
│
▼
✦ Enhance ──► five-dimension diagnosis ──► model rewrite (streaming)
│ │
│ ▼
│ findings auto-staged in the Knowledge tab
│ │
│ you decide: save as assumption card
│ │
└──────────────► verified cards feed future enhancements ◄┘
Installation
npm (recommended)
dsh plugin --profile web add dsh-promptkit
GitHub (pin a commit for reproducible installs)
dsh plugin --profile web add github:fsrmqi/dsh-promptkit#<commit-sha>
Build artifacts are committed to Git — GitHub installs work out of the box without building.
tarball (offline / audit)
npm pack && dsh plugin --profile web add ./dsh-promptkit-0.2.1.tgz
After installing, refresh the browser. You'll find ✦ Enhance beside the composer and Advanced Method Studio as a conversation tab. Node-side changes also require the host to reload the plugin or restart DSH; refreshing the page alone may keep cached server code. Registry installs select a published version; use a tarball to test this checkout. See the upgrade steps.
Interface modes
The plugin defaults to simple mode, prioritizing draft → enhance → result. After 3 successful enhancements it expands the full interface (method library, vault, statistics, strength levels). You can lock either mode in Settings → Interface mode. Onboarding stores only progress from 0 to 3, independently of optional detailed statistics; failures and cancellations do not count.
Using it as an npm library
The core depends on zero hosts. Four decoupled interfaces; hosts swap implementations via the contract:
| Interface | Responsibility | Default |
|---|
MethodProvider | Method source / composition / templates / history | StaticMethodProvider (21 built-ins + private methods) |
Composer | Read/write a target input box | TextareaComposer |
Enhancer | Semantic model call | OpenAIEnhancer (any OpenAI-compatible endpoint) |
AssetProvider | Vault save / search / backup | StaticAssetProvider |
import { PromptStudio, QuickEnhancer, StaticMethodProvider, StaticAssetProvider, TextareaComposer } from 'dsh-promptkit/browser'
const methodProvider = new StaticMethodProvider()
const assetProvider = new StaticAssetProvider()
const composer = new TextareaComposer(document.querySelector('textarea'))
<QuickEnhancer methodProvider={methodProvider} assetProvider={assetProvider} composer={composer} messages={messages} />
Missing optional capabilities are hidden or degraded. Bundlers supporting the browser condition can also import the root package; its default Node entry additionally exports DSH registration. Runtime requirements are Node >=22.6 / React 17+; source development and tests use Node 24.15+. See the embed contract and verified upgrade combinations.
Embedding in other hosts (Embed Protocol v1)
Any React host can compose PromptKit's components via the standard artifact ui/embed.js: an IIFE that only exposes the PromptKit namespace, with the pk-* visual namespace isolated from host themes. Contract locked by tests. See docs/EMBED.md.
Privacy
| Access | Purpose | Can disable |
|---|
localStorage | Vault, diagnosis inbox, favorites, history, preferences, onboarding progress, optional detailed statistics | Detailed statistics default off; back up assets before clearing browser data |
| Target input box | Buttons or shortcuts apply results; custom hosts can explicitly wire automatic sending | Without a send hook, only the draft is changed; asynchronous writes check for newer edits |
| Local plugin requests | Semantic bridge, file-name completion, optional project-memory search | Triggered by the corresponding feature; file search does not read content |
| Configured model endpoint | Processes semantic drafts and selected context | Use lightweight mode to avoid model calls |
Zero telemetry. Usage statistics are not uploaded. Semantic enhancement is not offline: the configured model service may be external, so review drafts and selected context for sensitive information. Project-memory and custom-adapter data handling depends on the host implementation.
Contributing
Issues and PRs welcome — see CONTRIBUTING.md. To add a thinking method, drop a Markdown file under methods/ (frontmatter + a ## Prompt block) and run npm run build.
License
MIT
