dsh-blunt-dev

dsh-blunt-dev

English | 中文

A DeepSeek Harness plugin that injects one system-prompt section: a blunt senior-developer voice. Install it and every session in that profile stops writing like a college entrance exam essay and starts answering like the colleague at the next desk.

It targets DeepSeek's specific failure mode — not customer-service flattery but essay voice: ### headings over three-sentence answers, three bolded phrases per paragraph, emoji bullet leaders, a famous quote nobody asked for, and a closing that elevates the topic to the temperature of technology.

What it changes

• Banned words, in five intent groups, each covering its synonym family (openers, connectors, literary flourishes, teacher voice, closing prompts), plus a reword clause so "in general → ultimately" is not a way around the list.
• Fifteen banned behaviours: headings, bold spam, emoji, unrequested tables and quotes, closing elevation, metaphor stacking, self-Q&A, parallel triads, four-character idiom spam, question restatement, replaying the reasoning trace, six answers to one question, trailing "anything else?".
• Positive requirements with real before/after examples, including a faithful reproduction of the default DeepSeek output.
• Exaggeration guards: accuracy beats style, fabrication is the red line (say "not sure, go verify" instead), formal prose still available on request.
• Deep-think clause: the reasoning trace is scratch paper, never the reply.

Install

From GitHub:

dsh plugin --profile web add github:rosemary-ewe/dsh-blunt-dev

From npm, once this package is published there:

dsh plugin --profile web add dsh-blunt-dev

Restart DSH after installing. The row lands in the profile's bundle stack; from there the market's enable/disable toggle and the profile patch file both work on it.

Configure

The insert row in cordis.patch.yml is yours to edit:

- id: blunt-dev
  name: dsh-blunt-dev
  config:
    enabled: true            # false disables without uninstalling
    extraBanned: []          # your personal blacklist, highest priority
    extraRules: ''           # extra rules appended after the style block
    order: 46                # after persona (0), before tool guidance (1000+)

extraBanned is the useful one: when a phrase starts grating, add it to the array and the model is told to rewrite the whole sentence if it shows up — synonym substitutions included. Values in extraBanned and extraRules are inserted into the system prompt verbatim, so only put text there whose instructions you would actually trust.

Verify

dsh --profile web --dump-config   # the blunt-dev row is in the composed tree

After cloning the repo, npm test covers registration, section order, and the config options.

Compatibility

Built against dsh 0.2.0-rc.2 and @deepseek-ai/dsh-system-prompt 0.2.0-rc.2. The plugin only calls ctx.systemPrompt.section(), the registry API that has been stable across the rc line. No runtime dependencies, no build step, no install scripts.

License

MIT