Introduction

Most design review plugins use a fixed rule set immediately after installation. Their evaluation criteria remain unchanged no matter how many times they are invoked. dsh-palate takes the opposite approach: it maintains an accumulating “taste corpus” and encoded principles. By supplying examples and principles to the plugin and recording review feedback, its evaluation criteria sharpen with use. The plugin is maintained by guo6x and belongs to the memory-type plugin category.

Core Features

The plugin provides a toolchain for observing, organizing, confirming, reviewing, and calibrating design taste.

  1. Observation and organization

    • palate_intake: stages structured visual analysis as candidate examples or principles. It does not directly modify the taste library; it is used only to submit content awaiting confirmation.
    • palate_candidates: shows pending, accepted, or rejected visual training candidates along with their source sessions.
    • palate_decide: based on an explicit user decision to accept or reject, formally admits or retains a candidate in the taste library. This is the only path that converts a candidate into a taste library record.

  2. Review and learning

    • palate_review: when reviewing a new design, invokes the accumulated taste (principles + relevant past examples) as context, so evaluation judgments are based on learned judgment rather than a generic checklist.
    • palate_add: directly adds examples to the corpus (marked as good/bad/note + reason + tags).
    • palate_learn: distills new principles from experience and encodes them into the recorded taste.

  3. Reference and feedback

    • palate_packs: views available visual reference packs and their application status.
    • palate_seed: applies one or more visual reference packs once (such as apple-product-storytelling), without overwriting the existing taste library.
    • palate_feedback: records review outcomes by review_id (whether effective and why); only accepted principles accumulate evidence.

  4. Management

    • palate_list: browses the accumulated corpus.
    • palate_principles: lists encoded principles, sorted by evidence.
    • palate_effectiveness: views data on which principles have been accepted or rejected in real review feedback.
    • palate_stats: displays accumulated statistics (such as the number of examples analyzed and the number of principles distilled).

Typical Usage

1. View and apply reference packs

First view the available reference packs, then explicitly apply them.

# View available reference packs
palate_packs

# Apply the Apple and X reference packs (once only, without overriding existing taste)
palate_seed apple-product-storytelling x-direct-utility

2. Visual intake and decision

Use a browser or visual tools first to analyze the page, then submit the decision.

# Assuming a vision tool analyzed the screenshot or page
palate_intake {hierarchy: ..., typography: ..., color: ..., spacing: ...}

# View pending candidates
palate_candidates

# After user confirmation, execute the decision
palate_decide {candidate_id: "...", decision: "accept"}

3. Tag-based review

Use tags during review to filter specific examples and principles within a particular style scope.

# Review a product launch page, citing only Apple-style examples and principles
palate_review {subject: "...", tag: "apple"}

4. Record feedback

At the end of a review, record the outcome to update effectiveness data.

# Record whether the review was effective, and accept or reject the relevant principles
palate_feedback {review_id: "...", outcome: "effective", accepted_principles: [...]}

Installation and Enablement

In a DeepSeek Harness environment, install it using the following command:

dsh plugin add guo6x/dsh-palate

The plugin requires Node.js >= 22 and uses the built-in node:sqlite for storage. After installation, if the local taste database is empty, it automatically inserts 12 base principles and 4 educational examples.

Storage and Characteristics

  • Storage path: the plugin stores the SQLite database under $DSH_HOME/palate/ and generates human-readable mirror files (such as taste.md, principles.md, feedback.md, and training.md).
  • Startup packs: by default, it includes two optional visual reference packs: apple-product-storytelling and x-direct-utility. These are transparent abstractions from publicly available pages dated August 27, 2026, and are not scraped assets, brand copy, or templates.
  • Safety mechanism: installing or upgrading the plugin does not overwrite existing taste library data. Only candidates confirmed through palate_decide are added to the taste library.

Suitable Scenarios and Cautions

The plugin is suitable for scenarios that require ongoing improvement of design judgment. When using it, note the following:
* The plugin does not automatically scrape screenshots or perform visual recognition by URL. You must first use a browser or visual tools (such as modlens_read_image) to obtain evidence, and then pass it to palate_intake or palate_review.
* Because the plugin runs with the permissions of the current DSH process, it is recommended to review the source code and license (MIT) before installation.

Summary

dsh-palate provides a closed-loop workflow from observation to confirmation and from review to feedback. It does not impose a fixed set of design standards; instead, it provides the ability to accumulate taste. Through explicit decision paths and feedback records, an Agent can reference more project-specific judgment criteria in subsequent reviews.