dsh-tech-selection

Stop letting your AI guess. Get a method that finds the truth.

A model-agnostic research protocol for technology selection and solution comparison. Works with any AI agent — DSH, Claude Code, Cursor, Codex, or plain ChatGPT.

中文说明见 README.zh-CN.md

Why

AI is making technology decisions every day — choosing a database, comparing frameworks, vetting open-source options. Yet most agents research unpredictably: a strong model today, a weak model tomorrow, missing sources, trusting blog spam, no verification. The quality of the answer depends on the model's luck, not a method.

This skill fixes that: the method, not the model, guarantees coverage.

What

A fixed, model-agnostic 6-step protocol:

1. Requirement clarification — quantify constraints (scale/latency/availability/compliance/budget). No quantified constraints, no conclusion.
2. Multi-source retrieval — parallel queries; source tiers T1–T6 (primary docs > institutions > experts > editorial > community > content farms); GitHub API for live facts (stars/archived/updated).
3. Gap checklist — competitors × maintenance × ecosystem × local reachability × security/compliance × cost × migration × freshness.
4. Weighted evaluation matrix — weights aligned with the decision-maker before any verdict.
5. Traceable output — decision table + source tier & URL per key fact + "as of YYYY-MM" dates; no source, no conclusion.
6. Retro & iterate — capture gaps, backfill the protocol (CHANGELOG).

Quality gates between stages: no ≥2 independent sources → back to retrieval; weights not aligned → no verdict; facts not traceable → dropped.

Failure modes checked: SEO farms, early-retrieval contamination (57% of source errors happen in the first retrieval round), stale-as-current, survivorship bias, citation loops.

Install

DSH (DeepSeek Harness):

mkdir -p ~/.dsh/skills/tech-selection-research
cp tech-selection-research/SKILL.md ~/.dsh/skills/tech-selection-research/

Claude Code / Cursor / any agent:

mkdir -p .claude/skills/tech-selection-research
cp tech-selection-research/SKILL.md .claude/skills/tech-selection-research/

Or just paste SKILL.md into your agent's system prompt — it's self-contained.

Methodology sources

• MiniMax-AI/skills — technology-selection framework (requirements quantification + weighted matrix) · MIT
• source-evaluation — T1–T6 source credibility tiers · MIT
• Architecture Decision Record (ADR) — decision-recording concept by Michael Nygard

License

MIT