Taste

taste skill

Analyze a set of repositories and produce a TASTE.md document that captures the shared engineering conventions, architectural decisions, and product philosophy across the corpus.

features

• analyzes 2–20 repositories from a shared author, org, or domain to extract a transferable style guide
• explains what makes a codebase family cohesive before contributing or building something similar
• bootstraps new projects by inheriting conventions from repos you admire

usage

/taste owner/repo1 owner/repo2
/taste --html --output ./reports owner/repo1 owner/repo2
/taste --html --slides ~/Projects/tool-a https://github.com/owner/tool-b

Pass --html to produce both TASTE.md and a polished TASTE.html. Pass --slides (requires --html) to also produce TASTE-SLIDES.html — a fullscreen slide deck summarizing the corpus.

Do not invoke for a single file, a single tiny repo, or repos with no shared context.

workflow

step 1 — intake

Parse all arguments. Recognized flags:

• --html → generate TASTE.html in addition to TASTE.md
• --slides → also generate TASTE-SLIDES.html (requires --html)
• --output <dir> → write outputs to <dir> instead of ./

For each non-flag argument:

• Starts with https:// or git@ → treat as full git URL
• Matches owner/repo (no slashes except the one separator) → expand to https://github.com/owner/repo
• Starts with /, ./, ~/, or is an existing path → treat as local path
• Otherwise → try as https://github.com/<argument>

Create a staging directory:

TASTE_STAGE="${HOME}/.taste/corpus/$(date +%Y%m%d-%H%M%S)"
mkdir -p "${TASTE_STAGE}"
echo "Staging in ${TASTE_STAGE}"

step 2 — clone

For each git URL:

REPO_NAME="$(basename "${URL%.git}")"
DEST="${TASTE_STAGE}/${REPO_NAME}"

if [ -d "${DEST}/.git" ]; then
  echo "[cached] ${REPO_NAME}"
else
  echo "[clone]  ${REPO_NAME}"
  git clone --depth 1 --single-branch --quiet "${URL}" "${DEST}" \
    || echo "[unavailable] ${REPO_NAME}: clone failed"
fi

For local paths: use the path directly without cloning. Mark as [local] in the scope table.

If a clone fails: note [unavailable] in the scope table, skip that repo, continue.

step 3 — inventory each repo

For each available repo, collect:

A. Language composition — count source files by extension, excluding generated/dependency dirs:

find "${REPO}" -type f \
  \( -name "*.go" -o -name "*.ts" -o -name "*.tsx" -o -name "*.js" \
     -o -name "*.py" -o -name "*.rs" -o -name "*.swift" -o -name "*.rb" \
     -o -name "*.java" -o -name "*.kt" -o -name "*.cs" -o -name "*.cpp" \
     -o -name "*.c" -o -name "*.zig" \) \
  -not -path "*/node_modules/*" -not -path "*/.git/*" \
  -not -path "*/vendor/*" -not -path "*/dist/*" \
  -not -path "*/build/*" -not -path "*/__pycache__/*" \
  2>/dev/null | sed 's/.*\.//' | sort | uniq -c | sort -rn | head -5

B. Framework signals — search dependency manifests:

[ -f "${REPO}/go.mod" ] && \
  grep -iE "cobra|kong|bubbletea|urfave/cli|spf13" "${REPO}/go.mod"
[ -f "${REPO}/package.json" ] && \
  grep -oE '"(commander|yargs|meow|inquirer|clack|chalk|vitest|jest|mocha)[^"]*"' "${REPO}/package.json"
for f in "${REPO}/requirements.txt" "${REPO}/pyproject.toml" "${REPO}/setup.py"; do
  [ -f "${f}" ] && grep -iE "click|typer|argparse|fire|rich|pytest|ruff|black" "${f}"
done
[ -f "${REPO}/Cargo.toml" ] && \
  grep -iE "clap|structopt|tokio|serde|anyhow|thiserror" "${REPO}/Cargo.toml"

Also scan for MCP (Model Context Protocol) SDK dependencies:

# Go MCP SDKs
[ -f "${REPO}/go.mod" ] && \
  grep -iE "mcp-go|modelcontextprotocol|go-sdk" "${REPO}/go.mod"
# TypeScript / Node MCP
[ -f "${REPO}/package.json" ] && \
  grep -oE '"(@modelcontextprotocol/sdk|@anthropic-ai/mcp)[^"]*"' "${REPO}/package.json"
# Python MCP
for f in "${REPO}/requirements.txt" "${REPO}/pyproject.toml"; do
  [ -f "${f}" ] && grep -iE "^mcp[>=\[]|mcp-server" "${f}"
done
# MCP server entry point files
find "${REPO}" -maxdepth 3 \
  \( -name "mcp*.go" -o -name "mcp*.ts" -o -name "server.go" -o -name "mcp_server*" \) \
  -not -path "*/.git/*" 2>/dev/null | head -5
# MCP client config examples documented in README
grep -l "mcpServers\|mcp_servers\|\"command\":" "${REPO}/README.md" 2>/dev/null

C. Key top-level files present — check for each: README.md, CHANGELOG.md, LICENSE, Makefile, .goreleaser.yml, .goreleaser.yaml, package.json, Cargo.toml, pyproject.toml, go.mod, tsconfig.json, vitest.config.ts, jest.config.*, .golangci.yml, .github/, AGENTS.md, CLAUDE.md, CODEX.md, docs/, Dockerfile, install.sh, Formula/, mkdocs.yml, docusaurus.config.js, docusaurus.config.ts, astro.config.mjs, _config.yml, hugo.toml, hugo.yaml, conf.py, typedoc.json, .readthedocs.yaml

D. Test file count:

find "${REPO}" \
  \( -name "*_test.go" -o -name "*.test.ts" -o -name "*.spec.ts" \
     -o -name "*.test.js" -o -name "*.spec.js" -o -name "*_test.py" \
     -o -name "test_*.py" -o -name "*_spec.rb" \) \
  -not -path "*/node_modules/*" -not -path "*/.git/*" \
  2>/dev/null | wc -l

E. Badge inventory — extract full badge markdown from README.md to capture both the image URL and surrounding link/style metadata:

# Linked badges: [![alt](img-url)](link-url)
grep -oE '\[!\[[^]]*\]\([^)]+\)\]\([^)]+\)' "${REPO}/README.md" 2>/dev/null | head -30

# Unlinked badge images: ![alt](img-url)
grep -oE '!\[[^]]*\]\(https://[^)]+\)' "${REPO}/README.md" 2>/dev/null \
  | grep -viE '\[!\[' | head -10

# Banner check — first 10 lines, any image (badge or not)
head -10 "${REPO}/README.md" 2>/dev/null \
  | grep -oE '!\[[^]]*\]\([^)]+\)'

For each badge markdown found, record:

1. The full markdown expression (preserves link target, alt text, and raw URL with all query parameters)
2. The image URL domain — classify by type:
   • CI/Build — github.com/*/actions or */workflows/* badge
   • Coverage — codecov.io, coveralls.io
   • Docs — pkg.go.dev, godoc.org
   • Quality — goreportcard.com, snyk.io, deps.rs
   • Version/Release — shields.io/github/v/release, shields.io/npm/v
   • License — shields.io/github/license
   • Social — stars, forks, downloads
   • AI/LLM Docs — deepwiki.com, or shields.io badges whose link target points to deepwiki.com
3. From the raw query string, extract and note:
   • style= value if present (flat-square, flat, for-the-badge, or absent)
   • color= or labelColor= value if present (hex or named)
   • logo= value if present (the Simple Icons slug, e.g., npm, swift, node.js, homebrew, apple)
   • logoColor= value if present
4. Whether the badge is linked (wrapped in [...](url)) and what the link target domain is
5. Banner image: if the banner-check grep returns an image whose URL does not contain a known badge domain (shields.io, codecov.io, badgen.net, github.com/*/workflows, deepwiki.com), treat it as a project banner — record its relative path and alt text separately

Record badges in row order as they appear in README.md — order is itself a convention signal.

F. README section headings — capture the structural outline of README.md:

grep -E '^#{1,3} ' "${REPO}/README.md" 2>/dev/null | head -40

Record heading level, text, and order. Normalize for cross-repo comparison (e.g., "Installation", "Install", "Getting Started" → install slot).

step 4 — sample files by priority

For each repo, read files in this order. Stop after ~50 files per repo (token budget). Read up to 150 lines per implementation file; read config/CI files in full if under 200 lines.

P1 — CLI entry points and wiring (most signal-dense):

• cmd/*/main.go, main.go, cmd/root.go, internal/cli/root.go
• src/index.ts, src/cli.ts, bin/*.ts, bin/*.js
• main.py, __main__.py, src/main.rs, Sources/*/main.swift
• First file in cmd/, commands/, or internal/cli/ after root
• MCP server entry points: cmd/*/mcp*.go, internal/mcp/*.go, src/mcp*.ts, src/server.ts, any file matching *mcp*server*

P2 — Config and architecture:

• internal/config/*.go, src/config*.ts, config/*.py
• Any file matching *config* or *settings* in internal/ or src/
• Schema/migration files: *schema*, *migration*, prisma/schema.prisma

P3 — Tests (3 representative files):

• Prefer files named root_test, cli_test, main_test, cmd_test
• One unit test, one integration/CLI test, one config test if distinct

P4 — Tooling and CI (read in full):

• .github/workflows/ci.yml (or ci.yaml)
• .github/workflows/release.yml (or release.yaml)
• Any .github/workflows/ file whose name contains docs, pages, deploy, or preview — these reveal documentation CI patterns
• Doc-site config files if present (detected in Step 3C): mkdocs.yml, docusaurus.config.js, docusaurus.config.ts, astro.config.mjs, _config.yml, hugo.toml, hugo.yaml, conf.py, typedoc.json, .readthedocs.yaml
• Makefile (first 80 lines)
• package.json (scripts and devDependencies sections)
• vitest.config.ts, jest.config.*, pytest.ini, pyproject.toml
• .golangci.yml, .eslintrc.*, oxlint.json, biome.json, ruff.toml
• .goreleaser.yml or .goreleaser.yaml

P5 — Docs (first 100 lines each):

• README.md
• AGENTS.md, CLAUDE.md, CODEX.md (full, they are usually short)
• List docs/ and read 2–3 representative files

P6 — Output helpers (if present):

• Files in internal/out/, internal/output/, pkg/format/, src/output*, src/format*

P7 — Code evidence snippets (required, 6-12 short examples total):

• Import ordering and grouping from representative implementation files.
• Function/class docstring style from public APIs or framework extension points.
• Typing style: annotations, generics, overloads, Protocols/interfaces, typed config, or explicit untyped choices.
• Variable declaration and naming style: constants, module globals, local names, env/config names.
• Error handling style: wrapping, custom exceptions, stderr/stdout behavior, warning policy, validation failures.
• Test style: fixtures, parametrization/table tests, snapshot tests, assertion helpers, naming conventions.
• CI/linter examples: exact formatter/linter/type-check/test commands from config or workflows.

For each snippet, capture the repo, file path, and why the example is worth copying or avoiding. Keep snippets short: usually 5-20 lines, never more than 30 lines unless the surrounding context is essential.

step 5 — detect signals

For each signal family, record which repos show the pattern and classify:

• Strong (3+ repos) → top-level section in output
• Moderate (2 repos) → Style Guidelines mention
• Single (1 repo) → Repo-Specific Notes only

Analyze all 10 families: CLI surface, Architecture/data, Testing, Tooling/shipping, Documentation, Code philosophy, Code conventions, Project structure, AI collaboration, MCP integration.

See reference/signal-families.md for the complete signal definitions.

step 6 — write taste.md

Write ./TASTE.md in the current working directory (not the staging directory). Follow the output template exactly. Be prescriptive: write "do X because Y", not "they did X". Keep repo-specific observations in the repo-specific notes section; everything else should be cross-repo synthesis.

The report must include concrete evidence. Do not only describe style; show it. Include at least 6 short code/config snippets across at least 2 repos, and prefer all available repos when the corpus is 3+ repos. Every snippet must have a short caption explaining the convention it demonstrates. Good snippet topics: imports, docstrings, typing, constants/config variables, error handling, tests, and CI/linter commands. If a corpus has no usable snippet for a category, say that explicitly and explain what replaced it.

When mentioning specific tools by name — linters (ruff, black, flake8, golangci-lint, oxlint, biome), formatters (isort, gofmt, prettier), test frameworks (pytest, vitest, jest, inline-snapshot), release tools (goreleaser, semantic-release), and CI utilities (alls-green, zizmor, act, CodSpeed) — look up and include the canonical GitHub or documentation URL as a Markdown hyperlink on first mention in the Tooling section. Format: [tool-name](https://github.com/owner/repo). Prefer the source repository URL over a package registry page (PyPI, npm, pkg.go.dev). Subsequent mentions in the same section may use plain text.

step 7 — generate taste.html (only if --html was passed)

Fetches reference files from the thariqs HTML effectiveness corpus, selects appropriate visual components, and writes a self-contained TASTE.html with at least 5 distinct component types. The result should feel like an editorial analysis artifact — scannable, visual, and opinionated.

See reference/html-generation.md for the full design system reference, component selection guide, and rendering rules.

step 8 — generate taste-slides.html (only if --slides was passed)

Produces an 8–12 slide browser-native deck using scroll-snap, covering: title, scope, top patterns, pattern deep-dives, project structure, tooling/CI, best practices, anti-patterns, playbook quickstart, and close.

See reference/html-generation.md (Slides section) for the full rendering rules.

output template

See reference/output-template.md for the exact TASTE.md output format.

edge cases