Dx Roaster

dx-roaster

Roasts a GitHub repo's developer experience. Produces four files: roast.md, action-plan.md, ideal-readme.md, score.json. Built for OSS maintainers and dev-tool founders preparing for Show HN, Product Hunt, or any launch where the first five minutes matter.

---

Critical rules (read before every run)

1. Read all three reference files BEFORE scoring. references/scoring-rubric.md, references/roast-voice-guide.md, references/readme-template.md. The rubric is non-negotiable — use the exact 10 dimensions.

1. Score every dimension on the 0–10 scale. Never skip. Missing data = score 0, not "N/A".

1. Roast must cite specific findings. Quote bad lines verbatim. Reference file paths. Cite line numbers when possible. Never write generic platitudes like "documentation could be better".

1. Tone matters. Default honest. Switch voice per references/roast-voice-guide.md. Brutal is funny but never personal. Kind is encouraging but still specific.

1. Action plan is impact-ranked, not severity-ranked. Use impact × (1 / effort). A 5-minute fix that adds 10 points ranks above a 2-hour fix that adds 12 points.

1. Never invent URLs. When writing ideal-readme.md, use `` syntax for missing links. Roast will tell maintainer to fill in.

1. Never name maintainers or contributors. Roast the work, not the people.

1. Output exactly four files in dx-roast/. Do not write to other locations. Do not skip files.

---

Step 1: Intake

Required:

• repo — local path (default .) OR https://github.com/owner/repo URL

Optional:

| Parameter | Default | Description | |---|---|---| | tone | honest | brutal / honest / kind | | output_dir | dx-roast/ | Where to write the 4 output files | | compare_to | none | Optional second repo URL for benchmark comparison |

If repo is a URL: clone to a temp dir using git clone /tmp/dx-roast-. Scan from there. Delete temp dir after run.

If repo is . or local path: scan in place, do not modify anything.

---

Step 2: Discover

Read these files (each may be absent — note absence):

• README.md — full contents, structure (headings), code blocks, image refs, badge URLs
• CONTRIBUTING.md, CODE_OF_CONDUCT.md
• .github/ISSUE_TEMPLATE/, .github/PULL_REQUEST_TEMPLATE.md, .github/CODEOWNERS
• package.json (description, homepage, repository, license, scripts), pyproject.toml, Cargo.toml, go.mod
• LICENSE (existence + type)
• docs/ folder contents (just file names + sizes)
• examples/ or samples/ or cookbook/ folder
• .github/workflows/*.yml (just names + brief CI presence check)

If repo is a URL, also fetch via gh:

gh api repos// --jq '{stars: .stargazers_count, forks: .forks_count, watchers: .subscribers_count, pushed_at, description, homepage, topics, default_branch}'
gh api repos///commits --jq '.[0].commit.author.date' | head -1
gh api repos///contributors --paginate --jq '[.[] | .login] | length'

Catch and continue if any GitHub call fails.

---

Step 3: Score 10 dimensions

For each of the 10 dimensions in references/scoring-rubric.md, output {score: 0–10, evidence: [...]} where evidence is a list of 1–3 short strings citing what you found.

Example:

"visual_proof": {
  "score": 4,
  "evidence": [
    "Found 1 screenshot at line 89 (below the fold)",
    "No animated demo / gif",
    "No video"
  ]
}

Sum scores → total (0–100) → grade (A/B/C/D/F per rubric).

Save to dx-roast/score.json with this exact shape:

{
  "repo": "",
  "scanned_at": "",
  "dimensions": {
    "time_to_first_success": { "score": 7, "evidence": ["..."] },
    "what_is_this_clarity": { "score": 5, "evidence": ["..."] },
    "visual_proof": { "score": 4, "evidence": ["..."] },
    "install_simplicity": { "score": 9, "evidence": ["..."] },
    "runnable_quick_start": { "score": 8, "evidence": ["..."] },
    "documentation_depth": { "score": 6, "evidence": ["..."] },
    "examples": { "score": 5, "evidence": ["..."] },
    "community_signals": { "score": 3, "evidence": ["..."] },
    "trust_signals": { "score": 7, "evidence": ["..."] },
    "marketing_signals": { "score": 4, "evidence": ["..."] }
  },
  "total": 58,
  "grade": "D",
  "tone": "honest"
}

---

Step 4: Write the roast

Read references/roast-voice-guide.md. Apply the requested tone. Write 250–500 words to dx-roast/roast.md.

Required structure:

# DX Roast: 

**Grade:  (/100)** · Tone: 

Every roast must reference at least 3 specific findings from score.json evidence arrays.

---

Step 5: Generate action plan

Compute priority = impact / effort_minutes for each candidate fix. Take top 5–10. Sort desc by priority.

For each fix, write to dx-roast/action-plan.md:

### N. 

**Why:** 

**Impact:** + points · **Effort:** 

**How:**

1. 
2. 
3. 

Open action-plan.md with:

# Action Plan — 

Current grade:  (/100)
Top fixes ranked by impact × effort. Do these in order.

---

---

Step 6: Generate ideal README

Read references/readme-template.md. Fill the template per "Filling rules" section. Write to dx-roast/ideal-readme.md.

Use placeholders `` etc. for missing values. Length cap: 100 lines.

---

Step 7: Terminal summary

Print this exact format to the user:

DX Roast complete — Grade:  (/100)

Top 3 fixes (impact-ranked):
  1.                                 + pts  · 
  2.                                 + pts  · 
  3.                                 + pts  · 

Files:
  → dx-roast/roast.md          (shareable verdict)
  → dx-roast/action-plan.md    (prioritized fixes)
  → dx-roast/ideal-readme.md   (target state)
  → dx-roast/score.json        (raw scores)

If compare_to was provided, add a comparison line:

Benchmark:  scores /100 (grade ) — gap of  points.

---

Self-QA (every run, before printing the terminal summary)

• [ ] All 10 dimensions in score.json have score + evidence
• [ ] Total equals sum of dimensions
• [ ] Grade letter matches the total per rubric thresholds
• [ ] Roast references at least 3 evidence items from score.json
• [ ] Roast does not name any maintainer / contributor
• [ ] Action plan has 5–10 items, sorted desc by priority
• [ ] Ideal README is ` for any URL not found in scan
• [ ] All 4 files exist in output_dir
• [ ] No ` / ` placeholders in roast.md or action-plan.md

If any check fails, fix and re-run that step.

---

Error handling

These four failure modes get explicit handling — never let the skill silently produce a partial output without surfacing the cause.

| Case | Behavior | |---|---| | repo doesn't exist OR can't be cloned OR can't be read | Exit immediately with the message Repo not found or inaccessible: . Do not write any output files. | | README.md is missing | Run anyway. Score what_is_this_clarity, visual_proof, runnable_quick_start, and marketing_signals near 0. Use the absence as the opening of the roast ("No README is its own statement"). | | Repo is private + no GitHub auth available | Try a local clone (git clone ); if it fails, exit with Repo is private and no authentication is available. Suggest the user set GITHUB_TOKEN or pass a local path. | | compare_to URL fails (network, 404, private, etc.) | Continue without the benchmark. Add a single line to the roast: Note: requested benchmark against could not be fetched. Omit the Benchmark: line from the terminal summary. |

Any GitHub API error (rate limit, transient 5xx) during Step 2 metadata fetch is caught and logged in score.json under trust_signals.evidence as "github metadata unavailable". The rest of the run proceeds.

---

Examples

Good invocation:

> "Roast my repo at https://github.com/myorg/mytool — tone: brutal."

→ Skill clones, scans, scores, writes 4 files to dx-roast/, prints summary. Total runtime "DX audit on the current dir. Benchmark against https://github.com/vercel/next.js."

→ Scans both, scores current, includes benchmark line in roast and summary.

---

What this skill does NOT do

• Does not modify the target repo (read-only scan)
• Does not run security audits or dependency scans
• Does not generate marketing copy (action items are concrete, not "rewrite for tone")
• Does not produce images or assets (just placeholders)
• Does not post to social media (writes shareable text only)

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Varnan-Tech
• Source: Varnan-Tech/opendirectory
• License: MIT
• Homepage: https://www.opendirectory.dev

Install and usage instructions live in the source repository linked above.