Artifact Eval
A lightweight evaluation harness for the three artifacts that leave the system and reach a human: resume, cover-letter, outreach. It scores against a fixed spec, gives actionable fixes (never vague), revises only what's flagged, and re-judges independently until PASS. The generator never grades its own work.
When to use
- "Evaluate / grade / score this resume | cover letter | outreach message."
- "Is this good enough to send?" / "What would make this stronger?"
- Automatically as the final quality gate inside
resume-tailor,cover-letter, andnetworking-and-referralsbefore their output is handed to the user.
Inputs
- Reads: the artifact text (from the calling skill, a
workspace/path, or pasted), and theartifact_type(resume|cover-letter|outreach). For a resume/cover letter, also read the JD fromworkspace/pipeline/applications/<id>-*.mdso JD-match can be scored against real text. - Needs from user:
artifact_typeif not obvious. For outreach: who it's to and the intended ask, if not clear from the text.
Workflow
- Select the spec by
artifact_type(the three rubrics are inline below). If type is ambiguous, ask. Do not guess a rubric. - Stage 0,
writing-qualityDETECT. Run it first. Its output feeds the no-slop dimension AND is a hard input to the gate check (heavy slop can fail the gate outright). Never skip stage 0. - JUDGE (independent pass). Score every weighted dimension 0–100 and check each hard gate. Judging must be adversarial and specific: every deduction names the exact line/phrase and the fix. Vague feedback ("make it stronger," "add more impact") is not allowed. Say what to change to what.
- Compute weighted score = Σ(weight × dimension_score) / 100.
- Verdict:
- FAIL. Any hard gate fails (fabrication, over length, wrong/no recipient, etc.). Gate failures are non-negotiable. A beautiful résumé that lies or runs 3 pages FAILs.
- REVISE. All gates pass but weighted score < 80.
- PASS. All gates pass AND weighted score ≥ 80.
- If PASS → log and return (score + verdict + the artifact). Done.
- If REVISE (or a salvageable FAIL) → revise surgically. Fix only what the judge flagged and keep everything that scored well. Do not rewrite wholesale. That discards working material and drifts from facts. For a FAIL on fabrication, the fix is to cut or correct the false claim, never to dress it up.
- RE-JUDGE as a fresh pass. Re-run steps 2–3 on the revised artifact from scratch. The judge does not trust the reviser's self-assessment and does not rubber-stamp. Re-run stage 0 too (a revision can introduce new slop).
- Loop REVISE ↔ RE-JUDGE until PASS or until two revise cycles yield no score gain. Then stop and hand the user the specific blockers (usually a missing real fact only they can supply). Never fabricate to force a PASS.
- Log every cycle. Including failed ones, because the failures are the trend data.
The three specs (inline)
resume
- Hard gates (any fail → FAIL):
- No fabrication. Every title, date, metric, and claim traces to the user's real history/
profile/. No invented numbers. - ATS-parseable. Single-column-friendly, standard section headers, no tables/text-boxes/images/headers-footers that break parsers, real bullet characters, a plausible keyword match to the JD.
- ≤ 2 pages (≤ 1 for early-career).
- No fabrication. Every title, date, metric, and claim traces to the user's real history/
- Weighted dimensions (→ /100):
- JD / keyword match, 30 (does it mirror the JD's real requirements and language without stuffing?)
- Quantified impact, 25 (bullets show outcome + metric, not duties)
- Relevance / cut, 20 (irrelevant roles and bullets trimmed, with the top third targeting this JD)
- Clarity & scannability, 15 (strong verbs, consistent tense, skimmable in 10 seconds)
- No-slop, 10 (from stage 0)
cover-letter
- Hard gates:
- One page (~250–400 words).
- No fabrication.
- Addressed to a real person (named hiring manager/team lead). A generic "Hiring Team" is acceptable only when no name is findable. "To Whom It May Concern" fails.
- Weighted dimensions:
- Angle strength, 30 (a specific reason-to-believe / throughline, not a résumé in prose)
- Proof specificity, 25 (concrete evidence + numbers, tied to their needs)
- Company fit / insight, 20 (shows real knowledge of the company/role, not boilerplate)
- Voice + no-slop, 15 (sounds like a person and passes stage 0)
- Close, 10 (specific, confident next step, no empty CTA)
outreach
- Hard gates:
- < 90 words.
- Exactly one ask (one clear, low-friction request, not two and not zero).
- Genuine personalization (a real, specific detail about this person/company, not a mail-merge token).
- Weighted dimensions:
- Hook, 30 (the opener earns the next line, and is specific rather than "I hope this finds you well")
- Specificity / evidence, 30 (a concrete reason for reaching out, with credible relevance)
- Clarity of the ask, 25 (the one ask is unmistakable and easy to say yes to)
- No-slop, 15 (from stage 0)
Output
- Writes to: the calling skill's output file by default. If run standalone, it returns to chat (and appends to an eval log at
workspace/output/eval-log.md). - Format, every run:
- Score. Weighted /100 + each dimension's score.
- Verdict. PASS / REVISE / FAIL, with gate results (pass/fail per gate).
- Fix list. Specific, actionable, line-referenced. No vague notes.
- The revised artifact (if a revision ran), the final PASS-ing version.
- Cycle log line.
artifact-eval: <type> | cycle N | score M | verdict(all cycles, including failures).
Rules & quality bar
- The generator never grades itself. JUDGE and RE-JUDGE are fresh, adversarial passes. The reviser's opinion of its own work counts for nothing.
- Gates are absolute. No weighted score buys back a failed gate. Fabrication, over-length, or no real recipient = FAIL, full stop.
- Fixes must be specific. Every deduction says which line and what to change it to. Ban "tighten this" / "add impact" with no target.
- Revise surgically. Touch only what's flagged and preserve what scored well. No wholesale rewrites.
- Never fabricate to pass. If the only path to ≥80 is inventing a metric or a recipient, stop and tell the user what real input is needed.
- Log every cycle, failures included. That's the trend signal.
- PASS = all gates pass AND weighted ≥ 80. Both, always.
Related skills
- Runs:
writing-quality(DETECT) as stage 0. - Called by:
resume-tailor,cover-letter,networking-and-referralsbefore they finalize.interview-followupandlinkedin-presencerunwriting-qualitydirectly (no dedicated spec here yet). - Sibling pattern to: the ID/quality discipline in
application-tracker(structured, logged, no fabrication).
