| name | cvboost-audit |
| description | The CVBoost audit contract — how to fetch the profile, the rubric for the 6 criteria, report structure, hart publishing, and the exact issue-comment format the status page parses. |
Input shape (URL-based)
The issue body carries a Profil: URL and, optionally, a ## Profil complet block
(full profile text the user pasted).
First step, always: fetch the public data with cvboost-li-fetch "<url>". It emits
name, full headline, current company, education, location, and the public summary — for
any profile, from this worker's residential IP (the web app can't fetch it: LinkedIn
999-blocks the datacenter IP, so fetching is the worker's job).
- If a ## Profil complet block is present, audit it as the primary source; the
cvboost-li-fetch output is a cross-check.
- If only the public fetch is available, audit what you can (Titre fully; Cohérence
partially, from headline + companies + schools). For criteria you cannot judge from
public data (À propos, Expériences detail, Compétences, Réseau), assign a neutral 50
and state in that criterion: "Non évaluable depuis le profil public — pour un score
précis, relancez en collant le texte complet (Ctrl+A / Ctrl+C sur votre profil)."
Never invent content that is not in the input.
- Note near the top of the report whether it was based on the public profile only or on
the full pasted text.
The rubric (score each /100, justify by quoting the profile)
- Titre — searchable keywords + impact. Penalize: bare job title, buzzword soup, no value proposition.
- À propos — a pitch that convinces in 3-5 lines. Penalize: third-person, generic passion statements, no numbers, walls of text.
- Expériences — quantified, result-oriented. Penalize: responsibility lists with no outcomes, no metrics, unexplained gaps.
- Compétences — visible, aligned with the target role. Penalize: 50 unfocused skills, missing the profile's obvious core stack.
- Réseau — activity that reflects expertise. If the input gives no signal, say so and score 50 (neutral) rather than inventing.
- Cohérence — one credible narrative from title to last experience. Penalize contradictions and positioning drift.
Global score = weighted: Titre 20%, À propos 20%, Expériences 25%, Compétences 15%, Réseau 5%, Cohérence 15%. Round to integer.
Report structure (audits/.md, in French)
Audit LinkedIn — #
(one line: audit basé sur le profil public / sur le texte complet fourni)
Score global : NN/100
| Critère | Score | — table of 6
Diagnostic (exactly 3 sentences)
Analyse par critère (one ### per criterion: justification + 1 quick win)
Plan d'action (5-8 items, ordered by impact, each starting with a verb)
Footer: généré par un agent autonome mago · cvboost.intrane.fr
HTML report for hart
Self-contained: ONE file, all CSS in a tag, system-ui font, no external URLs
(fonts/images/scripts) — hart's CSP blocks them. Simple layout: score badge, 6
horizontal bars (width = score%), diagnosis, actions. French. Mobile-friendly
(max-width 720px, margin auto).
Delivery contract (the status page parses this — format matters)
- hart publish FIRST, then PR (body includes
Closes #<n>), then the issue comment.
- The issue comment must contain these two lines exactly (no bold, no backticks):
PR: https://github.com/javimosch/cvboost-audits/pull/
Artifact: https://hart.intrane.fr/a/cvboost/audit-
- Touch ONLY audits/.md in the PR. The HTML file is NOT committed — it only goes to hart.
- hart returns JSON; use the
url field verbatim.
