AI Calibration
AI Calibration is how Skima AI turns your Job Description into the structured scoring rubric candidates are evaluated against. Instead of scoring against raw JD text, Skima extracts criteria and sorts them into four tiers:
- must-have
- preferred
- nice-to-have
- must-avoid
Every candidate is scored against this rubric, not the JD itself.
AI Calibration works only based on the Job Description and the skill evidence available in the candidate's resume and application data. If a requirement is not clearly written in the JD or cannot be found in the candidate's visible information, the AI cannot consider it during screening. Candidates may show unmet must-have or must-avoid criteria even when the missing information is simply not visible in their profile, which can affect their match score and pipeline ranking.
Where AI Calibration Appears
The scoring rubric is built once, during job creation, and can be viewed or edited from three places in the app.
- Step 3: Scoring criteria during job creation is where the rubric is first generated and published. Refer to the Create Job guide for the full workflow.
- View JD → Edit Job Description takes you back into the same Scoring criteria step for an already published job, where you can update the JD and rubric together. Refer to the View JD section of the Tabs and Filters guide.
- AI Calibration button in the View Candidates tab opens a read-only Scoring rubric modal showing every active criterion the job is currently scoring candidates against, organized by tier. This is a view-only summary. To make changes, you must go through View JD → Edit Job Description.
The AI Calibration button shows the active rule count inline, for example AI Calibration 9 rules. This count updates each time you publish a new version of the rubric.
Compliance and Bias Protection
Before the rubric is generated, Skima scans the JD for language that references a protected attribute or a known proxy for one, such as age, gender, national origin, or family status. Flagged language is excluded from scoring by default, and every decision is recorded in an audit log.
This protects your hiring process against generating knockout or scoring criteria that would create legal exposure under EEOC, the EU AI Act, or local equal-opportunity law, even when the bias is embedded in soft or coded language rather than an explicit statement.
Overriding a Compliance Flag
If you believe flagged language is necessary for the role, you can override the exclusion. Overriding requires you to confirm you understand the risk, select which tier the criterion should be placed in, and provide a written reason. The override, the reason, and your account are permanently recorded in the audit log.
Overriding a compliance flag creates a permanent audit record against your account and may create legal exposure. Only override a flag if you have a specific, defensible business reason for the role, and consult your legal or compliance team first. Compliance mode and override permissions are managed by your workspace admin per jurisdiction.
A criterion that has been overridden appears in the rubric with an [Ignored] prefix and an OVERRIDE tag in amber, both in the Scoring criteria step and in the read-only AI Calibration modal, so it stays visibly distinct from standard criteria at all times.
Refer to the Compliance Flags section of the Create Job guide for the full flagging and override workflow, including how to review the audit log.
How to Set Up or Update Calibration
- Navigate to 'Jobs' from the left sidebar and open the job you want to calibrate.
- Click the 'View JD' tab, then click 'Edit Job Description' in the top right. This takes you into the Scoring criteria step.
- Update the JD text if needed. Skima highlights the passages each existing criterion was extracted from, so you can trace what produced each rule.
- Click 'Generate scoring criteria' if the rubric needs to be rebuilt from the updated JD, or edit existing criteria directly: reweight them, move them between tiers using the up/down arrow icon, or delete them.
- Review any compliance flags before publishing. Resolve, dismiss, or override each one as appropriate.
- Click 'Publish scoring'. All candidates linked to the job are automatically re-scored against the updated rubric.
After publishing, check the View Candidates tab for any strong candidates who are now failing must-have criteria. If good candidates are being disqualified, the criterion may be too narrow. Adding equivalents under Also counts is often enough to fix match quality without changing the rubric structure.
Writing a JD for Better Calibration
Skima AI is capable of extracting must-have, preferred, nice-to-have, and must-avoid criteria from a JD even when requirements are written conversationally rather than as explicit bullet points. However, writing requirements clearly gives you more predictable control over which tier a criterion lands in, and criteria can always be moved to a different tier afterward if Skima's placement is not what you intended.
Minimum Years of Experience
State the number directly so Skima can apply it as a precise threshold.
Must have a minimum 3 years of B2B SaaS sales experience.
Candidates are required to have at least 5 years of experience in full-cycle recruitment.
Strictly 4+ years of hands-on experience in enterprise account management.Specific Tool or System Knowledge
Name the exact platform instead of using broad wording like "comfortable with tools."
Must have hands-on experience using Workday for recruitment operations.
Candidates are required to have experience working with Salesforce CRM in a sales role.
Experience with Greenhouse is mandatory for this role.Industry Experience
State the industry directly rather than writing a general preference.
Must have prior experience working in the healthcare SaaS industry.
Candidates are required to have experience selling to enterprise clients in the logistics domain.
Candidate should strictly have experience in B2B fintech sales.Certifications or Licenses
Write the certification as a strict condition when the role cannot move forward without it.
Candidates must mandatorily hold a valid Six Sigma Green Belt certification.
Project Management Professional certification is mandatory.
Must have a current AWS Solutions Architect certification.Location, Shift, or Availability
Write location and availability conditions clearly when they are non-negotiable.
Must be available to work in the Mumbai office five days a week.
Candidates must be willing to work night shifts in the US time zone.
Notice period should be strictly 30 days or less.Avoid writing any requirement tied to age, gender, national origin, family or marital status, religion, disability, or other protected characteristics, even indirectly through proxies like graduation year, physical stereotypes, or cultural fit language. Skima will flag and exclude this language from scoring by default. Writing compliant JDs from the start avoids compliance flags altogether and keeps your hiring process defensible under EEOC, the EU AI Act, and local equal-opportunity law.
In the Chrome Extension
The Chrome Extension shows the same read-only Scoring rubric used in the app. AI Calibration itself is configured only in the Skima AI app.
- Install and log in to the Skima AI Chrome Extension.
- Open any job where calibration has already been applied.
- Click 'AI Calibration' to view the active rule count and the full rubric broken down by tier, including any [Ignored] or OVERRIDE tagged criteria.
- Candidates who fail must-have or must-avoid criteria show a DISQUALIFIED tag with the specific disqualification reasons listed underneath, for example "Hard Skills - No mention of HubSpot or Pipedrive in the resume."
Best Practices
Separate must-haves from nice-to-haves. Mixing both in the same sentence makes it harder for Skima to weight them correctly. State mandatory conditions plainly and keep softer preferences in a separate line.
Use specific and measurable language. Broad wording like "good experience" or "strong background" gives Skima less to work with. Mention exact years, tools, industries, or certification names wherever possible.
Keep must-have and must-avoid criteria realistic. Only mark conditions that would truly remove a candidate from consideration. Too many strict conditions can shrink your shortlist further than intended and filter out strong candidates.
Review compliance flags carefully before overriding. An override is a legal and audit event, not a formatting choice. Only override with a genuine, defensible business justification and involve your legal or compliance team first.
Update calibration when the role changes. Skima scores against the rubric, not the JD. If requirements change, update the JD and republish the rubric so scoring reflects the current requirements.
If you are unsure how to write compliant, well-calibrated criteria for your specific domain or role type, reach out to [email protected] for guidance.