Tailor Resume¶
The Tailor step uses Claude AI to rewrite your resume sections based on the specific JD, with strict guards against hallucination.
Usage¶
How It Works¶
flowchart TD
Base[Base Resume] --> Split[Split into Sections]
JD[JD Analysis] --> Prompt[Tailoring Prompt]
Split --> |Tailorable sections only| Prompt
Prompt --> LLM[Claude AI]
LLM --> Modified[Modified Sections]
Modified --> Diff[Side-by-Side Diff]
Diff --> |User approves| Save[Save Tailored Version]
Diff --> |User rejects| Discard[Discard]
Split --> |Protected sections| Merge[Merge]
Save --> Merge
Merge --> Final[Complete Tailored Resume]
Anti-Hallucination Guards¶
The LLM prompt enforces strict rules:
- Only rephrase, reorder, or emphasize existing content
- Never invent skills, experiences, or achievements
- Never modify: Education, dates, company names, contact info
- Only modify: Summary, Technologies/Skills, experience bullet points
The system works by:
1. Sending only tailorable: true sections to Claude
2. Including the JD analysis (keywords, required skills) as context
3. Asking Claude to optimize for ATS keyword matching while preserving truthfulness
4. Setting temperature to 0.3 for controlled creativity
The Diff View¶
After tailoring, a side-by-side diff is shown for review:
+-- Original ----------------+-- Tailored -------------------+
| | |
| Experienced software | Experienced software |
| developer with expertise | developer with deep expertise |
| in Rust and Python. | in Rust, async systems, and |
| | cloud-native Python services. |
| | |
| - Built CLI tools in Rust | - Built high-performance CLI |
| | tools in Rust with async |
| | I/O and error handling |
+-----------------------------+-------------------------------+
| [y] Approve [n] Reject [e] Edit manually [r] Retry |
+------------------------------------------------------------+
Resume Versions¶
- Base resume is never modified
- Tailored versions only store the changed sections
- Full resume is reconstructed by merging base + tailored sections
- Export a tailored version:
kairos resume export <tailored-id>(--format pdf,htmlortex)
Export styles¶
pdf and html use the designed theme (src/kairos/render/themes/default.css): one A4
page with an accent rule, right-aligned dates, stack chips and track subheadings.
WeasyPrint prints the PDF. tex is plain pandoc LaTeX for anyone who wants to typeset it
themselves.
The theme reads a few Markdown conventions. Each is ordinary text anywhere else:
# Ada Example
**AI Agent Engineer** <- role line under the name
Sydney · open to Shanghai <- contact lines, links kept
ada@example.com · [ada.dev](https://ada.dev/)
## Work Experience
### Software Engineer — Acme, Sydney (Jun 2025 – Present) <- dates and place on the right
Multi-tenant ERP platform <- context line
#### Autonomous agents <- track subheading
monitoring → fix → eval → merge <- pipeline line
- Grew it from **5 to 40 fixes a day**. <- a bold number is highlighted
## Projects
### Site · portfolio · [ada.dev](https://ada.dev/)
Stack: Next.js, PostgreSQL <- chips
## Technologies
Agents: Claude Code, MCP <- two-column list
Tailoring keeps these lines, so every tailored version prints in the same design.
Gaps, coverage and reasons¶
kairos jobs analyze <job-id>
kairos jobs gaps <job-id> # have / supported / gap per job skill, before tailoring
kairos resume tailor <job-id> --intensity standard
kairos resume report <tailored-id> # the same report again
- Gaps. JD analysis stores each skill's English name and narrow synonyms. Even a
Chinese posting can then be checked against an English resume. No LLM runs: a skill
is
havewhen its name is in the base resume, andsupportedwhen only a synonym is, with the resume line quoted. Otherwise it is agap. Gaps are for you to raise in a cover letter or interview, never for the resume. - Coverage. After tailoring, the same check runs on the tailored resume: present, thin or missing.
- Reasons. Every changed line comes with one sentence naming the job requirement it serves. Lines that only moved are not listed.
- Claims to prepare. Every line with a number or an ownership word, since those are what an interviewer asks about.
- Intensity.
lightchanges wording only and keeps every bullet in place (checked).standardrephrases and reorders.strongrewrites bullets as action and result. Every level keeps the facts. Set the default with[resume] intensity.
Why: decision 0005.
Language, checks and page budget¶
[resume]
language = "" # en or zh for every tailored resume; "" = the base resume's
pages = 1 # 1: half a page to one page; 2: one and a half to two
- One language. A tailored resume is written wholly in
language. Job keywords in another language are translated, never copied or glossed in brackets. - Checks. Each tailored section is checked against the base section. It must not
add numbers, new names or technical terms, bracketed keywords, Chinese text in an
English resume, or buzzwords such as "spearheaded" or "leveraged". It must not add
ownership or scale words the original lacks ("led", "owned", "end-to-end",
"expert"), and must not change
###headings. A failing section is retried once with the failures listed. If it fails again, the export uses the base text. - Page range. A designed PDF must land in its range: half a page to one page with
pages = 1, one and a half to two withpages = 2. It is cut to fit in this order: lower projects; older roles' later bullets; the newest role down to half; then the rest. Tailoring puts the most relevant project and bullet first, so the cuts take the least relevant. A resume still outside its range after cutting is refused, with what to change, instead of being padded or squeezed (decision 0008).
Why these rules and where they come from: decision 0004. Why this design and engine: decision 0003.