aimiaosu-blip Evidence-First Job Search
Build application materials that are persuasive because every important claim is traceable to evidence.
Choose the mode
Infer the smallest mode that satisfies the request:
- Discover: find current roles, verify them, and produce a ranked shortlist.
- Analyze: decompose one or more JDs and map requirements to candidate evidence.
- Tailor: create a role-specific resume/CV and a change log.
- Prepare: build evidence-backed interview stories, questions, and talking points.
- Full pipeline: run discovery through preparation when the user requests an end-to-end workflow.
If the user supplies a JD, do not repeat discovery unless current status or external facts matter. If the user requests live roles, browse current sources; do not answer from memory.
Establish the source of truth
Locate the candidate's base resume, evidence ledger, project notes, and constraints. Preserve the original resume unless the user explicitly asks to edit it.
When no evidence ledger exists, create one from the supplied material using assets/evidence-ledger-template.md. Assign stable IDs such as E001 to facts that may support application claims.
For each evidence item, record:
- exact fact and source;
- candidate ownership: led, co-owned, supported, or observed;
- metric definition, baseline, unit, and comparison period where applicable;
- confidence: verified, candidate-reported, or needs confirmation;
- safe wording and wording to avoid.
Read references/evidence-governance.md when the inputs contain metrics, ambiguous ownership, conflicting versions, or inferred skills.
Discover and verify roles
When searching live jobs, read references/discovery-and-verification.md.
Use the user's target geography, function, seniority, industries, languages, work authorization, schedule, and exclusion criteria. Search multiple relevant sources, prefer primary employer or ATS pages, and capture the retrieval date.
Treat a role as verified only when a current primary posting is accessible. Label aggregator-only or stale-looking results as unverified. Do not claim a role is open from a search snippet alone.
Hard gates override fit scores. Separate:
- eligibility and logistics;
- evidence match;
- career value;
- application cost and urgency;
- unknowns requiring confirmation.
Analyze the JD
Read references/jd-analysis-and-ranking.md for multi-role ranking or detailed analysis.
Separate explicit requirements from inferred preferences. Map each material requirement to evidence IDs and classify it as:
- Strong: direct, recent, and defensible evidence;
- Partial: adjacent or transferable evidence with a clear limitation;
- Gap: no support in the current evidence base;
- Unknown: evidence may exist but has not been provided.
Scores are decision aids, not ATS predictions. Always show the assumptions and let hard gates supersede the numeric result.
Tailor without drifting from the truth
Read references/application-packaging.md before generating a resume or complete application package.
Tailor by selecting, ordering, and clarifying supported evidence. Mirror JD terminology only when the underlying experience is genuine. Never invent employers, titles, dates, tools, ownership, scope, metrics, production use, customers, or outcomes.
Preserve distinctions such as:
- relative change versus percentage-point change;
- individual result versus team-level result;
- prototype or coursework versus production deployment;
- contributed to versus led;
- source documents versus customers, records, or projects.
Every new or materially changed resume claim must cite one or more evidence IDs in the change log. The final resume itself may omit the IDs for presentation.
Prepare for interviews
Build STAR or equivalent stories from the same evidence map. For each story include:
- suitable interview questions;
- a concise spoken version;
- likely follow-ups;
- metric and ownership boundaries;
- missing details the candidate should confirm;
- claims or phrasing to avoid.
Do not turn a proposed approach into past experience. Label transferable methods as “how I would approach it.”
Output contract
For a full pipeline, create one folder per role using assets/application-package-template.md:
applications/<company-role>/
|-- job-record.md
|-- fit-analysis.md
|-- resume-tailored.md
|-- resume-change-log.md
|-- interview-brief.md
`-- claim-audit.md
Also maintain a shortlist containing the primary URL, verification status, retrieval date, hard gates, evidence match, career value, deadline, and recommended next action.
Before delivery:
- Check every material claim against the evidence ledger.
- Mark unresolved fields explicitly; do not silently guess.
- Confirm that no hard gate is hidden by a high fit score.
- Explain what changed in the tailored resume and why.
- If files were generated, verify they open and contain the expected text.
- Run
python scripts/audit_claims.py <evidence-ledger.md> <application-file-or-folder>when the package uses[E###]citations.
Authorization boundary
Searching, analyzing, and drafting do not authorize submitting applications, messaging recruiters, editing external trackers, or changing public profiles. Stop before those actions unless the user explicitly authorizes them.