Application Engine

Skill: application_engine

Identity

Full-cycle job application skill for Agent Friday. Given a JobListing (and optional intel pack), the engine drives the entire pipeline:

  intel ──→ resume tailor ──→ cover letter ──→ form completion ──→ submission
                                                                       │
                                                                       ▼
                                                                  tracker log

Lives at skills/application_engine/engine.py. Designed to be invoked by the chat ("apply to this") or autonomously when auto_apply=true is set on a high-confidence priority job.

When this skill fires

• The user says "apply to that one" / "submit application" / "tailor for X".
• Auto-apply queue hits a job with confidence ≥ auto_apply_threshold.
• Re-application: an old listing reposts with substantially changed

details and quality_gates_passed includes relevance.

What this skill does

1. Intel pass. Pull company brief from cache or generate one. Stash

3 distinctive talking points.

1. Resume tailor. Pick a resume variant via A/B logic, swap in

role-specific bullets, ensure must-include items are present.

1. Cover letter. Draft against the listing's stated priorities, cap

at cover_letter_max_words. Run brand-voice check before saving.

1. Form completion. Detect ATS platform (Greenhouse, Lever, Workable,

SmartRecruiters, other) and emit a field-by-field plan.

1. Submission.

• Below salary floor → block and notify the user.
• Above confirmation threshold → require the user to OK before submit.
• Otherwise → submit and log.

1. Tracker log. Persist an ApplicationRecord with the resume +

cover letter variants used and which quality gates passed.

Inputs

• job_id — the JobListing to apply to
• force_confirm (optional) — request the user's explicit OK even when

under the confirmation threshold

• resume_variant (optional) — override A/B selection
• dry_run (optional) — produce all artifacts but skip submission

Outputs

{
  "application_id": "app_...",
  "job_id": "job_...",
  "status": "submitted" | "blocked" | "needs_confirmation" | "dry_run",
  "ats": "greenhouse" | "lever" | ...,
  "resume_variant": "AI_VP_v3",
  "cover_letter": "...",
  "cover_letter_word_count": 412,
  "quality_gates_passed": ["salary_floor", "must_include", "brand_voice"],
  "quality_gates_failed": [],
  "duration_ms": 11234
}

Quality gates

| Gate | Pass condition | |-------------------|---------------------------------------------------| | salary_floor | salary_max >= 150000 (configurable) | | salary_ceiling | If salary_max >= 300000, require confirmation | | must_include | Every must-include item appears in resume | | cover_voice | Cover letter passes brand-voice check | | cover_length | Word count ≤ cover_letter_max_words (450 default)| | dedup_apply | No prior application for the same job_id | | ats_supported | Detected ATS is in the supported list |

A failed salary_floor is a hard block; everything else surfaces a warning but doesn't stop submission unless dry_run=true.

A/B testing

The engine maintains a small bandit over resume variants. On every application, it picks a variant with epsilon-greedy (default ε = 0.10) and records the variant in the ApplicationRecord. When responses come back (positive or negative), record_response() updates the variant's score so future picks improve.

ATS handling

We support the common four major ATSes with platform-specific field maps:

| Platform | Detection signal | Notes | |-----------------|----------------------------------------|------------------------------------| | Greenhouse | boards.greenhouse.io in URL | Mostly text fields + dropdowns | | Lever | jobs.lever.co in URL | Has its own EEO block | | Workable | apply.workable.com in URL | Often gates by domain whitelist | | SmartRecruiters | jobs.smartrecruiters.com in URL | Heavy on cover letter parsing |

Anything else is unknown — we emit the field plan and ask the user to finish manually.

Quality bar

• Zero misfires on the salary floor gate.
• No application submitted twice for the same job_id.
• Cover letter brand-voice score ≥ 0.75 before submission.

How this skill improves itself

Each application is recorded to SkillOpt. When response data lands (interview, rejection, ghost) the engine attributes outcomes back to the application and folds them into:

• accuracy — was the application a real fit? (response_kind score)
• user_satisfaction — explicit user feedback
• completeness — did we fill every required field?
• cost_usd — token spend on tailoring

If the rolling outcome score drops, autoresearch proposes:

• Adjusting must-include lists,
• Trying a new resume variant,
• Cover letter style shifts (terser / more storytelling),
• A different brand-voice profile per role family.

Success criteria

A run is successful if it:

1. Surfaces a complete artifact bundle (resume + cover + field plan),
2. Passes all configured quality gates,
3. Either submits or hands off with all information ready for the user.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: FutureSpeakAI
• Source: FutureSpeakAI/Agent-Friday
• License: MIT
• Homepage: https://futurespeak.ai

Install and usage instructions live in the source repository linked above.