Experience Bank

Build and maintain a deep, reusable pool of the candidate's achievements so resume-tailor can surface role-relevant experience that doesn't fit on the (necessarily short) base resume.

This is a periodic maintenance skill — run it once to seed, then refresh every quarter/semester or after shipping something notable. It is not part of the per-application flow.

experience-bank  ──writes──▶  .mercury/experience/   (deep, tagged pool)
                                      │ read-only
                                      ▼
profile-optimizer → job-scout → resume-tailor → recruiter-outreach
                                      └─ selects role-relevant entries

Prerequisites

• LinkedIn MCP — get_person_profile (seed material, gap detection)
• The user's base resume (if present) and the existing bank in .mercury/experience/

Data Directory

.mercury/
└── experience/
    ├── {slug}.md        # one achievement/project per entry (frontmatter + STAR body)
    └── index.md         # rollup table for quick scanning

Entry frontmatter schema (tags drive cheap per-role retrieval):

---
title: Cut checkout p99 latency 40%
slug: checkout-latency-40pct
skills: [performance, distributed-systems, observability]
tech: [java, aws, sqs, dynamodb]
domain: [payments, e-commerce]
role_type: [backend, platform]
metrics: ["p99 -40%", "2M events/day"]
scope: "team of 6, 2M daily orders"
dates: "2024-Q3"
on_resume: false        # is this already on the base resume?
source: interview       # interview | freeform | linkedin | resume
---

Workflow

1. Load What Already Exists (never start from scratch)

• Read every entry in .mercury/experience/ (if any) → know what's already captured
• Read the base resume (.mercury/base/resume.*) if present
• Pull get_person_profile(...) for experience/projects already on LinkedIn
• Build a mental set of "covered" achievements so you only ask about gaps and new material

If the directory doesn't exist, create it — this is the seeding run.

2. Grill the Candidate (interactive elicitation)

Conversational interview loop using a STAR frame. Probe for the things people leave off resumes:

• Situation/Task — what was the context and the problem?
• Action — what did you specifically do? (separate your work from the team's)
• Result — quantified impact: metrics, %, scale, revenue, latency, users
• Scope — team size, budget, scale (req/s, data volume, $)
• Tech — the stack actually used (not aspirational)

Rules:

• Ask one focused question at a time; follow the thread before moving on.
• Push for numbers when answers are vague ("how much faster?", "how many users?").
• Only spend questions on new or thin material — skip anything already well-captured.
• Stop when the user is out of new material or says they're done.

Also accept freeform dumps: if the user pastes a raw story, structure and tag it without re-interviewing.

3. Structure & Store (incremental, idempotent)

For each new/updated achievement:

• Write .mercury/experience/{slug}.md with the frontmatter schema + a tight STAR body
• Dedupe: if it matches an existing entry, update it rather than create a duplicate
• Set on_resume: true for things already on the base resume (so tailor knows what's "extra")
• Regenerate index.md — a table of all entries (title, skills, tech, domain, metrics)

4. Persist to Mercury

Log the maintenance run so the dashboard reflects it:

mercury activity log --skill experience-bank --summary "Captured {N} new entries; bank now {total}"

If mercury isn't installed, just write the files — the bank still works standalone; only the dashboard activity log is skipped.

Truthfulness (non-negotiable)

• Capture real stories and real numbers — the skill structures and reprioritizes, it never invents achievements, metrics, or scope.
• If the user can't quantify something, store it qualitatively rather than fabricating a number.

Integration with resume-tailor

At application time (no interview), resume-tailor:

• Matches the job's requirements/stack against entry tags (skills, tech, domain, role_type)
• Selects the best-fitting entries — even if they aren't on the base resume
• Keeps the gap report honest: distinguishes "you have this, it's in your bank (not your resume)" from a genuine miss

Example

User: "grill me — I shipped a few things this quarter"

Agent:
1. Loads existing bank + base resume + LinkedIn profile
2. "You don't have anything tagged 'observability' yet. Did you do any
    monitoring/alerting work recently?"
3. … STAR follow-ups, pushing for metrics …
4. Writes .mercury/experience/checkout-latency-40pct.md + updates index.md
5. mercury activity log --skill experience-bank --summary "Captured 3 new entries; bank now 14"