kochab

Files (22)

• SKILL.md
• VERSION
• profile.example.json · assets/
• resume_builder.py · assets/
• cover_note.md · references/
• gap_analysis.md · references/
• intervention_dial.md · references/
• interview_prep.md · references/
• legitimacy_check.md · references/
• momentum.md · references/
• network.md · references/
• offer.md · references/
• pipeline.md · references/
• process_health.md · references/
• resume_builder_notes.md · references/
• scan_protocol.md · references/
• scheduled_task_template.md · references/
• setup_interview.md · references/
• stage_guidance.md · references/
• story_bank.md · references/
• study_plan.md · references/
• writing_up_a_role.md · references/

SKILL.md (first lines)

# Job Scan

The skill runs in several modes. Detect which one applies before doing anything.

- **Setup mode.** No `Master_Profile.md` in the working folder. Run Step A, then stop and confirm before scheduling anything.
- **Scan mode.** `Master_Profile.md` exists. Run Step B. This is what the scheduled task calls.
- **Draft mode.** She asks for a cover note, an application email, or to build the story bank, for a specific role. Run Step C. See `references/cover_note.md` and `references/story_bank.md`.
- **Plan mode.** She asks what to study, how to qualify for a role, or for a study plan. Run Step D. See `references/study_plan.md`.
- **Interview mode.** She asks to prepare for a specific interview. Run Step E. See `references/interview_prep.md`.
- **Pipeline mode.** She reports an application event ("I applied to," "I heard back from," "I withdrew"), or asks where her search stands or what her next step is. Run Step F. See `references/pipeline.md`.
- **Momentum mode.** She asks whether or when to follow up, for help getting a drafted application sent, what to do after a call, or to pass on a role she is not interested in. Run Step G. See `references/momentum.md`.
- **Stage-guidance mode.** She asks what a stage expects or what comes after it, or for help with references (who to ask, when, how to brief them). Run Step H. See `references/stage_guidance.md`. Prep for a specific interview at any stage stays Interview mode (Step E).
- **Offer mode.** She has an offer, or asks about evaluating or negotiating one. Run Step I. See `references/offer.md`.
- **Process-health mode.** She reports a rejection, wants to withdraw from a live process, or asks how her search is going overall (funnel). Run Step J. See `references/process_health.md`.
- **Intervention dial.** She asks the tool to be more or less hands-on ("dial the coaching up/down," "just hold things and let me ask"). Update the Intervention level in `Master_Profile.md` and confirm. See `references/intervention_dial.md`.
- **Network mode.** She wants to bring in her contacts (LinkedIn export or paste), asks whether she knows anyone at a specific employer she intends to apply to, or wants help reaching out to a contact. Run Step K. See `references/network.md`. Matching is intent-gated to one employer; never run it across the whole scan.

The scan is only as good as its honesty. A padded scan that reports four "strong matches" every day is worse than useless, because the day a real one appears it will not stand out. Most days produce little. Say so.

---

# Step A: Setup (first run only)

## A1. Read the resume

Ask for the resume file if one has not been provided. Read it. Extract, without embellishing:

- Every role: title, employer, dates, what she actually did.
- Education, degrees, field, institution, dates. Note anything incomplete or in progress.
- Certifications, licenses, patents, publications, talks.
- Tools, systems, languages, domains.

Resumes undersell and oversell in predictable places. Ask about anything that looks compressed ("what did that actually involve?") and anything that looks inflated ("were you the one doing that, or were you managing the person doing it?").

## A2. Interview for what the resume cannot say

Work through `references/setup_interview.md`. One question at a time. Do not batch them into a wall of text.

The interview covers: what she wants next, what she is moving away from, location and commute reality, remote tolerance, compensation floor and whether comp or mission ranks higher, hard nos, and how much of a level change she will accept in either direction.

## A3. Capture the honesty guardrails

This is the part people skip, and it is the part that makes the output trustworthy.

Ask directly: **"For each thing on your resume that sounds more impressive than it was, tell me the honest version."**

Concrete prompts that surface real answers:

- Where were you the owner, and where were you adjacent to the owner?
- Which technology do people assume you built, that you actually used or managed?
- Which title overstates the scope, and which understates it?
- Is there a credential people assume you have that you do not?
- Is there anything on here you would be embarrassed to be asked about in detail?

Write each answer as a rule in the profile, in the form *"say X, never say Y."* For example: *"Coordinated the vendor rollout, working closely with the engineering team. Never imply she wrote the integration code."*

Every future draft, resume variant, and cover note obeys these rules. They are not suggestions. If a role's screen requires overclaiming past a guardrail, the correct output is to name the gap, not to blur it.

## A4. Capture voice rules

Ask for words, phrases, and constructions she would never write, plus two or three samples of her own writing if she has them. Record as a Voice section in the profile. If a writing-style skill is available, use it too.

## A5. Write `Master_Profile.md`

Canonical source. Every future document tailors *down* from this file and never rebuilds from scratch. Sections:

`Contact` · `Search priorities` (including the comp-versus-mission ranking, in her words) · `Voice rules` · `Honesty guardrails` · `Intervention level` (how hands-on she wants the tool: light default, standard, or high; see `references/intervention_dial.md`) · `Experience` (reverse chronological, full detail) · `Skills inventory` · `Education` · `Publications / patents / credentials` · `Open items to resolve`

The story bank lives in its own file, `Story_Bank.md` (built in A5b), not in the profile.

## A5b. Seed the story bank

Follow `references/story_bank.md`. Read the resume and pull a draft story stub from every role or bullet that implies a real accomplishment. Then run a short interview that turns three to five of those stubs into full situation-task-action-result stories, capturing the honest framing for each (what she did versus what the team did). Write `Story_Bank.md`.

This is the raw material for cover notes and, later, interview prep. It does not have to be finished at setup; seed a few strong ones and let it grow as she applies. Ask how many she wants to do now, and offer to add more any time.

## A6. Build the target-employer list

Do not guess at this. Research it.

From her profile, derive her target titles and her target sectors. Then search for the organizations that employ people with those titles, in her geography and remote. Group them into **lanes** the way a search naturally divides. Most searches have two or three, for example: the core role in her primary industry; an adjacent application of the same skills in a different sector; and a part-time, contract, or teaching variant.

Write `Target_Employers.md`, organized by lane, with a one-line note on each employer explaining why it is there. Aim for breadth over precision. This is a seed list, not a cage.

Create `Discovered_Employers.md` (empty ledger, same columns as described in Step B) and `Application_Tracker.md` (empty, with the header and the three tables in `references/pipeline.md`: Active, Watching, Closed). Fill the capacity header from her setup answer (Part 3, question 9b); if she has no number in mind, leave it unset rather than guessing.

## A7. Hand off the schedule

Show her `references/scheduled_task_template.md` with her folder path and lanes filled in. Tell her to create a scheduled task with that text as the prompt.

Recommend daily. A daily scan that honestly reports "nothing new" most days is doing its job; a weekly one misses postings that close in five days.

**Do not create the scheduled task yourself unless she asks.** Confirm the profile is right first. A scan running against a wrong profile is worse than no scan.

---

# Step B: The scan (every run)

## B1. Read state before searching. Mandatory.

1. `Application_Tracker.md`. Every role in it, across all three tables (Active, Watching, Closed), is **already known**. A role is a duplicate if the same employer and role family already appear, even when the title wording, posting ID, or location differs. When unsure, it is a duplicate. Duplicates go under "already tracked, skipped," never in fresh finds. Schema in `references/pipeline.md`.
2. `Master_Profile.md`. Canonical content, voice rules, and honesty guardrails.
3. `Target_Employers.md` and `Discovered_Employers.md`. Both are in scope this run.

## B2. Discover, then sweep

Discovery first, because the seed list only finds what she already knew about. Full protocol in `references/scan_protocol.md`. In brief:

- **Role-first.** Her target titles crossed with qualifiers and the current month and year, across LinkedIn, Indeed, Glassdoor, Greenhouse, Lever, Ashby, Built In, and any board specific to her field.
- **Location-first.** Her titles plus her metro, her state, and neighboring metros. This is how local employers surface that no list would have named.
- **Aggregators and portfolio boards.** VC portfolio boards aggregate hundreds of companies at once.
- **Funding and news.** Recently funded organizations hire program and operations staff early. New organizations did not exist when the seed list was written.

Then sweep a **rotating subset** of `Target_Employers.md`. Rotate so coverage broadens across runs. Note which channels you used this run.

Favor postings from the last seven to ten days.

## B3. Score and assess each new role

From btmoriarty/kochab, MIT.

More from btmoriarty/kochab