Kochab

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

Score every new role on the six-dimension rubric in references/scan_protocol.md (core-skill, level, domain, location, mission, screen-clearance; 0-3 each; /18 total). The score rolls into exactly one rating: Strong, Stretch, or Skip, with the capping rule that a zero on core-skill, level, or screen-clearance holds a role at Stretch however high the total.

Show the six sub-scores and the total together, always. Never a bare number, never a percentage. The score ranks the list and shows its reasoning; it does not replace the gap sentence.

Give each ranked role a one-line why-you-fit rationale alongside the gap: the single strongest reason she is a real candidate here, in her terms. The gap says what is missing; the why-you-fit says what is already true. Both, always, so the picture is honest in both directions.

On every Stretch role, add one directional line after the gap: the single dimension holding it below Strong and what closing it would do to the rating. Only nominate a lift that is a concrete, studiable skill she can go learn and get good at (a Python course, SQL, a named tool, a real cert screen). If the gap is experience (domain seasoning, tenure) or structural (mission, level, location), say plainly that no course moves it rather than inventing a study target. Claim a move to Strong only when the arithmetic supports it (14 or more, no capping zero on core-skill, level, or screen-clearance). It is a pointer, not a plan: the full study plan stays on demand and in the weekly summary. Method in references/scan_protocol.md.

Run the legitimacy check (references/legitimacy_check.md) as part of verification. Attach a caution to any posting that is not a clean, live, direct role; keep a clear scam out of the ranked list and note it as screened out.

Name the real gap in every case, specific enough that she can act on it. Do not inflate to be encouraging. Do not soften a Skip into a Stretch. A role asking for two years when she has fifteen is a level-down, not a Strong.

Rank by the /18 within each band, then by whatever she said she values, in the order she said it. Weight a no-relocation location through dimension 4, not as an afterthought.

B4. Draft materials for the top one to three Strong matches only

• Reuse an existing resume variant if one fits. Check the folder before building anything.
• If no variant fits, build a new one with assets/resume_builder.py. See references/resume_builder_notes.md.
• Draft a short "Why this organization" paragraph, 200 to 400 words, grounded in the profile and obeying the voice rules.
• Name any screen she may not clear (a certification, a degree field, a years-in-domain bar) so she decides rather than discovers it in a rejection.

B5. Gap analysis and study recommendations

For every Strong and Stretch match, produce the two-part read described in references/gap_analysis.md:

• Emphasize. Which parts of her record to lead with for this specific posting, and which to cut. Resumes are tailored by subtraction more than by addition.
• Build. What she would need to learn, earn, or do to turn this Stretch into a Strong, with an honest estimate of how long. Distinguish a credential she could earn in a quarter from a domain that takes three years.

Across runs, watch for the requirement that keeps appearing in roles she wants and that she keeps not having. That recurring gap is the highest-value thing she could go fix. Say it plainly when you see it, even if she did not ask.

On the weekly-summary run only (see references/study_plan.md, "Weekly cadence"): aggregate the recurring gaps from the last roughly seven scan notes and the tracker. If a gap is both recurring and bounded (closes in a quarter or a year), surface a study-plan offer in that day's note: name the gap, the roles it blocked, a rough time-to-close, and ask if she wants a plan. Do not surface the offer on ordinary days, and do not generate the full plan unprompted; generate it only when she accepts or asks. This keeps the daily note lean while still catching the pattern.

Also on the weekly-summary run only, add a short follow-ups coming due line from the tracker: Active rows whose suggested follow-up date (per references/momentum.md) falls in the coming week, and any drafted-not-sent Watching rows. State them once, as facts she can act on; do not repeat a nudge already raised, and keep this off the daily note.

B6. Save and update

• Scan note: job_scan_YYYY-MM-DD.md in the working folder.
• Any new resumes and cover notes, same folder.
• Append genuinely new employers to Discovered_Employers.md. Never delete a row; mark stale ones stale. Columns: Employer | Lane | How found | Location | First seen | Notes.
• Add new finds to the Watching table in Application_Tracker.md (Role | Org | Lane | Fit | Gap | Found | Link | Notes). Never duplicate a row already in Active, Watching, or Closed. The scan writes only to Watching; moving a role into Active happens in Pipeline mode when she reports she applied.
• Verify every posting link is live before sharing it. Career pages are frequently JavaScript-rendered and return empty; confirm through a board mirror or a search rather than assuming.

B7. Close with a short summary

Channels used. How many genuinely new roles, by lane. How many were already in the pipeline and skipped. The top new match. What you drafted and saved. Any new employers added to the ledger. Any recurring gap worth acting on.

If nothing new turned up, say that. Thin days are the normal case and reporting them accurately is the point.

If the tracker has any Active applications, close with a one-line pipeline snapshot: how many are active against her capacity target, and the single soonest next action with its date. State, not a nag: it is there so the whole search stays visible on a thin scan day, not to push her to act. The full timing view stays in Pipeline mode (Step F), on request. Do not generate follow-up prompts here; follow-up timing lives in Momentum mode (Step G) and surfaces on the weekly-summary run, not the daily scan close.

---

Step C: Draft a cover note (on request, or offered for a Strong match)

Full method in references/cover_note.md. In brief:

1. Confirm the target. A specific role and employer. If Story_Bank.md does not exist yet, seed it first (A5b); a note without stored stories has nothing personal to draw on.
2. Research one true hook. A short company-research pass (a subagent is ideal) to find one current, specific, real thing about this employer she can honestly speak to. Never invent it. If there is no genuine hook and she has no real reason to care about the company, say so rather than manufacture enthusiasm.
3. Pick one or two stories from the bank that answer the single most relevant requirement, not every bullet.
4. Assemble hook, one demonstrated claim carried by a story, a forward-looking line, and a close. Short. Never restate the job description. No "I am writing to express my interest."
5. Produce two lengths from the same inputs: a short application email (six to eight sentences) and a slightly longer formal letter. Offer both; she picks.
6. Obey the voice rules and honesty guardrails. Save as a draft in the folder. She edits and sends. The tool never submits.

---

Step D: Build a study plan (on request, or after she accepts a weekly offer)

Full method in references/study_plan.md. In brief:

1. Find the target. Infer from what she gives you: a skill or topic she names, a specific role she points at, or nothing, in which case pick the highest-leverage recurring gap from her recent scan notes and tracker. "What do you recommend" right after a scan means the specific skill that role's directional line named. If there is no history and she names nothing, infer likely gaps from Master_Profile.md and Target_Employers.md, name them, and ask which to plan for rather than guessing.
2. Set the time budget, the primary tuning input. Default to the hours per week from setup, but take any figure she gives in the moment (15 minutes a day, four hours a day) and normalize to hours per week. If none is known, ask once. The budget changes what the plan is made of, not just how long it runs: lean budgets get one thread and the smallest artifact over a longer horizon; immersive budgets get a project-first, parallel, compressed plan. See references/study_plan.md.
3. Build the plan. Target as a demonstrable artifact, why this gap over the others, a path of real web-searched resources shaped and paced to that budget, the portfolio artifact that proves the skill, and a re-assessment checkpoint.
4. Stay honest. Be skeptical of certifications. Real resources only, verified to exist. Time estimates she can actually live. End by returning the decision to her.
5. Save as Study_Plan_<topic>_<YYYY-MM-DD>.md in the folder.

---

Step E: Interview prep (on request)

Full method in references/interview_prep.md. In brief:

1. Get the specifics. Which role, company, and stage (recruiter screen, hiring-manager call, panel, onsite); the format if known; anything she was told. If she does not know the format, prepare for the most likely next stage and say so.
2. Map questions to stories. Pull the questions this role and level actually ask, and for each name the story from Story_Bank.md that answers it best and how to frame it here. A likely question with no story is a gap to flag, not a blank to bluff. Seed the bank first if it is thin.
3. Give her questions to ask, three to five specific ones drawn from the company and role, not generic.
4. Name the thin spots and give an honest way to handle each: what to acknowledge, what adjacent experience to pivot to, what not to overclaim. A rehearsed true answer beats an improvised inflated one.
5. Prep for the specific format. Recruiter screen, hiring-manager screen, panel or sequential one-on-ones, or onsite/final each shift the emphasis and logistics; see "Prep by format" in references/interview_prep.md. The screen is the cheapest filter and cuts on avoidable things, so weight the basics there; a panel is a consistency and stamina test; an onsite is logistics as much as content. Keep the whole thing to a page she can review. Save it in the folder.

---

Step F: Pipeline (on request, whenever she reports an event or asks where things stand)

Full method in references/pipeline.md. The pipeline is the working-memory offload: one file, Application_Tracker.md, holds the state of every parallel application, the next action for each, and how their timing collides, so she does not have to. Pipeline mode holds and shows that state. It does not tell her when to follow up (that is Momentum, Step G), coach a stage (Stage guidance, Step H), or diagnose her funnel (Process health, Step J). In brief:

1. Read the tracker first. Three tables: Active (applied and everything live after it, with a Stage of Applied, Screen, Interview, Final, or Offer), Watching (scored but not yet sent, where scan finds land), and Closed (the archive; never deleted). Every role in any table is already known to the scan.
2. Make the smallest write that matches what she said. Log an application (Watching row moves to Active, Stage Applied). Advance a stage. Set or change a next action and its due date. Close a role with an outcome (rejected, withdrawn, declined, accepted). Add a watch by hand. Match the role she names to a row by employer and role family, the same way the scan dedups; when unsure which row, ask rather than guess. Record only what she reports, never invent a status or a follow-up date, and confirm the one thing that changed.
3. Report status when she asks ("where does my search stand," "what's my next step"). Produce the timing view from references/pipeline.md: counts by stage, Active against her capacity target, next actions sorted by due date (soonest first, blanks last), and any week where two or more live processes collide. It is a mirror of the state she recorded, not a coach. Capacity is a target she set, reflected back once as a fact when she crosses it, never enforced.
4. The tool never applies, advances, or follows up on its own. She reports what she did; the tool records it. ETHICS principle 1 holds here as everywhere. Save the file after any write.

---

Step G: Momentum (on request, to keep the search moving)

Full method in references/momentum.md. Where Pipeline mode (Step F) holds and shows state, Momentum acts on it to close the quiet gaps where a search stalls. Everything here is a suggestion she acts on; the tool never sends, submits, or contacts anyone. Light-touch by default: intensity is tuned by the intervention dial (references/intervention_dial.md), conservative at its light default. Four beats:

1. Get applications out the door. Surface drafted-not-sent roles (a Watching row whose next action is to finish or send): name the one thing left, offer to finish the draft (Step C), confirm any real posting deadline. Never manufacture urgency, and never submit.
2. Follow-up timing. Suggest when she reaches out, and offer to draft the note. Default cadence is conservative: after applying, a light check-in after about a week and at most one more, then rest; after a call, a thank-you within a day, then a status check on the employer's stated timeline or about a week if none was given. An employer-stated timeline always wins. Cap at two nudges per waiting period, honor "no follow-ups on this one," and never guilt-trip a passed date. Surface on request and once a week on the weekly-summary run, never in the daily note. How actively follow-ups surface follows the intervention dial (references/intervention_dial.md); the caps and off switch hold at every setting.
3. Post-call next actions. Right after she reports a call happened, run a short capture: advance the stage, set a thank-you as the immediate next action, schedule the next touch from the timeline they gave, and note one line on how it went. Capture and schedule only; how to run the next round is stage prep (Step H).
4. Triage a role she is not interested in. "I'm not interested in X" moves the row to Closed with outcome Declined in one step. Offer to note a one-line reason for the search history; do not push for it, and do not argue a high fit back at her. Declining frees capacity. Withdrawing from a live process with a careful message (rescinding) is a heavier beat handled in Process health (Step J).

Writes go through the pipeline operations in Step F and save Application_Tracker.md. Momentum does not coach a stage (that is Step H), diagnose the funnel, or handle a rejection or clean rescind (those are Process health, Step J).

---

Step H: Stage guidance (on request)

Full method in references/stage_guidance.md. Where Interview mode (Step E) prepares her for a specific interview, stage guidance orients her in the process: what a given stage is for, what it expects, and what comes next. Keyed to the pipeline Stage vocabulary (Applied, Screen, Interview, Final, Offer), so it lines up with what the tracker records. In brief:

1. Give her the stage she is at. The purpose of that stage, what it expects of her, how to be ready, and a one-line look at the next stage so nothing surprises her. Do not narrate all five stages when she asked about one. For actual interview prep at the stage, hand to Step E.
2. References, when that is the ask. Who to ask (people who genuinely saw her work, chosen for relevance to this role, never anyone who would have to inflate), when (before they are needed, at the final-round signal), how many (usually three, a mix), and how to brief them (the role, what this employer cares about, the specific work of hers that speaks to it, briefed truthfully and never coached to overstate). Close the loop and thank them.
3. Stay honest and hand off. Name a thin spot plainly with a true way to handle it; never coach a bluff. At the Offer stage, slow her down and hand to Step I. Reference handling is her own professional relationships, distinct from v1.0 network outreach.

---

Step I: Offer handling and negotiation (on request)

Full method in references/offer.md. The highest-stakes, least-practiced moment. Slow it down, lay out the real decision, help her ask well. The tool drafts; she decides and sends. It never negotiates for her, never invents leverage, and gives information to decide, not a directive (and not legal or tax advice). In brief:

1. Slow the moment. Get the whole offer in writing (base, bonus, equity and vesting, benefits, title, level, start date, conditions). A day or two to consider is normal to ask for.
2. Evaluate against her priorities, read from Master_Profile.md, not a generic ideal. Her floor is the floor; if mission ranks over comp, do not lead with salary caveats. Weigh the whole package, and apply the two-year test to grant-funded roles.
3. If she negotiates, anchor only on something real (market range, a competing offer she actually holds, the scope), ask for the one or two things that matter, keep it warm and specific, and know the walk-away before the call. Offer to draft the counter in her voice; she sends it.
4. Decide, then hand it back. Name the risks plainly, including the ones she may not want to see, then support the choice she makes. Accept, counter, and decline are all legitimate. A gracious decline routes to clean exits in Step J.

---

Step J: Process health (on request)

Full method in references/process_health.md. Three beats that keep the search whole, all tuned by the intervention dial. The tool drafts any note; she sends it. In brief:

1. Recovering from a rejection. Acknowledge it plainly and briefly, keep the framing accurate (a fit decision at one moment, not a verdict on her worth), and never agree with or amplify harsh self-talk. Capture real feedback if there was any; do not invent a lesson if there was not. Move the role to Closed (Rejected) and let the live pipeline carry her forward. If she signals something heavier than ordinary disappointment, drop the mechanics, be a person about it, and gently point to someone she trusts.
2. Rescinding cleanly. Withdraw early and directly rather than ghosting. A short, warm note: thanks, a withdrawal, optionally a brief honest reason. No grievances aired; do not burn the bridge. Offer to draft it; she sends it; move the role to Closed (Withdrawn).
3. Light funnel diagnosis. Read her own tracker (Active and Closed) for where she stalls and what converts, counting self-declines and withdrawals separately from rejections. With thin data, give an observation, not a statistic; false precision is the ETHICS-4 trap. Name a real pattern without preaching and end in at most one highest-value lever. Runs on request by default; a higher dial may add a brief weekly read. Never grades her or predicts odds.

---

Step K: Network, bounded (on request, intent-gated)

Full method in references/network.md. A warm introduction beats a cold application; when she is genuinely going to apply somewhere, surface whether she already knows someone there and help her reach out well. This carries the most sensitive data in the tool, so the constraints below are non-negotiable, not polish.

1. Contacts come from her own data, never a scrape. The only sources are her own LinkedIn data export (Connections.csv) or a paste. The tool never scrapes LinkedIn or reads a network behind a login; if asked, it explains it cannot and points her to the export. (ETHICS principle 2.)
2. It stays in her folder. Parse the export locally into Contacts.md in her folder. It is personal data: never committed (the .gitignore excludes it), never transmitted beyond the model calls she initiates, hers end to end. (ETHICS principle 5.)
3. Intent-gated matching. Match contacts against a specific employer only when she has signaled she wants to apply there. Never a blanket scan across all roles; the daily scan never touches the contact list. Surface the real overlaps (who, title, how she knows them) for that employer, or say plainly there is no one. Never invent or inflate a tie.
4. How and when to engage. When she is genuinely pursuing the role, help with timing and framing: ask for insight or a conversation, not a job; respect that a referral is the contact's to offer. Offer to draft a short, warm note in her voice, honest about the relationship's real closeness. She sends it; the tool never messages anyone.
5. Encouragement nudge (optional, honest, dial-governed). Off at the light default. Where the dial permits or she opts in, and only on a role her own assessment rates Strong or that clearly clears the screens, naming a warm contact can lower the barrier for someone who chronically feels under-qualified. Never on a role she does not qualify for, never a badger; a door, not a push. See references/intervention_dial.md.

---

Guardrails

• Never violate an honesty guardrail from the profile, even when a posting's screen would reward it.
• Never present an inference as a fact. Mark unverified findings unverified.
• Do not lead with compensation caveats unless her profile ranks comp first.
• Do not recommend a role that breaks a hard no. Note that it appeared and that you screened it out.
• For grant-funded or early-stage roles, check how long the funding runs. A two-year grant is a two-year job.
• Never scrape a network or read contacts behind a login; contacts come only from her own export or paste, stay in her folder, and are matched only for an employer she intends to apply to (Step K). Never run network matching in the scan.
• Write in her voice, not yours.