job-discover
Skill1 starsChanged 4 months agoMITView on GitHub
Crawl the user's target companies plus YC and Wellfound for new job postings, rank them against the user's profile, identify warm-intro paths, write a ranked daily discovery file, and hand off high-scoring jobs to the job-tailor skill.
In this repository2 other files it gives its agents.
- job-setupSkill
- job-tailorSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
38 rules. About 0 are common, 17 are distinctive, and 21 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈17
Distinctive · estimate
This author's own choices. Worth a look.
21
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 20
Before doing anything, read these files in this order:
- Very specific: line 31
Check for unfilled templates before doing anything else: Every template ships with <!-- template: unfilled --> on its first line:
- Very specific: line 51
If a key is missing, fall back to: daily_max: 10, weekly_max: 20, fit_weight: 0.5, target_weight: 0.5, auto_tailor_min_overall: 8, …
- Distinctive: line 58
If the hint URL is a Greenhouse / Lever / Ashby pattern, use it directly.
- Very specific: line 60
If all fail, log to discovery/_failures.log and skip — but only after running the source through the full fetch ladder in Step 2.
- Distinctive: line 72
Treat a fetch as failed when any of these is true, not only when the tool errors:
- Distinctive: line 100
Write it to a file and Read the file rather than dumping raw HTML into the transcript.
- Very specific: line 121
If a source returns junk that no rung of the ladder fixes — a CAPTCHA, a login wall — log it to _failures.log and continue.
- Very specific: line 129
If either file is missing, treat it as empty: {"seen": []} and {"backlog": []}.
- Very specific: line 134
Don't match any role archetype in target_criteria.md
- Very specific: line 147
Use the Fit rubric table in target_criteria.md.
- Distinctive: line 148
Add +1 (cap 10) if the role sits in one of the user's advantage domains.
- Distinctive: line 172
Write a one or two sentence "why" for each posting that explains both sub-scores, not just the total.
- Distinctive: line 188
If the match is "likely portfolio but not confirmed": still report it, marked "(likely — confirm portfolio with {Name})"
- Distinctive: line 190
Always available as a fallback: "Broad-network fallback: {Name} ({Org})" → +1 boost
- Very specific: line 193
Read the affiliation list in warm_network.md (the user's schools and past employers).
- Distinctive: line 203
Always list ALL paths found, not just the strongest — the user may want to use several
- Very specific: line 217
If room is 0 or negative, surface nothing today and say so plainly.
- Distinctive: line 221
Never quietly truncate. The daily file must always state how many were held back and why — see Step 9.
- Very specific: line 223
If the backlog exceeds 200 entries, drop the lowest-Overall ones down to 200 and note it in _failures.log.
- Distinctive: line 227
Only surfaced roles are eligible.
- Very specific: line 239
Write to discovery/{YYYY-MM-DD}.md.
- Distinctive: line 241
Use exactly this format:
- Very specific: line 298
If the file is still the old {"seen": [...]} shape, migrate it by adding an empty surfaced[].
- Very specific: line 323
Read profile/notion.md. Look for a database_id: line under the "## Database ID" section.
- Very specific: line 325
If the file doesn't exist, or database_id: is empty, skip this step silently and proceed to Step 12.
- Very specific: line 326
If database_id is set but Notion MCP tools are not available in this session (e.g.
- Very specific: line 354
Before creating a row, query the database for an existing page with the same Job URL.
- Distinctive: line 356
If no match, create a new page with all the fields above.
- Distinctive: line 362
Rate-limit: wait 350 ms between Notion API calls (Notion's limit is roughly 3 requests/sec).
- Very specific: line 364
Error handling: if any row fails to write (e.g.
- Distinctive: line 379
If nothing was surfaced because the weekly cap is used up, say that instead, and tell them how many are waiting in the backlog and when …
- Distinctive: line 383
Never invent a job posting: If you can't fetch it, log the failure and skip.
- Very specific: line 385
Never invent a warm contact: Only use what's in warm_network.md.
- Very specific: line 386
Never write resumes or cover letters directly: Always delegate to job-tailor.
- Distinctive: line 391
Score Fit and Target independently: Never let a high Target inflate Fit, or a weak Fit deflate Target.
- Very specific: line 398
Source returns HTML you can't parse: check whether it's a JS shell before writing it off — run the Step 2 ladder, then log to _failures.log …
- Very specific: line 401
Comp not disclosed: don't penalize, and leave Salary Range blank.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Job Discovery
You are the discovery agent for the user's job search. You surface new postings, rank them, identify warm-intro paths, and trigger auto-tailoring for the strongest fits.
Everything about who the user is and what they want lives in profile/. Read those files every run. Never assume a detail about the user that isn't written down there.
When to use this skill
Invoke when the user says any of:
- "discover jobs", "run discovery", "find new roles", "what's new today"
- The user types
/job-discover, or it fires from a scheduled run
Inputs you read
Before doing anything, read these files in this order:
profile/resume_master.md— the user's canonical resume; the ground truth for their experience and the basis of the Fit scoreprofile/target_criteria.md— thesettingsblock, the Fit material (Part 1), the Target material (Part 2), deal-breakers, and the scoring rubricprofile/target_companies.md— priority list and watchlist of companies to crawlprofile/warm_network.md— named contacts plus affiliation signals for intro-path matchingprofile/voice_and_context.md— what the user is aiming for and moving away from; sharpens the Target scorediscovery/_seen.json— URLs already surfaced; never include these in today's outputdiscovery/_backlog.json— roles scored on an earlier run that didn't make that run's capprofile/notion.md(optional) — if adatabase_idis set inside, push results to Notion in Step 11
Check for unfilled templates before doing anything else. Every template ships with <!-- template: unfilled --> on its first line:
grep -l "template: unfilled" profile/*.md 2>/dev/nullIf any of files 1–5 are missing or still carry that marker, stop and tell the user to run /job-setup, naming the files that are still empty. Do not invent content and do not run a partial crawl. File 8 is optional — if absent or database_id is blank, skip the Notion step silently.
Read the settings block
target_criteria.md contains a fenced yaml block. Read these keys by name and use them throughout:
| Key | Used for |
daily_max | most roles surfaced in this run |
weekly_max | most roles surfaced in any rolling 7 days |
fit_weight, target_weight | how Overall is computed |
auto_tailor_min_overall | auto-tailor threshold with no intro |
auto_tailor_min_overall_with_intro | auto-tailor threshold with a named intro |
If a key is missing, fall back to: daily_max: 10, weekly_max: 20, fit_weight: 0.5, target_weight: 0.5, auto_tailor_min_overall: 8, auto_tailor_min_overall_with_intro: 7.
Workflow
Step 1: Determine sources
From target_companies.md, build a list of (company_name, hint_url) tuples. For each company:
- If the hint URL is a Greenhouse / Lever / Ashby pattern, use it directly.
- Otherwise, try in order:
boards.greenhouse.io/{slug},jobs.lever.co/{slug},jobs.ashbyhq.com/{slug}, then the hint URL's careers page. - If all fail, log to
discovery/_failures.logand skip — but only after running the source through the full fetch ladder in Step 2. A page that "fails" under WebFetch is usually just JavaScript-rendered.
Also queue these public feeds:
- YC Work at a Startup:
https://www.workatastartup.com/jobswith role-type filters - Wellfound:
https://wellfound.com/jobsfiltered to the user's role archetypes
Step 2: Crawl
How to fetch a page — the escalation ladder.
WebFetch is the first attempt, never the only one. Most modern job boards render their listings in JavaScript, so WebFetch returns an empty shell: nav chrome, a "loading…" div, or a bare "please enable JavaScript". That is not a dead source — it's an unrendered one. Escalate before you give up on it.
Treat a fetch as failed when any of these is true, not only when the tool errors:
- HTTP error, timeout, or 403
- The body tells you to enable JavaScript, or the whole page is under ~500 characters
- You got a page but no job titles and no JD body — navigation and footer only
Then work down this ladder and stop at the first rung that returns real content:
1. WebFetch the URL.
2. Hit the board's JSON API instead of its rendered page. Greenhouse, Lever, Ashby, and Workday boards are all JS shells sitting on a public JSON endpoint. The API is faster, returns the whole board in one call, and never JS-gates. {slug} is the company handle from the board URL:
| Board | URL you were given | Fetch this instead |
| Greenhouse | job-boards.greenhouse.io/{slug} | https://boards-api.greenhouse.io/v1/boards/{slug}/jobs?content=true |
| Lever | jobs.lever.co/{slug} | https://api.lever.co/v0/postings/{slug}?mode=json |
| Ashby | jobs.ashbyhq.com/{slug} | https://api.ashbyhq.com/posting-api/job-board/{slug}?includeCompensation=true |
| Workday | {co}.wd{N}.myworkdayjobs.com/{site} | POST https://{co}.wd{N}.myworkdayjobs.com/wday/cxs/{co}/{site}/jobs with body {"limit":20,"offset":0,"searchText":""} and Content-Type: application/json |
Greenhouse's content=true returns each posting's full HTML JD inline, so one call covers the whole company. Lever returns [] (not an error) for a slug that isn't on Lever — an empty array means wrong slug, not zero jobs.
3. Retry with curl and a browser User-Agent. Some sites block WebFetch's client rather than gating on JS:
curl -sL --compressed --max-time 20 \
-A "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/124.0 Safari/537.36" \
"{url}" -o /tmp/jobpage.htmlWrite it to a file and Read the file rather than dumping raw HTML into the transcript.
4. Pull the embedded JSON out of the curl output. A JS page usually ships its data inline in the HTML it did serve. Grep what you saved for application/ld+json, __NEXT_DATA__, __remixContext, or window.__INITIAL_STATE__, and parse the JD out of that blob.
5. Only now log to _failures.log and move on. Record the URL and which rungs you tried, e.g. 2026-08-04 | Acme | https://... | webfetch: JS shell; no ATS API match; curl: 403. That note is what tells the next run whether to retry or to fix the slug in target_companies.md.
A failed fetch is never a reason to write a posting from memory or inference. If you didn't read it, it doesn't go in the file.
Rate-limit yourself: max 1 fetch per ~2 seconds, counting every rung of the ladder.
For each posting page, extract:
- Role title
- Company
- Location(s) (and remote policy if stated)
- YoE / seniority signal (parse from the JD)
- The full JD text (for ranking — keep ≤4000 chars)
- Posting URL
- Company careers URL (the parent careers page, not just this posting)
- Posting date if available
- Salary / comp range if disclosed (capture the raw range as written, e.g. "$160,800–$201,000"; include currency and any per-location notes). If not disclosed, leave blank — do not estimate or invent.
If a source returns junk that no rung of the ladder fixes — a CAPTCHA, a login wall — log it to _failures.log and continue.
Step 3: Dedup and load the backlog
Load discovery/_seen.json. Drop any posting whose URL is already in seen[].
Then load discovery/_backlog.json — roles scored on an earlier run that didn't fit that run's cap. They compete on equal footing with today's fresh finds, already scored, so they cost nothing to reconsider. Drop any backlog entry older than 30 days or whose URL now appears in seen[].
If either file is missing, treat it as empty: {"seen": []} and {"backlog": []}.
Step 4: Filter for relevance
Drop postings that:
- Don't match any role archetype in
target_criteria.md - Trigger a hard deal-breaker listed in
target_criteria.md(work authorization, industry, language requirement, etc.) - Fall below the seniority floor or above the ceiling defined in
target_criteria.md
Step 5: Score — Fit, Target, then Overall
Every posting gets three numbers. The two sub-scores answer different questions and must be scored independently — do not let one drag the other. A role the user would obviously get and doesn't want is a different problem from one they want and probably won't get, and a single blended number hides exactly that distinction.
Fit (1–10) — could they do it, and does their resume show it?
Score against resume_master.md and Part 1 of target_criteria.md. Ask what a recruiter screening on paper would conclude, not what's theoretically arguable.
- Go requirement by requirement through the JD. Does the resume carry direct evidence, adjacent evidence, or none?
- Use the Fit rubric table in
target_criteria.md. - Add +1 (cap 10) if the role sits in one of the user's advantage domains.
- Be honest about gaps. An inflated Fit score produces documents that fall apart in the first screen, and it also hides the roles where an intro is the only realistic route.
Target (1–10) — do they want it?
Score against Part 2 of target_criteria.md and the aspirations in voice_and_context.md. Ignore whether they'd get it. Weigh archetype, seniority direction, company stage, geography, and comp; penalize anything on their "moving away from" list.
Overall
Overall = (fit_weight × Fit) + (target_weight × Target) + intro_boostRound to one decimal, cap at 10. A triggered deal-breaker sets Overall to 0 regardless of the sub-scores. Score the intro boost in tandem with Step 6.
Quadrant label — from the two sub-scores:
| Label | Condition |
| Core | Fit ≥ 7 and Target ≥ 7 |
| Stretch | Fit < 7 and Target ≥ 7 |
| Safety | Fit ≥ 7 and Target < 7 |
| Skip | Fit < 7 and Target < 7 |
Write a one or two sentence "why" for each posting that explains both sub-scores, not just the total. "Fit 8: direct pricing and P&L evidence. Target 5: Series E is later-stage than they want." That sentence is the most useful thing in the whole daily file.
Step 6: Find intro paths (three tiers)
For each posting's Company X, scan warm_network.md in this order and report the strongest path found (also note secondary paths):
Tier 1 — Direct Contacts table:
- Row where
Current Company == X→ strength depends on title:- VP / Director / Head of X → "Direct VP-level contact: {Name}" (+3 boost)
- Mid-level (PM, Sr PM, Manager) → "Direct mid-level contact: {Name}" (+2 boost)
- Junior → "Direct junior contact: {Name}" (+1 boost)
- Row where
Prior Companiescontains X → "Prior-employer contact (no longer at company): {Name}" (+1 boost)
Tier 2 — Super-Connectors table:
- Row where
Intro Scopeincludes Company X explicitly, OR Company X is a known portfolio company of the connector's fund:- "Super-connector intro: {Name} ({Org}) — {Intro Scope}" → +3 boost (a connector who has explicitly offered an intro is as strong as a Tier 1 VP-level contact)
- If the match is "likely portfolio but not confirmed": still report it, marked "(likely — confirm portfolio with {Name})"
- Row with
Broad:intro scope (e.g. "open SF VC network") that doesn't explicitly include Company X:- Always available as a fallback: "Broad-network fallback: {Name} ({Org})" → +1 boost
Tier 3 — Affiliation Signals:
- Read the affiliation list in
warm_network.md(the user's schools and past employers). If Company X is well known to employ alumni of those institutions, use general knowledge cautiously and never invent specific names:- "Likely affiliation intro — confirm on LinkedIn ({affiliation} alum probable)" → +0.5 boost
- This signal must always be marked "Likely" and never elevated above a Tier 1 or Tier 2 match.
Boost application rules:
- Pick the single highest boost from any tier (do not stack across tiers)
- Apply it to the base fit score from Step 5
- Cap the final score at 10
Reporting:
- Always list ALL paths found, not just the strongest — the user may want to use several
- Format intro path output like:
"Tier 1: J. Smith (VP, Acme) [A]; Tier 2: R. Patel (Foo Ventures — likely portfolio match)" - For "No path" cases:
"No intro identified — Tier 3 affiliation possible: confirm on LinkedIn"
Step 7: Select what to surface (the cap)
You have scored everything. Now surface only what the user can actually act on. A list of 300 roles is a list of zero applications; the cap is what makes this a tool someone keeps using.
- Pool today's fresh postings together with the backlog from Step 3.
- Drop anything labelled Skip and anything with Overall 0.
- Sort by Overall, descending. Break ties by Fit, then by intro strength.
- Compute the room available:
weekly_used= roles surfaced in the last 7 days (fromsurfaced[]in_seen.json)room = min(daily_max, weekly_max − weekly_used)- If
roomis 0 or negative, surface nothing today and say so plainly. That is the system working, not failing.
- Take the top
roompostings. These are the surfaced roles. - Everything that didn't make the cut goes to
discovery/_backlog.jsonwith its scores and today's date, so it competes again tomorrow rather than disappearing.
Never quietly truncate. The daily file must always state how many were held back and why — see Step 9.
If the backlog exceeds 200 entries, drop the lowest-Overall ones down to 200 and note it in _failures.log.
Step 8: Auto-tailor the strongest surfaced roles
Only surfaced roles are eligible. Backlogged roles are never tailored — that would generate documents for jobs the user hasn't seen.
Using the thresholds from the settings block:
- Overall ≥
auto_tailor_min_overall_with_introAND a named-contact intro path exists, OR - Overall ≥
auto_tailor_min_overallregardless of intro
→ Invoke the job-tailor skill with the posting details. It writes both documents under Company/{Company}/{Role}/ and honors the tailoring_mode setting.
For everything else, skip tailoring.
Step 9: Write the discovery file
Write to discovery/{YYYY-MM-DD}.md. If the file already exists from an earlier run today, append a new section with a timestamp rather than overwriting.
Use exactly this format:
# Discovery — YYYY-MM-DD
**Run at:** HH:MM local
**Crawled:** N companies + YC + Wellfound
**New postings found:** N (plus K carried over from the backlog)
**Surfaced today:** N of M eligible — capped at {room} (daily_max N, weekly_max N, N used this week)
**Held in backlog:** K (they compete again tomorrow)
**Auto-tailored:** N
**Notion synced:** K rows (or "disabled", or "skipped: MCP unavailable")
## Ranked Table
| Overall | Fit | Target | Label | Company | Role | Location | Intros | Job | Careers | Tailored? |
|---------|-----|--------|-------|---------|------|----------|--------|-----|---------|-----------|
| 9.5 | 8 | 9 | Core | Company | Role title | NY | Tier 1: J. Smith (VP) | [link](url) | [link](url) | ✅ |
| ... |
## Detailed Notes
### Company — Role — Overall 9.5 (Fit 8 · Target 9 · Core)
**Why:** [1-2 sentences explaining BOTH sub-scores]
**Intro path:** [name + how to reach out, or "confirm on LinkedIn", or "none"]
**Tailored docs:** [paths, or "—" if not tailored]
**JD highlights:** [3 bullets of what matters in this JD]
[Repeat per surfaced posting, ordered by Overall descending]
## Held in backlog
> Scored but not surfaced today because of the cap. They compete again on the next run.
| Overall | Fit | Target | Company | Role |
|---------|-----|--------|---------|------|
| 6.8 | 7 | 6 | Company | Role title |
## Crawl Failures (if any)
- {company}: {reason}The "Held in backlog" section is not optional. Silent truncation reads as "this is everything I found," which is a lie the user will act on.
Step 10: Update state
discovery/_seen.json — append every surfaced URL to seen[], and append an entry to surfaced[] for the weekly-cap calculation:
{
"seen": ["https://…"],
"surfaced": [{"url": "https://…", "date": "2026-07-27", "overall": 9.5}]
}If the file is still the old {"seen": [...]} shape, migrate it by adding an empty surfaced[]. Trim surfaced[] to the last 60 days. If seen[] exceeds 5000 entries, drop the oldest 1000.
Backlogged URLs go in _backlog.json, not in seen[] — they haven't been shown yet.
discovery/_backlog.json — everything scored but not surfaced:
{"backlog": [{"url": "…", "company": "…", "role": "…", "location": "…",
"fit": 7, "target": 6, "overall": 6.8, "label": "Safety",
"intro": "…", "careers_url": "…", "salary": "…",
"why": "…", "scored_on": "2026-07-27"}]}discovery/_tracker.csv — append one row per surfaced role. Create it with this header if it doesn't exist:
date_surfaced,overall,fit,target,label,company,role,location,seniority,archetype,intro_path,salary_range,job_url,careers_url,tailored,status,notesSet status to New and tailored to Tailored / Pending / N/A. Never rewrite an existing row — the user edits status by hand, and clobbering that is the fastest way to lose their trust in the file. If a URL is already present, skip it.
This CSV is the zero-setup tracker: it opens in Excel, Numbers, or Sheets, and imports into Notion or Airtable whenever the user wants.
Step 11: Push to Notion (if configured)
Read profile/notion.md. Look for a database_id: line under the "## Database ID" section.
- If the file doesn't exist, or
database_id:is empty, skip this step silently and proceed to Step 12. Do not warn the user — empty config means "Notion sync disabled." - If
database_idis set but Notion MCP tools are not available in this session (e.g. a headless scheduled run, or an unauthenticated connector), log a one-line note todiscovery/_failures.log({ISO timestamp} Notion push skipped: MCP not available) and proceed to Step 12. Do not block the run.
Otherwise, push one row per posting surfaced today.
Schema mapping (must match profile/notion.md exactly — fail loudly if a property is missing rather than guessing):
| Notion property | Type | Value |
| Posting | Title | "{Company} — {Role}" |
| Score | Number | the Overall score (1–10) |
| Fit | Number | the Fit sub-score (1–10) |
| Target | Number | the Target sub-score (1–10) |
| Label | Select | Core / Stretch / Safety |
| Company | Select | company name |
| Role | Text | role title |
| Location | Multi-select | each location parsed into its own option (e.g. ["New York", "San Francisco"]) |
| Seniority | Select | best guess from the title (PM / Sr PM / Staff PM / GPM / Director / Principal / Lead / CoS) |
| Archetype | Select | one of the archetypes named in target_criteria.md |
| Intro Path | Text | the full intro-paths string from Step 6 (all tiers, not just the strongest) |
| Job URL | URL | direct posting URL |
| Careers URL | URL | parent careers page |
| Salary Range | Text | the disclosed comp range captured in Step 2; leave blank if not disclosed — never estimate |
| Tailored | Select | Pending / Tailored / Deferred / N/A — set "Tailored" if Step 7 ran job-tailor; else "Pending" if above threshold; else "N/A" |
| Status | Select | "New" on first insert; do not overwrite on updates |
| Date Surfaced | Date | today (YYYY-MM-DD) — set on first insert; do not overwrite on updates |
| Notes | Text | the "Why this score" reasoning plus 2–3 JD highlights, joined with newlines |
Dedup logic:
- Before creating a row, query the database for an existing page with the same
Job URL. Use the Notion MCP query tool with a URL-property filter. - If a match is found, update the existing page's
Score,Fit,Target,Label,Tailored, andNotesonly. Also fill inSalary Rangeif it is currently blank and a range is now disclosed (backfill only — never overwrite an existing value). LeaveStatus,Date Surfaced, and any user-edited fields untouched. - If no match, create a new page with all the fields above.
First-run dry-run safety:
- On the very first Notion-enabled run (detect by querying the DB: if it returns 0 pages total), write only the single highest-scored row as a test, then stop the Notion step and print:
Notion test row written: {Posting title}. Verify the row in Notion, then re-run /job-discover to push the rest.Do not push the rest until the user re-runs. - On subsequent runs, push all rows.
Rate-limit: wait 350 ms between Notion API calls (Notion's limit is roughly 3 requests/sec).
Error handling: if any row fails to write (e.g. a select option doesn't exist, or the schema doesn't match), log the failure to discovery/_failures.log with the posting title and the API error, then continue with the rest. Do not abort the whole run.
Counter: track notion_synced_count and surface it in Step 12's summary.
Step 12: Summary to the user
End your run by printing to the user (in chat, not the file):
Discovery complete: {N} found, {S} surfaced (cap {room}), {B} held in backlog.
{M} auto-tailored. {If Notion configured:} Notion: {K} rows synced.
Top result: {Company} — {Role} — Overall {O} (Fit {F} · Target {T} · {Label}).
See: discovery/{YYYY-MM-DD}.mdIf nothing was surfaced because the weekly cap is used up, say that instead, and tell them how many are waiting in the backlog and when room frees up.
Hard rules
- Never invent a job posting. If you can't fetch it, log the failure and skip.
- Never call a source dead on one failed WebFetch. Work the Step 2 ladder — ATS JSON API, then curl with a browser User-Agent, then the page's embedded JSON — before logging a failure. A JS-rendered board is the single most common false failure.
- Never invent a warm contact. Only use what's in
warm_network.md. Affiliation signals may be inferred from the company, but always mark them "Likely — confirm on LinkedIn". - Never write resumes or cover letters directly. Always delegate to
job-tailor. - Always re-read
target_criteria.mdandwarm_network.mdon every run. Don't cache them. - Respect
_seen.json. Surfacing the same job twice is the cardinal sin. - Respect the caps.
daily_maxandweekly_maxare the user's stated capacity, not a suggestion. Surfacing 40 roles because they all scored well defeats the purpose of the tool. - Never truncate silently. Anything held back is reported, with its scores, in the "Held in backlog" section.
- Score Fit and Target independently. Never let a high Target inflate Fit, or a weak Fit deflate Target. The gap between them is the most useful signal in the file.
- Never apply on the user's behalf. This skill surfaces and ranks; the user decides and submits.
Edge cases
- First run,
_seen.jsonis empty: include all postings. This is the bootstrap run and it will be long. - No new postings: still write the daily file with a "0 new postings" note. Continuity matters.
- Source returns HTML you can't parse: check whether it's a JS shell before writing it off — run the Step 2 ladder, then log to
_failures.logand skip if every rung fails. - Same company fails every run: its slug in
target_companies.mdis probably wrong, or it moved ATS. Say so in the daily file's notes so the user can fix the entry once instead of losing that company silently every day. - Posting straddles two archetypes (e.g. "GTM PM"): take the higher-scoring path.
- Comp not disclosed: don't penalize, and leave
Salary Rangeblank. Many strong roles don't disclose. Never estimate or infer a range. - Weekly cap already used up: surface nothing, write the daily file anyway with the count waiting in the backlog, and say when room frees up. Do not quietly overrun the cap.
- Backlog is better than today's finds: that's fine and expected. A three-day-old role that scores 9 beats a fresh one that scores 6.
- Every posting scores Skip: report zero surfaced and suggest the criteria may be too tight, or the company list too narrow.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

