agentleFS
CodesWhatCodesWhat/careerrat/.agents/skills/company-health/SKILL.md

company-health

Skill4 starsChanged 14 days agoMITView on GitHub

Role-scoped company-health/sentiment signal — web-search layoff risk, hiring momentum, financials, sentiment, and leadership stability for the candidate's target function, score a healthy|watch|risky rating with provenance, persist it to the tracker, and cross-cut the fit score only where it hits a stated candidate need.

The file

We split this skill into the rules it gives an agent

12 rules. About 0 are common, 5 are distinctive, and 7 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 7

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Use this skill when the user asks how risky, stable, or healthy a company is — layoffs, financials, morale, "is this a safe place to land" …

    line 17
  2. Very specific:

    Mode detection: run careerrat data status.

    line 48
  3. Distinctive:

    Read before doing anything else.

    line 53
  4. Very specific:

    Default interview keeps the set tiny (only companies you actually reached).

    line 88
  5. Very specific:

    If the gate says "don't fire," say so briefly (e.g.

    line 96
  6. Very specific:

    Run focused web searches (public sources only — no login-walled scraping), capped at max ~2 searches per dimension.

    line 114
  7. Very specific:

    No intersection → fitDelta: 0, crossCut: [].

    line 165
  8. Very specific:

    DB workspace: write it through the dedicated verb, not set-fields/upsert-batch — this is the required write path, and the only one that …

    line 198
  9. Distinctive:

    Verify + snapshot (the dashboard handoff):

    line 209
  10. Distinctive:

    If a dimension produced nothing credible:

    line 298
  11. Distinctive:

    Internal signal ONLY: Never leaks into an outbound artifact — no defamation-y phrasing in anything a human at the company could read.

    line 308
  12. Very specific:

    No hardcoded role families, companies, or domain terms in this skill's behavior.

    line 322

≈ “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.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

company-health

Use this skill when the user asks how risky, stable, or healthy a company is — layoffs, financials, morale, "is this a safe place to land" — or when they ask to factor company risk into a role. It also runs on its own when a tracked company reaches the configured firing stage (default: the interview band) with no fresh rating; STEP 1 is that decision.

The question is never "is this a good company." It is "is my function a risky place to land at this company right now." Role-scoped, time-stamped, judgment surface — a healthy company that just gutted the function you'd join is a red flag; a wobbly company doubling down on it may still be worth a shot.

Runs under AGENTS.md. These contracts bind without being restated here: Honesty Firewall, Privacy Invariant (current_base never outbound), Placeholder/Bracket Ban, Domain-Neutral Rule, Browser Automation Contract, Activity Pulse logging, and the Tracker Write Contract (stamp → verify → snapshot → activity). Inline reminders at point-of-use are intentional. One extra hard rail for this skill: the rating is an INTERNAL signal only — it must never appear in any outbound artifact (cover letter, recruiter reply, LinkedIn). No "your company is in freefall" energy reaches a human at the company. Bare candidate/ and workspace/ paths below are symbolic; resolve them per AGENTS.md's Path Resolution rule.

Agent voice. Read candidate/modes.yml#agent_voice (default standard) before presenting the rating. Apply the register from AGENTS.md#mode-switches. The persisted companyHealth object is always written in full — register governs the in-chat summary: exec-summary = the rating + the one signal that drove it; standard = rating + per-dimension one-liners + as-of; technical = signal analysis + sourcing; verbose = the full rationale + evidence list.


STEP 0 — Load config

Mode detection: run careerrat data status. Exit 0 → DB workspace — every write step below gives the careerrat data <verb> command (Data Write Contract, AGENTS.md). Nonzero exit → legacy workspace (no DB yet) — every write step below gives the existing direct JSON-edit instructions, unchanged.

Read before doing anything else. The skill must work with all of these absent (neutral defaults shown).

FileFields usedDefault if absent
candidate/modes.ymlcompany_health.{auto, fire_at_stage, manual_always, recheck_days}auto:true, fire_at_stage:interview, manual_always:true, recheck_days:14
candidate/targeting.ymlrole_buckets / role families (the target function), priorities or must_haves (the candidate's stated needs — drive the cross-cut)function = the role on the tracker row; no needs = no cross-cut
candidate/profile.ymlcandidate.domain, candidate.location (query context only) — NEVER read compensation.current_basederive domain from the role

Derive target_function from the role family on the tracker row (or targeting.role_buckets) — e.g. "AI/ML engineering", "supply-chain operations", "clinical research". Substitute it wherever a dimension below references <function>. Never hardcode a domain term.

Derive candidate_needs from targeting.priorities/must_haves (e.g. stability, growth, culture, comp). These — and only these — decide whether the rating touches the fit score (STEP 4). No declared needs ⇒ the rating stays a standalone signal.


STEP 1 — Firing decision (cost gate)

Deep web research is expensive; do not run it on every applied company. Decide whether to fire before spending a single search:

  1. Manual / explicit ask → always run (when manual_always: true, the default). The user asked; honor it regardless of stage.
  2. Automatic → run only when both:
    • company_health.auto is true, and
    • the tracked company has reached fire_at_stage. Map the coarse stage to the tracker status band: applied = any live row; interview = the row has entered the interview band (a real conversation / screen is booked or done — conversations[] non-empty or status at the screen band or deeper: screen/interview/onsite/panel/offer — a recruiter/phone screen counts, since classifyStage ranks screen at the interview threshold); offer = an offer is in hand. Default interview keeps the set tiny (only companies you actually reached). The app checks this on every status change and starts the run in the job's thread on its own, once per job and stage. If no AI is set up it leaves a short note instead.
  3. Freshness / decay → if a companyHealth object already exists on the row and its asOf is newer than recheck_days ago, reuse it (do not re-research) unless the user forces a refresh or the row just crossed into the offer band. Older than recheck_days, or a major stage transition → re-research and bump asOf.

If the gate says "don't fire," say so briefly (e.g. "Skipping — auto-research fires at the interview stage; run company-health <Company> to force it now.") and stop. Never burn searches on a ghosted/early-funnel company by default.


STEP 2 — Resolve company + function

  • Resolve the canonical company name and whether it is public or private (changes the financial dimension's sources).
  • Confirm the target_function from STEP 0. The whole rating hangs on it — a layoff that hit Sales is near-irrelevant to an Engineering candidate and decisive to a Sales one.
  • Note as_of = today (the rating is stamped and decays from here).

STEP 3 — Role-scoped research (the dimensions)

Run focused web searches (public sources only — no login-walled scraping), capped at max ~2 searches per dimension. Weight every read toward <function>. Skip a dimension cleanly if nothing credible surfaces (mark it needs-more-info, never fabricate).

DimensionWhat to findRole-scoping
Layoff risklayoffs.fyi, WARN notices, RIF news (last 12–18 mo)Did it hit <function>? Function-hit dominates the whole rating.
Hiring momentumopen-req count + trend for <function>, freezes, rescinded-offer chatterGrowing the team you'd join, or a backfill into a shrinking org?
Financial healthpublic: earnings, revenue trend, profitability, guidance, stock; private: last raise date + size vs headcount, down-/flat-round, stage-vs-headcountRunway risk = staleness for private cos.
SentimentGlassdoor / Blind / news on morale + public perception ("how the world feels")Prefer trend over snapshot + business-outlook / CEO approval over the star rating. Tempered — noisy/biased, down-weight.
Leadership stabilityexec churn in the line you'd report into, reorg frequencyIs the function's leadership a revolving door?
Acute catalysts (if present)lawsuit, breach, big customer loss, acquisition rumor, product flopA live catalyst should dominate the signal.
Is <function> strategic or a cost-center (synthesis)from the above + business modelLoad-bearing function survives downturns; a side-bet gets cut first.

Record each finding with its source + date for the evidence trail (STEP 5).


STEP 4 — Score (standalone rating + selective cross-cut)

4a — Per-dimension level. Rate each researched dimension good | mixed | poor (or needs-more-info) with a one-line note. functionHit:true on layoffs is a hard poor.

4b — Overall role-scoped rating. Roll up to one of:

  • healthy — no function-scoped risk signals; financials/hiring/sentiment hold up.
  • watch — mixed signals, or company-wide risk that hasn't reached <function> yet.
  • risky — a function-scoped hit (layoff in your org, freeze on your reqs, cost-center in a downturn) or a live acute catalyst. Bias toward <function>: a 4.5-Glassdoor company that just cut your exact org is risky, not healthy.

4c — Provenance. built-from-data (≥3 dimensions with credible sources), needs-more-info (thin), or stale (reused past recheck_days). Mirror compEstimate.

4d — Cross-cut with candidate needs (the ONLY path to the fit score). The rating is its own standalone score by default and does not touch fit. It moves fit only where a poor dimension intersects a candidate_need, the same way benefits feed fit only where the candidate values them:

Candidate needCross-cuts dimension(s)
stability (dependents, can't take a risky bet)layoff risk, financial health
growthhiring momentum, function-is-cost-center
culture / moralesentiment, leadership stability
compfinancial health (offer-stage risk)
  • Intersection found → set a small fitDelta (negative; e.g. -2 mixed, -5 poor, per intersecting need). Record which needs in crossCut[]. Never a hard kill on its own — it pairs with fit_floor + the cold-family down-weight; a severe cross-cut can drag fit below floor via the need it hits, but health alone never gates.
  • No intersection → fitDelta: 0, crossCut: []. The rating still shows as its own badge.

STEP 5 — Persist to the tracker (Tracker Write Contract)

dashboard-data.js runs in the browser (no server imports / file reads — see the dashboard data-pipeline constraint in AGENTS.md). So the rating must be DERIVED here and PERSISTED onto workspace/tracker.json; the renderer only reads it. Mirror the compEstimate provenance pattern.

  1. On the company's applications[] row (or sourced[] row), set:

    "companyHealth": {
      "rating": "healthy|watch|risky",
      "forFunction": "<target_function>",
      "asOf": "YYYY-MM-DD",
      "provenance": "built-from-data|needs-more-info|stale",
      "crossCut": ["stability"],          // needs it intersected; [] if none
      "fitDelta": -5,                       // applied to fit ONLY via crossCut; 0 if none
      "dimensions": {
        "layoffRisk":     { "level": "poor",  "note": "...", "functionHit": true },
        "hiringMomentum": { "level": "mixed", "note": "..." },
        "financial":      { "level": "good",  "note": "..." },
        "sentiment":      { "level": "mixed", "note": "...", "trend": "down" },
        "leadership":     { "level": "good",  "note": "..." }
      },
      "rationale": "one-paragraph why, role-scoped",
      "signals": [ { "source": "...", "date": "YYYY-MM-DD", "summary": "...", "url": "..." } ]
    }
    • DB workspace: write it through the dedicated verb, not set-fields/upsert-batch — this is the required write path, and the only one that validates the shape (rating/provenance enums, asOf format, fitDelta sign) and guards against a current_base leak before it touches the row:

      careerrat health record <applicationOrSourcedId> --file <rating.json> --write

      <applicationOrSourcedId> resolves against applications[] first, then sourced[] — works for either row without knowing which table it's in. Drop --write first to dry-run: it validates the payload and previews the target row without committing anything.

    • Legacy workspace (no DB): open workspace/tracker.json and set the object above directly on the matching row.
  2. Bump meta.lastUpdatedAt (the freshness stamp every writing skill bumps).
    • DB workspace: careerrat health record --write already bumped meta.lastUpdatedAt (and meta.version) in the same transaction — no separate action needed.
    • Legacy workspace (no DB): stamp it by hand in the same write as step 1.
  3. Verify + snapshot (the dashboard handoff):

    • DB workspace: step 1's call already persisted and auto-exported workspace/tracker.json + workspace/activity.jsonl. Run:

      careerrat data verify
      careerrat tracker --verify
    • Legacy workspace (no DB):

      careerrat tracker --verify

    Fix and re-run until it passes clean. Render must never write tracker.json.

Conversational web handoff

In conversational chat, this skill runs as an embedded session under the chat tool profile (CHAT_RUNTIME_TOOLS), which has no Bash — there is no shell to run careerrat health record ... --write from, and no server-side file access to hand-edit workspace/tracker.json either. Once the companyHealth object above is fully composed (same shape, same rating/provenance enums, same current_base guard — nothing about the payload itself changes), emit it as one exact fenced block instead of running steps 1-3:

{"kind":"company_health_result","targetType":"application|sourced","targetId":"<applicationOrSourcedId>","company":"<Company>","companyHealth":{ "rating": "healthy|watch|risky", "forFunction": "<target_function>", "asOf": "YYYY-MM-DD", "provenance": "built-from-data|needs-more-info|stale", "crossCut": [], "fitDelta": 0, "dimensions": {}, "rationale": "...", "signals": [] }}

The app renders this as a real Save to workspace / Discard control. Saving calls the confirm-first company.health-record intent, which runs the write server-side through the exact same companyHealthSet/validateCompanyHealth guards (rating/provenance enums, asOf format, fitDelta sign, current_base leak guard) step 1's careerrat health record CLI path uses — so a malformed or leaking payload is refused there too, not silently accepted. Do not tell the user the rating is saved until they confirm; a prose reply is never itself a write. Skip steps 1-3's CLI/manual-edit paths entirely in this mode — the confirmed intent already bumps meta.lastUpdatedAt and persists + exports workspace/tracker.json, the same single transaction careerrat health record --write runs. STEP 5's numbered steps above stay exactly as written for a one-shot, non-embedded (external-agent) run, where there is a real shell (DB workspace) or direct file access (legacy workspace).


STEP 6 — Log to Activity Pulse

  • DB workspace: step 5's careerrat health record --write call already logs the rating-specific event ("Company health: <Company> — <rating>", type research) in the same transaction — no separate careerrat data activity append call needed. The same is true in conversational chat: the confirmed company.health-record intent calls the exact same companyHealthSet verb, so the Activity Pulse event is already logged once the user confirms the save — no separate step.
  • Legacy workspace (no DB):

    careerrat activity append --type research --actor agent \
      --title "Company health: <Company> — <rating>" \
      --summary "<function>-scoped: <the one driving signal>" \
      --company "<Company>" --write

STEP 7 — Hand off

Tell the user, in the configured voice register:

  • The role-scoped rating + the as-of date + provenance.
  • The one signal that drove it (e.g. "Eng RIF 3 weeks ago — your exact org").
  • Whether it touched the fit score, and which need it cross-cut (or "standalone — didn't move fit").
  • Consumers: evaluate-job reads companyHealth.fitDelta into scoring; the dashboard shows the badge on the card + the health section in the drawer; research-company embeds the section in a dossier when one is built.

Required Output

Emit after STEP 5 completes:

COMPANY HEALTH: <Company> — <healthy|watch|risky> for <function>
AS OF: <YYYY-MM-DD> (<provenance>)
DROVE IT: <the decisive dimension + note>
CROSS-CUT: <needs, or "none — standalone, fit unchanged">  | fitDelta: <N>
DIMENSIONS: layoffs <lvl> · hiring <lvl> · financial <lvl> · sentiment <lvl> · leadership <lvl>

If a dimension produced nothing credible:

NEEDS MORE INFO: <dimension(s)>

Honesty / freshness rails (hard)

  • Internal signal ONLY. Never leaks into an outbound artifact — no defamation-y phrasing in anything a human at the company could read. This is the Honesty Firewall applied to a risk read.
  • Always stamped + decaying. Show asOf + provenance; re-check past recheck_days or at the offer band; never present a lagging/noisy signal as certain.
  • Tempered sentiment. Glassdoor/Blind are noisy and biased — down-weight them, prefer trend over snapshot, and never let a single bad review set the rating.
  • Judgment surface, not a hard cut. The rating never kills a role by itself; it informs and, via a stated need, nudges. The candidate decides.

Domain-Neutrality Rule

No hardcoded role families, companies, or domain terms in this skill's behavior. The <function> and candidate_needs come entirely from targeting.yml / the tracker row. With no config, the skill scopes to the role on the row and treats the rating as standalone (zero fitDelta). See the Domain-Neutral Rule in AGENTS.md.

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.