research-company
Skill4 starsChanged 13 days agoMITView on GitHub
Web-search a company across six domain-neutral axes → compose a cited workspace/research/company-slug.md artifact that feeds interview-prep and evaluate-job FIT context.
4 other projects use a version of this.
Adopted by
Most recently taken up by
- spirosbax/insaighttheir own version, first seen Oct 13, 2026
- Ben-Bot-22/job-hunt-kittheir own version, first seen Oct 9, 2026
- GetBrew/growth-engineertheir own version, first seen Sep 30, 2026
- sree-data/job-search-agenttheir own version, first seen Sep 11, 2026
Other versions4 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 4
- GetBrew/growth-engineer★ 218same name
- spirosbax/insaight★ 32same name
- Ben-Bot-22/job-hunt-kit★ 2same name
- sree-data/job-search-agent★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository31 other files it gives its agents.
- AGENTS.mdAGENTS.md
- templates/AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- templates/CLAUDE.mdCLAUDE.md
- answer-questionSkill
- apply-jobSkill
- calendar-syncSkill
- coach-gapsSkill
- company-healthSkill
- configureSkill
- discover-companiesSkill
- email-commsSkill
- evaluate-jobSkill
- ingest-mailSkill
- ingest-messagesSkill
- ingest-profileSkill
- intake-extractSkill
- interview-prepSkill
- optimize-linkedinSkill
- reevaluate-strategySkill
- relationship-sourcingSkill
- report-issueSkill
- research-boardsSkill
- research-compSkill
- resume-extractSkill
- schedule-meetingSkill
- search-jobsSkill
- setup-searchesSkill
- sync-statusSkill
- tailor-applicationSkill
- track-outcomesSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
18 rules. 0 are common, 8 are distinctive, and 10 are very specific to this project.
0
Common
Most of the 4 edited versions we read keep these.
8
Distinctive
This author's own choices. Worth a look.
10
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 17
Use this skill when the user asks to research, look up, or pull intel on a company before an interview or application evaluation.
- Distinctive: line 30
Read the following before doing anything else:
- Very specific: line 42
Default six research axes (substitute <domain> and <role> from config):
- Very specific: line 51
If candidate/research-prefs.yml exists and specifies research_axes, use those instead.
- Very specific: line 59
Use --explicit when the user asked for this research.
- Very specific: line 76
If the command returns a non-stale artifact (within staleness_days), show the user the artifact summary and offer two choices:
- Distinctive: line 80
If the command returns a skip note ("no artifact" or "stale"), continue to STEP 2.
- Very specific: line 94
Run one WebSearch query per axis, capped at max_searches_per_company (default 6).
- Distinctive: line 108
If the snippet is rich enough to cite a specific fact, use it directly.
- Very specific: line 109
If the snippet is thin or ambiguous, WebFetch the top URL to read the full page.
- Very specific: line 114
Graceful degradation: If a search returns zero usable results for an axis, note it in ## Open Questions ("No reliable sources found for …
- Distinctive: line 220
Every body claim must carry a provenance marker.
- Very specific: line 258
Read the output. If record refuses (placeholder residue, missing frontmatter field, zero sources, current_base leak), fix the draft and …
- Very specific: line 276
Delete the draft file. The canonical artifact now lives at workspace/research/<company-slug>.md.
- Distinctive: line 288
If any axes landed in Open Questions, name them so the user can decide whether to supply missing context manually.
- Distinctive: line 295
If the user states during the research flow that this company is a no ("I don't want to apply there", "add them to the exclusion list", …
- Distinctive: line 324
If any axis produced zero results, add:
- Distinctive: line 378
All fields below have agent-side defaults so absent files keep working:
“Common” is measured: kept by at least 6 in 10 of the 4 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
research-company
Use this skill when the user asks to research, look up, or pull intel on a company
before an interview or application evaluation. Also use it automatically when
interview-prep or evaluate-job needs company context and no non-stale artifact
exists yet.
Runs under AGENTS.md. These contracts bind without being restated here: Privacy Invariant (
current_basenever outbound), Honesty Firewall, Placeholder/Bracket Ban, Gate Write-back, Domain-Neutral Rule, Browser Automation Contract, Activity Pulse logging, Tracker verify+snapshot, and Sent-Clears-Draft. Inline reminders at point-of-use are intentional; standalone restatements point back to the relevant AGENTS.md section. Barecandidate/,workspace/,config/, and.internal/paths below are symbolic; resolve them per AGENTS.md's Path Resolution rule.
Agent voice. Read
candidate/modes.yml#agent_voice(defaultstandard) before producing the research summary presented to the user. Apply the register from AGENTS.md#mode-switches. Theworkspace/research/<slug>.mdartifact is always written in full — register governs the in-chat summary:exec-summary= 3–5 bullet signals + file path;standard= short section-per-axis bullets + file path;technical= signal analysis + sourcing notes;verbose= full artifact mirrored in chat.
STEP 0 — Load config
Read the following before doing anything else:
| File | Fields used |
candidate/profile.yml | candidate.domain, candidate.location (string, for query context) — NEVER read compensation.current_base |
candidate/targeting.yml | role_buckets (for domain/role context), excluded_companies |
candidate/research-prefs.yml | research_axes, staleness_days, max_searches_per_company — works if absent; defaults are: the six axes below, 14 days, 6 searches |
Derive candidate_domain from profile.yml#candidate.domain (e.g. "logistics operations",
"financial services", "healthcare administration"). Substitute it wherever axes reference
<domain> below. Never hardcode a tech term.
Default six research axes (substitute <domain> and <role> from config):
- Overview — company mission, size, funding stage or public status, headquarters
- Trajectory — revenue trend, headcount growth or contraction, recent funding, IPO/M&A
- Recent Moves — product launches, layoffs, leadership changes, press coverage (last 12–18 months)
- Reputation and Culture — employee sentiment (Glassdoor, Blind, LinkedIn), DEI posture, stated values vs. reported reality
- Role-Relevant Challenges — open problems or pain points in
<domain>that<role>candidates are typically hired to solve - Competitive Position — main competitors, market share narrative, differentiation claims
If candidate/research-prefs.yml exists and specifies research_axes, use those instead.
Then check usage mode:
careerrat modes allows research:company --explicitUse --explicit when the user asked for this research. Their request authorizes this
one run and must not trigger another override question or change their saved usage mode.
For an automatic handoff from another skill, omit --explicit; stop if that verdict is
skip. If the explicit verdict is downshift, run a single overview query. If it is
run, continue normally.
STEP 1 — Resolve company and check freshness
- Canonicalize the company name as given by the user or inferred from the JD.
Run:
careerrat research read "<Company>"- If the command returns a non-stale artifact (within
staleness_days), show the user the artifact summary and offer two choices:- Reuse — proceed with the existing artifact; stop here.
- Refresh — continue to STEP 2 and overwrite.
- If the command returns a skip note ("no artifact" or "stale"), continue to STEP 2.
- If the command returns a non-stale artifact (within
STEP 2 — Run web searches per axis
[DELEGATE: subagent] The axes are independent, so this phase fans out — spawn one
subagent per axis (parallel; WebSearch/WebFetch has no one-browser limit). Hand each
its axis topic + company + candidate_domain and have it return a compact cited block
({axis, facts:[{claim, url, title, fetchedISO, confidence}]}); merge them in STEP 3. The
orchestrator already cleared STEP 0's config load and STEP 1 freshness, subagents
do not re-gate, do not read AGENTS.md, and do not write any file. Degrade to inline
sequential search when no subagent primitive exists. See the Delegation Contract in AGENTS.md.
Run one WebSearch query per axis, capped at max_searches_per_company (default 6).
Construct each query from the axis topic + company name + candidate_domain context.
Example query shapes (substitute real values — never paste these verbatim):
- Overview:
"<Company>" company overview size funding - Trajectory:
"<Company>" revenue growth headcount 2024 2025 - Recent Moves:
"<Company>" layoffs product launch leadership news site:techcrunch.com OR site:reuters.com - Culture:
"<Company>" glassdoor reviews culture employee sentiment - Role-Relevant:
"<Company>" <domain> challenges open problems <role> - Competitive:
"<Company>" competitors market share differentiation
For each search result:
- If the snippet is rich enough to cite a specific fact, use it directly.
- If the snippet is thin or ambiguous,
WebFetchthe top URL to read the full page. - Record the URL, page title, fetch date (today's ISO date), axis, and confidence (high = primary source or credible outlet; med = aggregator or secondary; low = forum, anecdote, or single-source claim).
Graceful degradation: If a search returns zero usable results for an axis, note it
in ## Open Questions ("No reliable sources found for [axis name]"). Do not retry in
a loop; move on.
STEP 3 — Record per-claim provenance
As you collect facts, tag every claim with one of three markers:
Verified evidence tier — do NOT use for research. Research artifacts are NOT
candidate/evidence.yml and must never feed directly into resume claims, cover-letter
claims, or interview-packet assertions.
Sourced-web fact — a datum from a specific cited URL. Marker format (colon and attribution are REQUIRED — a bare bracket with no colon is refused as placeholder residue):
[source: "<Title>" (<url>), fetched <ISO date>, confidence: high|med|low]Markdown links [Title](url) are also accepted inline.
Agent-inferred — synthesis across multiple sources with no single owning URL. Marker format (colon and attribution list are REQUIRED):
[AGENT-INFERRED from: <url1>, <url2> — candidate to verify]Conflicts — when two sources disagree on a material fact, record BOTH; never silently resolve:
[CONFLICT: <SourceA> says X (<urlA>, conf: high); <SourceB> says Y (<urlB>, conf: med). candidate to verify]Count sourced URLs, inferred claims, and conflicts as you go — you will need the totals for the Required Output block.
STEP 4 — Compose the artifact to a temp draft
Compose the full artifact to a temporary draft file. Use a path outside
workspace/research/ — for example workspace/research/.<slug>.draft — so it is
not itself listed by careerrat research list. Do NOT write a .md file directly
into workspace/research/; the record --write command is the only write path.
The artifact must follow this exact structure:
---
type: company-research
company: "<Canonical Company Name>"
slug: "<company-slug>"
fetchedAt: "<today ISO date>"
researchedFor: "<role title if known, else 'general'>"
staleness_days: 14
sources:
- url: "https://..."
title: "..."
fetchedAt: "<ISO date>"
axis: "<axis name>"
confidence: high|med|low
---
## Overview
<Paragraph. Every claim carries a [source: ...] or [AGENT-INFERRED from: ...] marker.>
## Company Trajectory
<Paragraph or bullets. Sourced markers on every claim.>
## Recent Moves (last 12–18 months)
<Bullets. Date each item. Sourced markers.>
## Reputation and Culture
<Paragraph. Include Glassdoor/Blind signals if found. Sourced or conflict markers.>
## Role-Relevant Challenges
<Bullets framed in candidate_domain terms — never hardcoded tech. Sourced markers.>
## Competitive Position
<Paragraph or bullets. Sourced markers.>
## Candidate Angle
<Entirely AGENT-INFERRED. Mark every sentence: [AGENT-INFERRED from: <urls> — candidate to verify].
This section connects observed company themes to role-relevant strengths the candidate
may choose to emphasize. It is synthesis, not sourced fact.
NEVER transfer any sentence from this section into evidence.yml, a resume,
a cover letter, or an interview-packet claim.>
## Open Questions
<Items where no reliable source was found, conflicts that need human resolution, or
axes that returned zero results. Format: "- <axis name>: <what is unknown>" — do NOT
wrap the axis in square brackets; a bare `[axis]` token is refused by the placeholder
lint as residue.>
## Privacy Note
Contains no candidate compensation data; no private compensation inputs are present in this artifact.Frontmatter staleness_days must match research-prefs.yml#staleness_days (default 14).
The sources[] list must have at least one entry; record --write will refuse an artifact
with zero cited sources.
Every body claim must carry a provenance marker. A section with unmarked prose will be refused by the placeholder lint.
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
research record ... --write from. Once the artifact in STEP 4 is complete and clean (same
frontmatter, same citation and placeholder-lint requirements — nothing about the artifact
itself changes), emit it as one exact fenced block instead of writing a draft file:
{"kind":"company_research_result","company":"<Canonical Company Name>","slug":"<company-slug>","markdown":"<the full STEP 4 artifact text, frontmatter and body>"}The app renders this as a real Save to workspace / Discard control. Saving calls the
confirm-first research.record intent, which runs the write server-side through the exact
same computeResearchWrite/writeResearch guards (citation-hygiene, placeholder lint,
current_base privacy) STEP 5's CLI path uses — so a bad artifact is refused there too, not
silently accepted. Do not tell the user the research is saved until they confirm; a prose
reply is never itself a write. Skip STEP 5's CLI commands entirely in this mode — the
confirmed intent already logs the Activity Pulse event and updates meta.lastUpdatedAt, the
same as record --write does. STEP 5's careerrat research record CLI path stays exactly as
written below for a one-shot, non-embedded (external-agent) run, where there is a real shell.
STEP 5 — Dry-run, fix, commit, clean up
External-agent / one-shot CLI runs only. In conversational chat, use the Conversational web handoff above instead — skip straight to STEP 6.
Dry-run to validate and preview:
careerrat research record "<Company>" --file <path-to-draft>Read the output. If
recordrefuses (placeholder residue, missing frontmatter field, zero sources,current_baseleak), fix the draft and re-run. Do not proceed to--writeuntil the dry run passes clean.Commit:
careerrat research record "<Company>" --file <path-to-draft> --writeLog the research to the Activity Pulse feed (see Activity Pulse in AGENTS.md):
careerrat activity append --type research --actor agent \ --title "Researched <Company>" --summary "<axes covered, e.g. 'product, funding, culture'>" \ --company "<Company>" --write- Delete the draft file. The canonical artifact now lives at
workspace/research/<company-slug>.md.
STEP 6 — Confirm and hand off
Tell the user:
- The artifact is saved at
workspace/research/<company-slug>.md. interview-prepwill read it automatically when preparing a packet for this company.evaluate-jobFIT context for this company is now available.- If any axes landed in Open Questions, name them so the user can decide whether to supply missing context manually.
STEP 7 — Handle mid-research exclusion
If the user states during the research flow that this company is a no ("I don't want to apply there", "add them to the exclusion list", "this is a no"):
- The user stated it outright, so write it now. It applies directly and is logged with an Undo.
Run:
careerrat gate exclude-company "<Company>" --write- Echo the CLI confirmation.
careerrat gatewrites SQLite in DB mode and legacy YAML only in legacy mode. - The research artifact can remain — it costs nothing and may be useful for competitive
intel — but note to the user that
evaluate-jobwill CUT any posting from this company.
Required Output
Emit this block after STEP 5 completes:
RESEARCH: <Canonical Company Name> — <company-slug>
AXES COVERED: <comma-separated list of axes researched>
SOURCES: <N> urls
CONFIDENCE: <X> high / <Y> med / <Z> low
CONFLICTS: <N>
ARTIFACT: workspace/research/<company-slug>.mdIf any axis produced zero results, add:
OPEN QUESTIONS: <axis1>, <axis2>Privacy Invariant
See Gates › Privacy invariant (hard) in AGENTS.md. The artifact is treated as outbound — write it as if a recruiter could read it;
current_basemust never appear in any research artifact, summary, or Candidate Angle section.
Inferred-Facts Rule (Citation-Hygiene Firewall)
Three fact tiers govern this skill. Understand and enforce all three:
Tier 1 — Verified evidence (candidate/evidence.yml): the only tier permitted in
resumes, cover letters, and interview-packet claims. Research never feeds this tier
directly. Never copy a finding from a research artifact into evidence.yml without the
candidate's explicit instruction and their own verification.
Tier 2 — Sourced-web fact: a datum from a cited URL, tagged with a [source: ...]
marker (colon and full attribution required). These are the body of the artifact.
Tier 3 — Agent-inferred: synthesis across sources, tagged [AGENT-INFERRED from: <urls> — candidate to verify]
(colon and url list required). The ## Candidate Angle section is entirely Tier 3.
The firewall: Tier 3 claims — especially anything in ## Candidate Angle — must
never be copied into candidate/evidence.yml, a resume, a cover letter, or an
interview-packet assertion. They are context for the candidate to verify, not sourced
facts. evaluate-job and interview-prep read research artifacts for FIT context only;
they do not promote research findings into evidence.
A bracket marker with no colon (e.g. a bare [AGENT-INFERRED] or [source]) is refused
by the placeholder lint as placeholder residue — every marker must carry its attribution.
Domain-Neutrality Rule
This skill is field-agnostic. Read profile.yml#candidate.domain and substitute it into
every axis query. A candidate in healthcare gets "healthcare operations" queries; a candidate
in logistics gets "freight and logistics" queries. Never hardcode tech terms, engineering
jargon, or industry-specific language anywhere in the artifact. The six axes are neutral
skeletons; their queries and body prose interpolate candidate_domain and role context
from the candidate's own config. A no-config run (absent research-prefs.yml) is always
possible — all defaults are field-agnostic.
Config dependencies
All fields below have agent-side defaults so absent files keep working:
candidate/research-prefs.yml#research_axes— if absent, use the six default axes above.candidate/research-prefs.yml#staleness_days— if absent, default 14.candidate/research-prefs.yml#max_searches_per_company— if absent, default 6.candidate/profile.yml#candidate.domain— if absent, omit domain qualification from queries (still run all six axes).candidate/targeting.yml#excluded_companies— checked in STEP 7 only; if absent, no exclusion check is performed.
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.

