company
Skill4 starsChanged 12 days agoMITView on GitHub
A company profile with an open contract and an optional private template. Keeps a current, source-backed picture of an organisation in `<org>/_company/` that people and agents can act on: create the skeleton, collect sources, check for drift, build one readable profile, export derived files. The structure is open; a question set, if you have one, is read from `company.template:` and is not part of this repository.
5 other projects use a version of this.
Adopted by
Most recently taken up by
- ckaico/reversiumtheir own version, first seen Nov 15, 2026
- varunmoka7/resume-hypergraphtheir own version, first seen Oct 2, 2026
- devmesh-in/claude-companytheir own version, first seen Aug 25, 2026
- dilanhendadura/company.mdtheir own version, first seen Aug 23, 2026
- MasatoshiSano/seigi-companytheir own version, first seen Mar 22, 2026
Other versions5 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 5
- devmesh-in/claude-company★ 8same name
- dilanhendadura/company.md★ 1same name
- MasatoshiSano/seigi-company★ 1same name
- ckaico/reversium★ 0same name
- varunmoka7/resume-hypergraph★ 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 repository13 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- analyticsSkill
- correspondenceSkill
- handoffSkill
- inboxSkill
- insightsSkill
- ops-baseSkill
- opsSkill
- outboxSkill
- preparationSkill
- tasksSkill
- transcriptSkill
- update-skillsSkill
The file
We split this skill into the rules it gives an agent
3 rules. 0 are common, 1 is distinctive, and 2 are very specific to this project.
0
Common
Most of the 3 edited versions we read keep these.
1
Distinctive
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 50
Before --display, check how the company itself writes its name (spelling, diacritics, casing) and use that.
- Distinctive: line 72
Keep any attribution the template states and show it to the owner with the first fetched material.
- Very specific: line 95
Synthesis: If company.template: is set, read it as described under init before this step.
“Common” is measured: kept by at least 6 in 10 of the 3 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
/company -- the company profile (CR-125)
Agents that write for a company (a proposal, a landing page, an agenda, an ad) keep re-deriving the same facts:
what the company is, who it sells to, what it may claim, how it sounds. /company keeps those facts in one
place, with the source of each, so every other skill can read them instead of asking again.
Open and optional. This skill carries the structure: where the profile lives, what its files are called,
what they carry, the rules, the checks. It carries no questions and no method. What to ask, in which order,
to what standard, comes from an optional template you point at with company.template:. Without one the skill
still works with the plain skeleton; it tells you so. Nothing in this repository reproduces a template.
Source-backed, not self-described (CR-167). A facet file is built from sources and records them: every
statement can be traced to where it came from, check reports when the sources have moved on, and a fact
nobody has is written as a gap rather than filled in. Where the only source is a person, that is recorded too
-- an owner's answer is a dated interview:<date> source, not an unmarked fact. A profile written once from
memory is a different object with the same filename: it is as true as the day it was written and has no way
to say so.
The contract is references/contract.md; read it before writing a profile file.
When to use it
| You want | Use |
| The standing facts about a company, with sources, readable by agents | /company |
| What was said and decided in a meeting | /transcript, /ops process |
| A living document with one person | /correspondence |
| Where one project stands | /ops orient <folder> |
A profile is current truth about the company, not a record of events. It is regenerated or edited when the facts change; events go to the meeting archive.
Subcommands
init <org>
Creates <org>/_company/ with the plain skeleton for the configured level (company.level, default basic):
one file per facet (all nine), README.md, AUDIT-LOG.md, empty sources.yaml and brands.yaml, .intake/.
Before --display, check how the company itself writes its name (spelling, diacritics, casing) and use that.
Run python3 "${CLAUDE_SKILL_DIR}/company_tool.py" init <org-folder> --vertical <v> [--display Name] [--level L] [--language xx] [--template ADDRESS].
Level, language and template default to the company: block of <org>/_ops.yaml (then the vault's _config/base.yaml); a flag wins.
${CLAUDE_SKILL_DIR} is this skill's own folder; where an assistant does not substitute it, company_tool.py sits in the same folder as this file.
It never overwrites an existing folder. With company.template: set, read the question set from there and say
which; without it, say that no question set is attached and stop at the skeleton.
Reading a template. company.template: is either a folder path or an MCP address. The script only records it;
the question set is read by you, the agent, and is never copied into this repository or into a shared place.
- Path: read the files the template names, in place.
- Items: when the set numbers its items, record every required one in
coverage.yaml(contract section 9): answered withwhere, orna/open/decision/workshopwith a reason. Classify with the ordered rules in contract section 9 (the same text must give the same class on every run). Map the profile's level to the set's withcompany.level_mapto know which are required. Runcheck --items <ids.json>so no item goes missing silently. - Address: the template is a read-only MCP server that needs an account token. The server is registered once in
the client (for Claude Code:
claude mcp add --transport http <name> <address> --header "Authorization: Bearer <token>", user scope; the token lives in the client's own configuration, never in a vault, a profile or a repository). If its tools are available in this session, start with the one that lists the template's levels and attribution, then fetch only what the facet in hand needs. If they are not available, or the server answers 401, say so, namecompany.template:as the cause, and continue with the plain skeleton. Do not guess a question set. - Keep any attribution the template states and show it to the owner with the first fetched material.
- Fetched questions are working material: use them to interview the owner and to judge what is missing, write
only the answers (with source markers) into the facet files, and put nothing from the template in
.intake/,README.mdorAUDIT-LOG.mdbeyond item identifiers.
extract <org>
Starts from the files the company handed over (PDFs, saved web pages) and turns them into text you can read and
cite. python3 "${CLAUDE_SKILL_DIR}/company_tool.py" extract <org-folder> [--source DIR] [--out DIR] reads
<org>/source-documents/ (or --source), writes one text file per document into <org>/.extracted/ with
=== page N === markers, the page count, the sha256 and, for a web page, its address when the page carries one,
and a _manifest.md. It lists the pages that are mostly images (fewer than 25 words): render those to read them
(pdftoppm -r 70 -f N -l N -png file.pdf prefix) and say in the facet that a fact comes from a page image. It
never changes a source and needs pdftotext (poppler) for PDFs. The original is the source in sources.yaml
(curation: original, pages: N, origin_url and as_of for a saved page); the extracted file is a working copy.
What extraction cannot carry (images, diagrams, colour swatches, tables) goes in the README under
## Reviewer notes and extraction losses.
intake <org>
- Sources first. Fill
sources.yaml: every place a fact may come from (vault paths, hosts, repositories, URLs, dropped material in.intake/), each withauthorityand the facets itfeeds; and anexcludelist of what is off-limits, with a reason. Read sources in place and never edit them. - Synthesis. If
company.template:is set, read it as described underinitbefore this step. For each facet in play, list per item: what is established (with source markers), what conflicts (both values, both sources, never resolved silently), what is missing, what needs a decision. Summarise the understanding for the owner to confirm before writing facet text. Put it in the README (## Synthesis, one row per facet). If the owner is not available, do not stop: write the synthesis, setOwner confirmation: pending, keep every facetdraft(neversynthesis-locked), and write on. - Write each facet from its sources, one facet at a time, within its own scope; link to other facets
instead of repeating them (contract rule 12 says which facet is the home of which fact). Open items read
_Not provided yet._Apply the never-say list (## Never sayinmessaging; contract rule 6 says how to make one when no template supplies it). Sources stay where they are:.intake/holds only dropped material that exists nowhere else. - Decisions are not made by the skill. Mark each as
**Decision needed:** <question> (<owner role>), list it under## Decisionsin the README and hand it to the owner; record the owner's answer as a datedinterview:<date>source.
check [org]
Report only; nothing is rewritten.
python3 "${CLAUDE_SKILL_DIR}/company_tool.py" check <vault-or-org-folder> [--copies] [--summary] [--strict] [--items ids.json] [--items-from items.json] [--leak-from items.json] [--max-age-days N] [--json]The script reports a stale facet (no review or generation within company.max_age_days, default 365, or status
stale), schema errors in the frontmatter and in sources.yaml / brands.yaml, a facet marked draft-locked or
final with no sources, a marker nobody defines, a facet H1 whose display name differs, a credential-looking value anywhere in a profile, and with --copies any profile
file outside _company/. --items-from items.json derives the required ids itself from the set's items (as the server returns them, saved as received), company.level_map and company.level_order; company_tool.py required-ids <org> --items-from items.json [--out ids.json] writes the same list to a file. --leak-from items.json takes the question wording of the attached set (as the server returns it) and reports every profile file in which that wording appears (item id and line, never the text): a profile holds answers, not questions. check reads company: from the config, so max_age_days and level_map (the question set's level names are accepted in frontmatter) apply without flags. --strict also reports what the overview graph needs (contract section 8): conflicts written C<n> with two
[s-id] sides and their facets, local markers that name a register id, decisions in the one form with an owner
role, feeds that match what a facet cites. Then read the files for what a program cannot decide:
- a status claim ("active", "live") with no recent evidence, or a paused business written as running;
- a fact without a source marker, or a number without a date and a definition;
- a contradiction between facets, or a never-say breach;
- a guide judged "missing" because the folder holding it was excluded: check
excludebefore concluding.
Offer fixes; apply none without being asked.
build [org]
Assembles the facets in play into one readable profile for a person or an agent: in facet order, sources
resolved, gaps and conflicts kept visible. Writes to exports/ only if asked; the facet files remain the source.
export <target> [org]
Writes a derived file for another system into exports/<target>/. Derived means regenerated, never hand-edited.
The first target is an external search-visibility tracker: the subject, one buying category, named competitors
(companies only, never the brand itself), six to twelve organic buying situations that do not name the brand,
and a short context line per market language. Where search data exists it beats the profile. Service keys are
referenced from company.services, never written into an export.
help
Print this skill's subcommands, the contract's location (references/contract.md), and whether company.template: is set.
Configuration
company: in <org>/_ops.yaml or the vault's _config/base.yaml (defaults in ops-config/base.yaml):
company:
level: basic # public | basic | standard | complete
language: en
max_age_days: 365 # a facet older than this is reported stale
template: null # path or address of an optional private template; null = the plain skeleton
level_map: {} # profile level -> the attached question set's own level name; no default is shipped
level_order: [] # the question set's own level names, lowest first; no default is shipped
services: {} # references only: {ranking: {account: acme, key_ref: "env:RANKING_KEY"}}A key value in any profile, sources.yaml, brands.yaml or repository is an error.
Readers
/ops (orient, prepare, check) and /preparation load the facets relevant to the task when a _company/
exists for the organisation, and say nothing when it does not. A reader never writes to a profile. Skills in
other repositories that write for a company (brand imagery, website scaffolding, change-request intake) read
the same folder under the same contract.
Privacy
A profile is customer data. It lives in the organisation's own area, is never an example in a skill or a test, and never carries credentials or private one-to-one content. Examples here use invented companies.
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.

