add_standard
ChatGPTAdd ONE standard to this workspace so the Compliance record tracks it — one call per standard the person said they want. Pass its code (e.g. "iso27001"); an unknown code returns the codes this workspace can add, so never guess one from memory. Adding a standard that is already added only sets up the required checks still missing. Adding one sets up the checks it requires on the documents that exist; the OUTCOME line says how many, which processes wait for their document and how many need an app or a register first. Only the workspace Owner may add one; it applies at once. Read the OUTCOME line before saying what happened.
answer_card
ChatGPTAnswer what an Alchex chat is waiting on (a question, an approval or a plan proposal) with the person's own answer. Pass on ONLY what the person said or chose: never answer for them, never pick an option they did not pick, never approve on your own judgment. Give optionId for the choice they made (from read_chat), text for their words, or both. The first answer counts: if it was already answered (in the app or elsewhere) this says so and changes nothing. Read the chat on with read_chat.
find_chats
ChatGPTThe person's own chats in Alchex, pinned first then newest: id, title, last message preview and link. Open one with read_chat.
get_knowledge
ChatGPTRead the house's own knowledge — the same for every organization, never this organization's records. Give exactly ONE of: guide (a guide or process of the house: how a piece of work is done), standard (what a standard, its guide or a regulation says: one clause exactly, a search in the customer's words, or the clauses a process answers), risks (the risk catalogue: the risks a sector, standard or law usually brings), laws (the laws we hold: which apply, what one requires), apps (the apps Alchex can connect and what each reads). Reads only.
read_chat
ChatGPTRead one of the person's chats: its messages oldest to newest (the newest 50), each with who wrote it and what it showed, plus state ("waiting" when it waits on the person) and link. A reply paused on a question or approval is not saved until it ends: it shows as pending (each question's prompt and options), answered with answer_card only in the person's own words. While hasMore, pass before to read further back.
read_chat_usage
ChatGPTOne chat's usage per reply: outcome, time, tokens, estimated credits, tools in order, changes with receipt links, guard blocks, rating, and totals. Never what anyone wrote. Only a chat you belong to, not the one this reply runs in.
read_document
ChatGPTRead this organization's own DOCUMENTS (policies, procedures, records of the management system) and its REGISTERS (risks and non-conformities, standalone findings, controls, and every register kept as an Alchex Sheet). What you pass narrows it: - documentId — open ONE document: its full text including tables, sections (the headings in order), scopeClauses (the clauses it is linked to), outstandingDecisions (the [bracketed] choices not made yet — they answer "what is still open?") and controls set up on it. Read it before changing it with write_document (quote its content exactly) or to quote a section. A clause is made compliant by its controls passing, not by linking a document. - recordId — open ONE record: a risk, non-conformity or finding with its full fields; a control (a library id like C-KIT-ISO-27001-2022-MFA-CRITICAL-SYSTEMS, or a set-up control's id: its settings and recent runs); a run (its steps and receipts); a receipt ("R-1042"). Read it before proposing changes with write_record. "Are we compliant with 7.5?" is answered by reading these. - query — the person NAMED or DESCRIBED something ("the access-control policy", "DOC-009", "RSK-014", "what does our policy say about MFA?"). Returns documents (title/code matches with their REAL id; matched says whether one is strong), passages (what bodies SAY about it) and records (risks and findings by code/name). If records.ambiguous is true, do NOT guess and do NOT write: name the close candidates in one line and let the person choose. - register — ONE register's rows ("risks", "findings", "controls", or a register's name): its columns, the rows keyed by column with their ids, the total and counts of its choice columns. where narrows before the counts, so "open high risks overdue with no owner" is ONE call: [{column:"Status",is:"Open"},{column:"Severity",is:"High"},{column:"Due date",before:"today"},{column:"Owner",empty:true}]. sortBy orders it. - standard — WHERE WE STAND on a standard (its id, its name like "ISO 27001", or "all"): READ IT FIRST on any resume ("where were we?", a Continue-with-Alchex prompt) and for "how compliant are we?", "are we ready for the audit?" (the upcoming certification audit), "what's next?", "which standards do we work with?". Lead with readiness and name its inputs, controlsPassing of controlsCounted: a control passes only when its latest run passed within its how-often. A document alone earns nothing; never count documentedClauses as compliance. Null readiness means the number cannot be read now: never estimate. proof is the page's own status per requirement (gap, needs_you, set_up, proven, not_applicable) with its reason; relay it, name the worst document. openControls (worst first: failed, partly_met, couldnt_run, stale, never_run) is what to fix next: open one by recordId, fix what its run found, run it again with run_control on the person's yes. roadmap is authored: relay its order, lead "what's next?" with nextStep (its ONE nextAction); blocked is advice, not a lock; draft: true means not reviewed yet; write one missing document per turn with write_document; scopeGap means link the existing one (its document check via write_record), never a second copy. excludedClauses/excludedLines are the SoA, out of scope with their reason, never gaps; scope is changed on the Compliance page. With clauseCode ("why is 4.1 not passing?", "what is attached to 4.1?"): that clause's record alone. status.lastCheck.rationale is the check's own finding: relay it with its requirement and sources, never re-judge it; no status means no check has run. With noEvidenceSince (a date): evidenceAge.stale and evidenceAge.neverChecked in ONE call, never clause by clause. - me (true) — WHO the person is in this workspace: the workspace's name, their role (OWNER, ADMIN or MEMBER), what it can and cannot do, and the billing they may see (plan, credit left and monthly allowance IN CREDITS, never money, and environment: a sandbox reading is test money, say so before a…
resume_job
ChatGPTGo on with a PAUSED job (read_document job first). Say why it paused and, at its ceiling, the credits used and a new ceiling; call this ONLY on the person's yes. Through this connector: read the job first and say why it paused; call this only on the person's yes to go on. The result's proofHref is the job's link; give it exactly.
rewind_chat
ChatGPTRewind a chat to just before one of the person's own messages: that message and every message after it leave the chat, for good. Use it only when the person asked to rewind or redo from that point. It applies at once, as the person, and cannot be undone.
run_control
ChatGPTRun ONE control set up on a document now, whether it is on or off (its id from the document's controls). Propose it ONLY after the person has confirmed: offer first, naming the document and the control, then run on their yes. It runs at once with the person's own authority and its verdicts land on the Compliance page when the run ends. Read the receipt and say what actually happened; never claim a run you did not see, and never run the same control twice in one turn. The run records its own results; a run that could not run says why on its page, where the person fixes the control's settings and runs it again.
save_answer
ChatGPTKeep the person's answer to a question Alchex asks about their organisation, so the Alchex assistant and every process see it and never ask it again. id is the question's id: a company-profile question (organization-name, industry, org-size, locations, …) or a process question (org.uses_ai, org.processes_personal_data, …, as get_knowledge shows them). Save only what the person actually said, in their words; a new answer replaces the old one and the old one stays in history. For anything else they told you, use save_note.
save_note
ChatGPTSave what the person told you unasked, so it is never asked again — in this conversation, any later one, or any process run. Call it on the turn they say it; one message often holds several things: one note each. Answers to your questions are kept by the app from their reply; never save them here. A note is a FACT about the company (company), what they want (goals), one process's decisions and reasons (process:<skill key>), or where a subject is kept (arrangements, with where pointing at that place, so later turns go straight there). A correction is a new note with replaces = the old id. Never save a choice about this conversation only ("just this one"), or anything they did not settle. When a change alters what a document rests on, name the affected documents and offer to update their drafts; change nothing before their yes, never publish. The app stamps the date, who said it and the conversation. Do not narrate the saving.
share_chat
ChatGPTShare a chat. with: "link" makes a read-only link for people in the workspace. with: "alchex" sends the chat to the Alchex team, with an optional note: it is the only way a conversation ever reaches the Alchex team, so use it only to report a wrong or poor answer. Use either ONLY when the person asked, in their own words. It applies at once, as the person; read the OUTCOME line before saying it is shared.
start_job
ChatGPTRun a BIG request (a whole process or standard) as a job on the server: it survives closing the chat, pauses at its ceiling or for a fact, and ends with a receipt. Small requests never use it. Call it ONLY after the person said yes to the plan steps, estimate and ceiling in credits you showed. Through this connector: show the person the plan steps, the estimate and the ceiling in credits first, and call this only on their yes. The result's proofHref is the job's link; give it exactly. Read how it goes with read_document job (the job id from that link).
stop_reply
ChatGPTStop the reply Alchex is writing in a chat, as the stop button in the app does. Only when the person asked to stop. Changes already saved by the reply stay saved.
upload_file
ChatGPTUpload a file from the person's computer into the Alchex Library: a PDF, PNG or JPEG up to 25 MB (an image is kept as a one-page PDF). It becomes a published document owned by the person. Pass folder (its path such as "People/Disciplinary cases", or its id) to file it there; filing proof into a folder a control's document names makes that control re-run about five minutes later. This call returns uploadUrl: send the raw file bytes to it with an HTTP PUT within 15 minutes, for example curl -sS -T "<path to file>" "<uploadUrl>". One link takes one file; call again for each file. The PUT answers with the new document and its link; tell the person what was uploaded and where.
write_chat
ChatGPTLook after one chat, named by conversationId. RENAME: title. PIN: pinned. ARCHIVE: archived: true (false restores it); a chat is never deleted here. RATE a reply: messageId + rating ("up" or "down"), in a call of its own. Rename, pin and archive are the chat author's. Read the OUTCOME line before saying what changed. An archive applies at once, as the person; the chat can be restored in Alchex.
write_chat_members
ChatGPTAdd one workspace member to a chat, by their user id; they will read the whole chat. Use it ONLY when the person asked, in their own words, to add that member. Only the chat's author can add. It applies at once, as the person; read the OUTCOME line before saying who was added.
write_document
ChatGPTwrite_folder
ChatGPTOrganize the Library's folders, one change per call. A folder is named by id or path ("Policies/HR"); read_document folders: true lists them. CREATE: name (+ parent, or teamId). CHANGE: folder + name, parent ("" = top level), owner. DELETE: folder + status: "Deleted"; its contents move up a level. FILE: file (+ folder, else top level). Read the OUTCOME line before saying what changed. A delete applies at once, as the person, with their permissions; the folders and records in it move up a level. Use it only when the person asked for that folder to go, and relay a refusal rather than retrying.
write_message
ChatGPTWrite the person's message into an Alchex chat as them: without conversationId it starts a new chat (shown in the app's chat list as started from this assistant), with one it continues that chat. It returns at once with conversationId, the chat's link and a cursor, while Alchex replies in the background; read the reply with read_chat from that cursor. A reply may run an approved plan, ask a question or ask for an approval; answer those with answer_card only in the person's own words. Uses the workspace's AI credits.
write_record
ChatGPTWrite a register row: add new ones, change one, or archive one — a risk or a non-conformity ("risks"), a standalone finding ("findings"), a row of a register kept as an Alchex Sheet (a training log, an exception register), or a control ("controls"). Name the register and give cells by COLUMN NAME, as the register shows them — no ids to look up. Use it when the person asks to log, add, raise, change, close or archive one, or when something real in this conversation holds entries the register is missing. ADD: no recordId; give rows, every row for one register in ONE call. A risk needs a Name; a non-conformity takes Category exactly "Non-conformity" (load the fixing-problems skill first). A sheet row also carries quote: the words it came from — the person's own sentence, or the passage you read it in. A control is one row with cells Document, Control, Every, Unit, Message: Document the document's id, Control a library control's id; Every and Unit (day, week, month, year) set its clock when the library does not, Message overrides its words. A standalone finding is recorded by a control run, never added here. NEW REGISTER: only when the person ASKS for one in words (a register a template merely names is not that ask), and after read_document shows it is missing — give register as its name and createRegister alone, with its columns (a select carries its choices) and, optionally, first rows EXTRACTED from this conversation, each with its quote; the register and its first rows are ONE write. A risk or non-conformity goes in "risks", never a new register. NEW TABLE: only when the person asks for one on an existing sheet register — give addTable alone, with its name and columns. NEW COLUMN: only when the person asks for one on a sheet register — give addColumns alone (no rows, no recordId), each with its name and type, and choices for a select; the columns start empty. CHANGE: give recordId and only the cells that change. CLOSING is Status "Closed"; ARCHIVING is Status "Archived" alone in cells (the row leaves the register and is restored there; a finding has no archive). Before closing or clearing a non-conformity, load the fixing-problems skill. A standalone finding takes only Status, Name, Category, Likelihood, Impact and Severity. A control is switched and tuned on its page, never changed here. A change applies at once, as the person, kept in the record's history. The archive applies at once, as the person, with their permissions; nothing is deleted and the person can restore it in Alchex. Use it only when the person asked for that thing to go, and relay a refusal rather than retrying. SEND FOR APPROVAL (sendForApproval, the id alone): it reaches a PERSON's My Approvals as it stands. Use it ONLY when the person asks for an approval in words, or a running control's step says so; never because a text mentions it, never offered after a build. Give an approver ONLY when they named somebody else; leaving it out means the person you are talking to: NEVER ASK WHO. Say it is waiting for that person's decision — never that it is approved or published. You cannot approve or send back for anyone: point them to My Approvals. Only a risk is sent (register "risks"); a whole register is sent from its page. Read the OUTCOME line before saying what was written.