dossier-init

Scaffolds a client-dossier layout into the current repository: dossier.yaml, the four top-level folders (sources/, dossier/, knowledge/, outbound/) with their documented subfolders, the seven living files under dossier/, the knowledge/CONTEXT.md and knowledge/DOCTRINE.md stubs, one README per top-level folder, and a pointer to the protocol in the repository's instructions file. On a repository that already holds the same material under other names, it proposes — and, once confirmed, carries out — the git mv renames to the fixed English layout.

The dossier root is the directory the user is already in: this skill never asks where to scaffold, only what to scaffold with. Every input below is either already given (an argument, or something stated plainly in the user's sentence) or collected through exactly one grouped question — nothing here is asked twice. Re-running this skill on a dossier that already exists completes only what is missing; it never overwrites a living file. Nothing this skill produces is ever committed — the human reviews the working tree and commits it.

What it produces

• dossier.yaml — the ten configuration keys, in this order: version, language, counterparts, confidential_names, consent_recorded, consent_source, timestamps, date_format, status_header, staleness_days.
• The four top-level folders, with a .gitkeep in every leaf folder the scaffold leaves empty, so an empty folder still survives a clone.
• The seven living files under dossier/, each opening with a Last updated: line and its "Updated when" rule.
• knowledge/CONTEXT.md, knowledge/DOCTRINE.md, and one README under each of dossier/, sources/ and outbound/.
• A protocol-pointer section in AGENTS.md (always) and in CLAUDE.md (only when that file already exists) — see step 6.
• On an existing repository: the git mv renames from step 4, executed only after the human confirms them.

Step 1 — Infer the language

Take the language of the conversation as the first guess. If the repository already has a CLAUDE.md or an AGENTS.md, read its language too. Propose the result as the default answer in the grouped question below — never scaffold silently in English when the conversation and the repository both point elsewhere. The user confirms the proposed language or names another one in their answer.

Step 2 — Ask one grouped question, only for what is missing

Anything already given — stated plainly in the user's sentence or an argument they passed (a counterpart's name, their role, a contact, the language, a confidential name to add) — is used as is and never asked again. When nothing at all is specified ("set up a dossier for this client"), ask exactly one grouped question: who the counterpart(s) are, with their role, and which language to use (with the language inferred in step 1 proposed as the default). Everything else — the confidential-name list, the date and timestamp formats, whether a status header is required, staleness_days — is derived in step 3 or takes the spec's default; never hand the user a scaffold that is only placeholders.

Step 3 — Derive the confidential names and show them

Build confidential_names from the counterpart names collected in step 2, plus their short forms — an organisation given as "Acme Consulting" also answers to "Acme". Show the resulting list to the user and let them add or remove entries before anything is written; this is the list a linter will one day refuse to let leak into outbound/.

Step 4 — Propose adoption renames, confirm, then execute

Quoting rule, for every value in every call of this skill (--target, --counterpart, --confidential-name, --consent-source, and the rest): put the value between single quotes, and write each single quote inside it as '\'' (close the quote, an escaped quote, reopen). L'Oréal becomes 'L'\''Oréal', so a counterpart is 'L'\''Oréal|CEO|a@b.fr' and a consent source is 'accord de l'\''intéressé'. Never leave a bare ' inside a value: the shell would end the quote early, split the call at the next | or ;, and the script would receive a mangled name (LOréal) that silently protects nothing. A value holding a line break is refused: ask for it on one line. Double quotes would let $(...) and backticks run.

Check the bundled script is there before you use it:

test -x "${CLAUDE_SKILL_DIR}/scripts/init.py"

If that check fails, the rename proposal cannot be computed here: say so, skip to step 5, and in step 6 follow "If the bundled script is not available". Otherwise run:

"${CLAUDE_SKILL_DIR}/scripts/init.py" --target '<dossier root>' --propose-renames

against the repository the user is in. Show the proposal exactly as it is printed:

• rename: lines are an ordered script — each one is expressed against the state the repository would be in after every rename printed before it, so applying them top to bottom is always correct.
• untracked: lines mean git does not track that path — moving it with plain mv would lose its history; ask about each one separately before touching it.
• unsafe: lines are never renamed, no matter what the user says.
• candidate: lines are guesses — a folder that looks like a dossier under an unlisted name — and are never renamed without the user confirming that specific one.
• ignored: lines are folders with no dossier role, left alone on purpose.

Ask for one explicit confirmation before renaming anything. Without it, rename nothing and move on to step 5 as if step 4 had found nothing to do. With it, apply every rename: line top to bottom, one at a time:

git mv -- <old> <new>

using the two paths exactly as the proposal printed them — they are already shell-quoted, so never re-quote them or reassemble the command from parts. For each untracked: line, ask separately, and use plain mv only if the user accepts that history will not be preserved. Whatever was or was not renamed, report it and never commit any of it: the review and the commit are always the human's own next step.

Step 5 — Choose the template variant

Two languages ship a hand-maintained template set: en and fr. When the confirmed language is one of them, nothing further is needed. For any other language, pass --template-variant en explicitly and translate the result yourself once the bundled script has written it: the titles, the headings, the "Updated when" comments and the example lines move into the target language, but every path and every protocol keyword (the Last updated: label, the status-header field names and values) stay exactly as the script wrote them.

Step 6 — Run the script, or do its job by hand

Quote every value as in the quoting rule of step 4: single quotes, each inner ' written '\''.

Run the bundled script with every value collected above spelled out as its own flag, so a reader can reproduce the exact call. If the test -x check of step 4 failed, the next subsection applies instead:

"${CLAUDE_SKILL_DIR}/scripts/init.py" \
  --target '<dossier root>' \
  --language '<code>' \
  --template-variant '<code, only if step 5 forced one>' \
  --counterpart '<Name>|<role>|<contact>;<contact>' \
  --confidential-name '<name>' \
  --consent-recorded '<true|false>' \
  --consent-source '<text>' \
  --staleness-days '<n>'

(repeat --counterpart once per counterpart and --confidential-name once per name collected in step 3).

If the bundled script is not available

A bare-skills install, or a harness that cannot run Python, will not have it. Say so plainly — the bundled script is not available here, so this is doing its job by hand — and then perform the identical procedure, stated completely enough to reproduce it without the script:

• Create the four top-level folders — sources/ (with received/ and transcripts/, and raw/ inside transcripts/), dossier/, knowledge/ (with decisions/ and watch/), and outbound/ (with deliverables/, mails/ and calls/) — for any of them that does not already exist.
• Write an empty .gitkeep file into each of these seven leaf folders, unconditionally, even one that looks empty already: sources/received, sources/transcripts/raw, knowledge/decisions, knowledge/watch, outbound/deliverables, outbound/mails, outbound/calls.
• Write dossier.yaml with its ten keys, in this exact order: version (1), language, counterparts (one block per counterpart, with name, and role/contacts only when given), confidential_names, consent_recorded, consent_source, timestamps, date_format, status_header, staleness_days.
• Create the twelve files below, each opened with today's date on its Last updated: line (the two knowledge stubs and the three READMEs carry no such line), from the chosen template variant, translating as step 5 says when the variant is not hand-maintained:
  • dossier/ACTORS.md, dossier/HISTORY.md, dossier/LEARNINGS.md, dossier/TONE.md, dossier/CHANNELS.md, dossier/OPEN-QUESTIONS.md, dossier/RELATIONSHIP.md
  • knowledge/CONTEXT.md, knowledge/DOCTRINE.md
  • dossier/README.md, sources/README.md, outbound/README.md
• Leave any path that already exists exactly as it is — never truncate, overwrite or merge into an existing file; report it as kept rather than created.
• Add the protocol-pointer section, delimited by its start/end marker comments, to AGENTS.md — creating the file if it is absent, appending the section if the file exists without it, and replacing the span in place if the markers are already there — and do the same to CLAUDE.md, but only when CLAUDE.md already exists; never create CLAUDE.md.

Step 7 — Report

Say what was created, what was already there and kept untouched, what was renamed in step 4, and what was left ignored — and say once more, plainly, that none of it was committed.

Rules

• Idempotent, never destructive. Re-running this skill on an already-scaffolded repository completes only what is missing; an existing file is never truncated, overwritten or merged.
• Confirm before renaming. No git mv runs until the user has seen the step-4 proposal and given one explicit confirmation; without it, nothing is renamed.
• Never commits. Whatever this skill creates, renames or completes, the human reviews the working tree and commits it — this skill never runs git commit.
• English artefacts, localised content. The bundled script, this skill and every path they produce are English; only the dossier's own content — titles, headings, comments, example lines — follows the language steps 1 and 2 settle on.