Announce Release To Customers

Announce Release to Customers

Describe a release that already happened, in words a customer can act on, with each item traceable both backwards to the commit that produced it and forwards to the customer signal it answers.

Two failure modes shape this skill. A release note written from a model's impression of a codebase announces work nobody shipped. A release note written only from commit subjects tells customers about refactors they cannot see. The change list is therefore an input, and the translation is the work.

Inputs

Require:

• the release reference: a tag, a version, or a commit range;
• a change list obtained deterministically, each item carrying its own reference — a commit hash, a merge request, or a CHANGELOG heading;
• the artifact directory holding signals.jsonl and decisions.jsonl, when a run exists to link against;
• the intended audience and channel, as a statement of intent only.

Obtain the change list by one of three means and record which one was used:

git log --no-merges --pretty=format:'%h %s' <previous-tag>..<release-tag>

reading the release section of CHANGELOG.md, or accepting a list the person provides. There is no fourth means. If none is available, stop and say so rather than reconstructing the release from the working tree.

Workflow

1. Record the release reference and how the change list was obtained.
2. Copy each change into internal_summary verbatim, keeping the team's own words as the source text a reviewer can compare the translation against.
3. Translate each change into customer language: what the customer can now do, or what stopped going wrong. Do not translate a change whose effect on the customer cannot be stated — mark it internal-only instead and leave it out of the customer summary.
4. Read signals.jsonl and decisions.jsonl and propose, per change, the signal or decision it answers.
5. Type every link with link_state. Use human_confirmed only when the linked decision or signal itself names this change, or a person confirms the link in a later user turn. Use model_inferred when the link comes from reading both records and judging them related. Use none when the change answers no recorded signal, and list it in unlinked-changes.md.
6. Write the release-level internal_summary from the change list and the release-level customer_summary as its translation.
7. Keep status at draft or awaiting_human_review. A person moves it to approved in a later user turn.
8. Leave publication_action_id null unless a person has recorded a publishing action. Sending is an external write and belongs to that action's approval, not to this note.
9. Append the note. Correct a published note by appending a new one that supersedes it, never by editing the earlier record.

Boundaries

• This skill does not invent changes and does not send anything. Every entry comes from a git log, a CHANGELOG, or a person, and the draft stays a file.
• Let the model translate developer language into customer language and propose which signal a change answers.
• Use deterministic checks for the change list, IDs, reference resolution, status transitions, and append-only history.
• Require a person to approve the wording, and require the existing external write boundary for any send.
• Do not treat a model_inferred link as evidence that the release answered a customer need. It is a hypothesis for a reviewer to accept or reject.
• Do not describe a change as an outcome. Shipping a change is not evidence that it worked; that judgement belongs to the metric and outcome records.
• Do not restate an unreleased roadmap item, a planned feature, or a decision that has not shipped.
• Do not write pricing, refund, contract, or other sensitive-topic wording from scratch. policies/default-policy.json restricts those topics to pre-approved templates, so route them to a person instead.

Outputs

Create or append:

• release-notes.jsonl
• customer-announcement-draft.md
• unlinked-changes.md

Read references/output-contract.md before writing them.

Add every new record to release-notes.jsonl with the append-record command, never by writing or editing the file:

python3 scripts/stg.py append-record <artifact-directory>/release-notes.jsonl '<json-object>'

It takes an exclusive lock and hash-chains each line to the one before it, so a later reader can tell whether an announcement was rewritten after the fact. In a packaged runtime the same command is signal-to-growth append-record.

Stop conditions

Stop when no deterministic change list is available, when a change carries no reference, when a linked signal or decision ID does not resolve, when the customer effect of a change cannot be stated without guessing, or when the request is to send rather than to draft.

Verification

Keep signals.jsonl, decisions.jsonl, and release-notes.jsonl together, then run:

python3 scripts/stg.py validate-artifacts artifacts/

Report the change count, how many changes carry a human_confirmed link, how many carry a model_inferred link, and how many carry none. A note in which every change is linked deserves more suspicion than one that reports gaps.