Agent skills

Status: draft

Proposals by Claude – not binding until Daniel approves them.

Proposal: Two skills in the Agent Skills format, shipped in the repository PROP_SKILLS_FORMAT

status: draft refines: DEC_SKILLS refined by: PROP_ROLLOUT_MIGRATE, Q_SKILL_INSTALL

skills/deciwa-work/SKILL.md and skills/deciwa-ask/SKILL.md, each a short instruction plus the CLI calls it uses. The skills only call the CLI with --json – never the files – and sign as agent:<name>.

deciwa-work

1. deciwa todo --json – the work list.

2. Per job: read the element and apply the comment, fix a rejected statement, or delete the element.

3. Never edit an element in state frozen; if a job collides with a freeze, skip it and say so.

4. deciwa resolve <id> --note "<what changed>".

5. deciwa request … --to <author> --question "Changed – OK?" for every element it touched, so the human sees the change.

deciwa-ask

When the agent writes something it is not sure about – an inferred wording, a number without a source – it writes the text as it thinks is right and puts a request on that element instead of a [CHECK] marker:

$ deciwa request 20_facts/company/founder.md --quote "since October 2026" \
    --to daniel --question "Is October 2026 the start date of the advisor role?"

Question: How do the skills get into a project? Q_SKILL_INSTALL

status: open to: Daniel refines: PROP_SKILLS_FORMAT

Options: (a) deciwa skills install copies them into .claude/skills/ of the current project; (b) a Claude Code plugin that bundles both skills; (c) both. Proposal: (a) for 0.1, (b) once the CLI is published.

Question: Are there quick actions besides delete? Q_QUICK_ACTIONS

status: open to: Daniel refines: DEC_CARD_VERDICTS

Rephrase and more details are now said in the comment (The feedback card has four ... (DEC_CARD_VERDICTS)). Candidates for extra one-click jobs: shorten, split, add source, translate. None, a fixed set, or configurable in .deciwa/config.toml?