Context
/planning:draft-goal-condition mechanizes goal drafting well but assumes a checkable condition exists. Most real goals are not npm test. The Anthropic loops webinar's technique for the rest [24:45]: impose a structural constraint (length, section count), enumerate required contents, and add a self-verification sub-step that is itself checkable ("write a report where you have verified that all facts are correct — do an extra lookup to make sure any citations link to a real website whose content supports the citation"). Source analysis (item 5, Tier 2, CORROBORATED by companion blog): https://github.com/melodic-software/knowledge-corpus/blob/main/sources/webinars/getting-started-with-loops/analysis/action-items.md.
Verified /goal mechanics relevant to the branch (2026-07-26, https://code.claude.com/docs/en/goal#how-evaluation-works): the evaluator is the configured small fast model (defaults to Haiku), it fires as a session-scoped prompt-based Stop hook each time Claude tries to stop, and it does not call tools — it can only judge what is already surfaced in the transcript. Conditions cap at 4,000 characters. This directly shapes the technique: the self-verification sub-step must leave transcript-visible evidence for the evaluator to judge.
Proposed work
- Add a non-quantifiable-goal branch to the skill: structural constraint + enumerated contents + transcript-demonstrable self-verification sub-step
- For the co-drafting technique (interview Claude about the goal before writing it): point to
/planning:interview rather than adding new prose
Acceptance criteria
References
Context
/planning:draft-goal-conditionmechanizes goal drafting well but assumes a checkable condition exists. Most real goals are notnpm test. The Anthropic loops webinar's technique for the rest [24:45]: impose a structural constraint (length, section count), enumerate required contents, and add a self-verification sub-step that is itself checkable ("write a report where you have verified that all facts are correct — do an extra lookup to make sure any citations link to a real website whose content supports the citation"). Source analysis (item 5, Tier 2, CORROBORATED by companion blog): https://github.com/melodic-software/knowledge-corpus/blob/main/sources/webinars/getting-started-with-loops/analysis/action-items.md.Verified
/goalmechanics relevant to the branch (2026-07-26, https://code.claude.com/docs/en/goal#how-evaluation-works): the evaluator is the configured small fast model (defaults to Haiku), it fires as a session-scoped prompt-based Stop hook each time Claude tries to stop, and it does not call tools — it can only judge what is already surfaced in the transcript. Conditions cap at 4,000 characters. This directly shapes the technique: the self-verification sub-step must leave transcript-visible evidence for the evaluator to judge.Proposed work
/planning:interviewrather than adding new proseAcceptance criteria
/planning:interview, not duplicated proseReferences
plugins/planning/skills/draft-goal-condition/SKILL.md