| name | weekly-synthesis |
| description | Produce a weekly team digest (decisions, scope moves, task deltas, risks) with
a rubric-gated self-correction loop: an independent verifier grades the draft
against .claude/rubrics/weekly-synthesis.md and the harness revises until the
rubric passes or its budget is spent. Use at sprint boundaries or when someone
asks "what happened this week?". Consumes decision-audit and scope-creep outputs
when present — do not re-derive their findings; transcript extraction stays with
transcript-decisions.
|
| version | 1.0.0 |
| kind | workflow-harness |
| workflow | weekly-synthesis.workflow.js |
| triggers | ["weekly synthesis","weekly digest","what happened this week","sprint summary","synthesize the week"] |
weekly-synthesis
The flagship self-correction loop harness (skills conventions #11–13):
Collect → Draft → Grade (independent verifier, fresh context, grades against
the rubric using the Read tool on sources) → Correct (revise only the failing
criteria) → loop until pass or budget exhausted.
Before running
Read .claude/memory/instincts.md (the harness passes distilled rules to both
the drafter and the grader). If the run fails its rubric after budget
exhaustion, record an incident in .claude/memory/incidents/ per
.claude/memory/README.md — that is how failures become rules.
How to run
Workflow({
scriptPath: ".claude/skills/weekly-synthesis/weekly-synthesis.workflow.js",
args: {
repoPath: "<absolute path to this team-ops repo>",
windowStart: "2026-06-03", // optional; default: 7 days ending windowEnd
windowEnd: "2026-06-10",
rubricPath: ".claude/rubrics/weekly-synthesis.md", // optional; default
budget: 3 // optional; default from rubric frontmatter
}
})
Returns { passed, loops_used, budget, grade_report, digest_markdown }.
A failed gate is a result, not an exception — the digest and its grade
report are returned either way. Review the grade report before publishing a
digest that didn't pass; never silently drop the report.
Design notes
- Independent grading — the verifier gets the rubric, the candidate, and
source paths; never the drafter's reasoning. Self-critique inside one context
is what this design exists to avoid.
- Correct, don't regenerate — revision preserves passing content; a full
regenerate can trade one failure for new ones.
- Read-once — sources are read in Collect and passed as excerpts; the
grader re-reads only what grounding checks require.
