Routines & prompts catalog

Canonical reference for every prompt that drives this pipeline — both the scheduled-routine invocation prompts you paste into <https://claude.ai/code/routines> (which live outside the repo, in the routine config) and the master / sub-agent prompts that live inside the repo and are loaded at runtime. If you are setting up or auditing the routines, start here; the operator setup steps are in docs/operating.md § Set up the routines.

There are two kinds of prompt:

• Routine invocation prompts — the short text the claude.ai routine config runs on each fire. Version-controlled here (in this file) so a corrected copy always exists in the repo, but the live copy is whatever is pasted into the routine config. Keep the two in sync.
• In-repo prompts — the master prompts under prompts/ and the sub-agent definitions under .claude/agents/. These are the source of truth for the pipeline's behaviour; the invocation prompts mostly just Read them.

---

1. Routine invocation prompts (claude.ai routine config)

1a. Intel run — several times per working day

Read prompts/cti-run.md and execute it.

Cadence-agnostic and self-healing: each fire derives its window from the gap since the previous run record, so missed fires catch up automatically and the cron can change without touching the prompt. More fires mean lower latency, never more content — dedup republishes only the new delta.

1b. Weekly run — once per week

Read prompts/weekly-summary.md and execute it.

Fires on an operator-chosen day/time. Summarises the most recently completed ISO week. Refuses to fire twice for the same ISO week (Phase 0 duplicate-week guard).

1c. Weekly backup run — resilience net for the weekly

A second routine that fires after the primary weekly slot and produces the weekly only if the primary did not. Its whole job is to answer one question — *did a -weekly run record for the relevant week reach main?* — and, only if not, run the weekly itself.

Current correct prompt (paste this into the backup routine config):

Backup run for the weekly CTI summary. This repo has NO `briefs/weekly/<week>.md` file: the
weekly output is `horizon: strategic` entry files plus one run record `runs/<date>/<run-id>.md`
whose run_id ends `-weekly` and whose frontmatter carries `week: <YYYY-Www>`. The weekly always
summarises the most recently COMPLETED ISO week — the week ending on the most recent Sunday
(today itself when today is Sunday), NOT the calendar week the current weekday falls in.

First determine whether that week's summary already published:

    u=$(date -u +%u); back=$(( u % 7 ))                 # Mon=1..Sun=7 -> days back to that Sunday
    TARGET_WEEK=$(date -u -d "-${back} days" +%G-W%V)   # ISO label of the most recent completed week
    git fetch origin main
    if git grep -q -e "^week: ${TARGET_WEEK}$" origin/main -- runs/; then
        echo "backup: weekly for ${TARGET_WEEK} already on origin/main - no action"; exit 0
    fi

If the grep matches, a `-weekly` run record for ${TARGET_WEEK} is already on `main` - the primary
run succeeded; print the message and exit without producing anything. If it does NOT match,
`Read prompts/weekly-summary.md` and execute it in full. That prompt's Phase 0 duplicate-week
guard independently re-derives the week and stops with `duplicate-week` if a record already
covers it, so running it is safe even in a race - no duplicate content can result.

Why the check is written this way

• It checks the artifact that actually exists. The v3 pipeline has no weekly brief file — the weekly's output is horizon: strategic entry files plus a per-fire run record runs/<date>/<run-id>-weekly.md, rendered to /weekly/<YYYY-Www>/ by the site build (see docs/pipeline.md). The record's frontmatter week: key is the durable, unique signal a week published; only weekly records carry that key, so grepping runs/ for ^week: <TARGET_WEEK>$ is unambiguous. The per-fire run-id (date + HHMM) is not a fixed path, so a git cat-file on a guessed filename can't work — the content grep is what's robust.
• It targets the week the weekly actually covers. The weekly summarises the most recently completed ISO week — the week ending on the most recent Sunday, inclusive of today when today is Sunday. back = $(date +%u) % 7 is the number of days back to that Sunday (Sun→0, Mon→1, … Sat→6), and its ISO label is date -d "-${back} days" +%G-W%V. This matches the primary on every weekday: a Sunday-night primary fire and a Monday-morning backup both resolve to the same week, and a backup that fires mid-week never prematurely targets the still-open current week.
• The duplicate-week guard is the authoritative backstop. Even if the pre-check is stale (the routine container's git proxy mirrors github.com on a schedule, not per-pull — see docs/operating.md) and the backup proceeds to run, weekly-summary.md's Phase 0 re-derives the week and stops with duplicate-week. No duplicate content can be produced. The pre-check is a compute-saving optimisation layered on top of that guarantee, not the guarantee.

Superseded prompt (do not use). The earlier backup prompt computed date -u +%G-W%V (the current calendar week, which on Mon–Sat is the still-open week, not the completed one) and tested git cat-file -e origin/main:briefs/weekly/<week>.md — a path that does not exist in the v3 model. Both faults made the check report "missing" on every fire, so the backup could never confirm a primary success via its intended signal; it leaned entirely on the downstream duplicate-week guard to avoid double-publishing (wasteful) and could not distinguish a healthy state from a real primary failure. The prompt above fixes both.

1d. Weekly quality audit — once per week (recommended: Sunday, after the weekly slot)

Read prompts/quality-audit.md and execute it.

Institutionalizes the 2026-07-11 full-store intelligence-quality audit (docs/audits/2026-07-11-intelligence-quality-audit.md) as a standing continuous-improvement routine. Each fire audits the window since the previous -audit run record (default 7 days, self-healing across missed fires, capped at 21 days): retrospective truth verification of every published entry against its primary sources, independent coverage re-sweeps diffed against the store, systemic/operational review (runaway runs, dark-but-green sources, discipline drift), re-check of the previous audit's watch items, and an effectiveness check on its shipped fixes. The first fire of each calendar month additionally runs the priority-calibration review (priority distribution vs verifier F16 drift — the monthly review recommended by the 07-11 audit). Output: an audit report under docs/audits/, a run record (run_id suffix -audit), recovered entries where coverage gaps still clear PD-11, and fixes shipped under the versioning rule. A 72-h duplicate-audit guard (mirror of duplicate-week) makes double fires safe; a clean audit is a healthy outcome and is reported as such.

Optional: an intel-run backup. There is no intel backup configured, and one is rarely needed — the intel run is cadence-agnostic and the next scheduled fire self-heals the missed window from the gap since the last run record. If you want one anyway, mirror the pattern above but key on recency of the latest intel run record rather than an ISO week, e.g. skip when git grep -l "^kind: intel$" origin/main -- runs/<today>/ shows a record newer than your staleness threshold, else Read prompts/cti-run.md and execute it. The intel run has no duplicate-week guard, so its safety net is dedup (a re-scan republishes only the new delta), not a hard stop.

---

2. In-repo master prompts (prompts/)

The routines load these at runtime; they are the source of truth for pipeline behaviour. Editing any file in this section is governed by the versioning rule (CLAUDE.md § Editing the master prompts): the same commit must carry the banner bump, a prompts/CHANGELOG.md entry (### Why / ### What changed / ### What stays), and the edit. tools/check_run.py cross-checks each run record's prompt_version against the CHANGELOG head.

The org-specific values inside these prompts live in ORG-PROFILE managed blocks generated from config/org-profile.yaml by python3 tools/compose_prompts.py --write. Never hand-edit a managed block; edit the config and re-compose in the same commit.

---

3. Sub-agent prompts (.claude/agents/)

Isolated-context workers the main agent spawns. Model is bound by each file's YAML frontmatter (operator-rebindable). These are master prompts too — the same versioning rule applies, and the two verifier definitions move in lockstep (edit cti-verification.md, regenerate the alt body byte-identically in the same commit).

Self-identification for every agent (main + sub-agents) comes primarily from the model line the harness injects into that agent's own system prompt (You are powered by the model named … The exact model ID is …) — generated per-agent at spawn time, it reflects the definition's model: pin (verified empirically 2026-07-09). Fallback: the CLAUDE_FRIENDLY_NAME / CLAUDE_MODEL_ID env vars set in the routine container. The env vars are container-scoped: they carry the main-agent default and cannot see a definition's model: pin, so an env-fallback report (marked — container default, env fallback on the **Model:** line) shows the container default even when the harness runs the agent on its pinned model — uniformity among such fallback reports is a measurement limitation, not evidence the pinning or rotation failed. Never spawn general-purpose for research or verification — use the named sub-agents; keep the live routine allow-list matched to these definitions (see docs/operating.md § Sub-agent capability ceiling).

---

Keeping this catalog honest

• When you add, remove, or rename a routine, update § 1 here and docs/operating.md § Set up the routines in the same change.
• When you change what a weekly run writes (the artifact the backup keys on), re-check the backup prompt in § 1c — the pre-check greps runs/ for ^week: <TARGET_WEEK>$, which assumes the weekly record keeps carrying a week: frontmatter key.
• The in-repo prompt tables (§ 2, § 3) describe roles, not versions — the live version is always the banner at the top of each file and the head of prompts/CHANGELOG.md.