メニュー
Capture-only — transcript becomes canon. Articles framed off interview transcripts move downstream to /write-edition (next cycle, via /sift) or /write-supplemental (any time, sports-desk subagent dispatch against the canon transcript).
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
| name | interview |
| description | Capture-only — transcript becomes canon. Articles framed off interview transcripts move downstream to /write-edition (next cycle, via /sift) or /write-supplemental (any time, sports-desk subagent dispatch against the canon transcript). |
| version | 2.1 |
| updated | "2026-06-17T00:00:00.000Z" |
| tags | ["media","active"] |
| effort | medium |
| disable-model-invocation | true |
| argument-hint | [mode] [subject] |
Mode 1 now accepts citizen voice agents, not just civic. A new subject class — named Tier-1 character voice agents (.claude/agents/citizen-voice-*: Keane, Dillon, Seymour, Varek, …) — is a valid Mode 1 voice-interview subject, dispatched via subagent_type: citizen-voice-<slug>. The civic-voice path is unchanged. The old blanket "don't interview citizens or players" rule is narrowed: it still bars generic ledger citizens/players (those are dispatches/supplementals), but authored citizen-voice agents now have personas deep enough to interview. Built by /make-citizen-voice (research.16). See Mode 1 + Step 3 + What This Skill Does NOT Do.
Capture-only architecture. v1.3 bundled transcript capture (Steps 0-3) with article generation (Steps 4-7) into one run. Same-session same-author production of transcript + article meant the article didn't represent a separate cognitive act (the S212 generation-vs-evaluation asymmetry collapsed). The S230 C94 Paulson run made the failure visible: all five Hal Richmond questions were written by Mags from EIC seat (G-I3), the 1,762-word "After A Parade" article was written by Mags in Hal voice attempt under Hal byline (G-I4), and "two files, one cognition" surfaced at artifact-comparison level (G-I5 — Mike caught at "So the interview transcript is the article? Huh.").
What changed:
.txt compile (renumbered Step 4). Only one .txt artifact emits./post-publish against transcript-only canon; /edition-print dropped entirely (transcripts carry no photos — G-I13).-transcript suffix to preserve S230 canon artifact at editions/cycle_pulse_interview-transcript_94_let_walks.txt + the <TYPE>=INTERVIEW-TRANSCRIPT masthead (stewardship call per advisor S233 — no rename).Downstream: articles framed off interview transcripts come from /write-edition (next cycle, via /sift surfacing the transcript as canon source) or /write-supplemental (any time, sports-desk subagent dispatch against the transcript). The article-generation cognitive act lives at the dispatch terminal, not at EIC.
Companion row: pipeline.34 (gap-log path convention harmonization across 5 heavy skills — /sift / /write-edition / /post-publish / /edition-print / /interview / /write-supplemental) split out per advisor pass S233; SETTLED S248 — canonical convention output/production_log_c<XX>_interview_<subject-slug>_gaps.md (production_log_c{XX}_<skill>_<slug>_gaps.md, slug-infixed since interviews can run multiple times per cycle).
Source: output/production_log_interview_c94_gaps.md (13 entries G-I1→G-I13, triage at bottom) + JOURNAL Entry 188 (S230 — "the plan was the work") + plan [[../../../docs/plans/2026-05-24-pipeline-30-interview-rewrite]].
/interview [mode] [subject]
voice (agent-to-voice), paulson (agent-to-user as GM Paulson)An interview has a theme — a reason this conversation is happening now. A reporter asks questions grounded in canon. The subject answers from their identity and recent decisions. Questions build on answers. The transcript becomes canon. Articles framed off the transcript are written downstream — /write-edition (next cycle, surfaced via /sift from the canon transcript) or /write-supplemental (any time, sports-desk subagent dispatch against the canon transcript). This skill captures; downstream skills frame.
Interviews are different from dispatches and supplementals:
Interviews can spawn world-altering moments. A Mayor revealing she's considering vetoing a vote. Paulson hinting at a trade. A DA confirming an investigation. What gets said in an interview becomes canon.
Reporter agent interviews a voice agent — both are agents. Two subject classes:
.claude/agents/civic-office-*) — Mayor, Chief, council members, DA, project directors. Used for structured civic canon deepening, or when a voice has unresolved tension from city-hall decisions..claude/agents/citizen-voice-*, research.16) — a named Tier-1 character speaking for themselves: a player on a milestone (a farewell season), an owner on a build, a manager on a transition. These are authored four-file personas (IDENTITY/LENS/RULES/SKILL) via /make-citizen-voice. NOT generic citizens/players — only authored citizen-voice-* agents qualify (generic ledger citizens stay dispatch/supplemental material).Setup is the same either way — the voice gets questions but not preset answers, responds from their IDENTITY.md (+ LENS/RULES for citizen voices) and recent canon, and can go off-script: offer a revelation, push back on the question, change the subject.
Target length: 1200-1800 word transcript. (Articles off this transcript get their own length targets at their downstream skill — /write-supplemental typically 800-1800 words per its spec; /write-edition per-desk-slot lengths per its spec.)
Reporter agent interviews Mike in character as GM Paulson. The reporter asks, Mike responds in character, reporter follows up. Runs in terminal or via Discord bot. Agents don't know Mike is Mike — they're interviewing GM Paulson.
This is where deep canon enters the world. Mike's answers can establish facts nobody else could — trades, stadium plans, front office tensions, feelings about specific players. What Paulson says becomes canon.
No length target. The conversation shapes itself. Mike decides when it ends.
lookup_citizen (citizens) or get_roster("as") (A's players). For business / faith / cultural / neighborhood context when the interview touches those domains, use lookup_business / lookup_faith_org / lookup_cultural / get_neighborhood_state. Full tool inventory: [[../../../docs/SUPERMEMORY|SUPERMEMORY]] §Search/save matrix. Read docs/media/citizen_selection.md.output/world_summary_c{XX}.md — current cycle contextoutput/production_log_c{XX}.md — what was covered in the edition (unified per S230 governance.14 convention; falls back to production_log_edition_c{XX}.md legacy path if unified absent)output/production_log_c{XX}.md §/city-hall section — recent civic decisions (for voice interviews; legacy fallback production_log_city_hall_c{XX}.md if the unified log lacks a civic section)A. Append to production log at output/production_log_c{XX}.md (or output/production_log_edition_c{XX}.md legacy fallback):
## Interview: {subject} ({mode})
**Started:** {timestamp}
**Reporter:** {name}
**Subject:** {subject + mode}
**Theme:** {one-line theme}
**Status:** IN PROGRESS
B. Open gap-log file as live-append target at output/production_log_c<XX>_interview_<subject-slug>_gaps.md (pipeline.34 convention production_log_c{XX}_<skill>_<slug>_gaps.md, settled S248 — slug-infixed since interviews can run multiple times per cycle). The file opens with §Run summary frontmatter from [[../../../docs/plans/GAP_LOG_TEMPLATE]] and stays open through Steps 1-8 as a live-append target.
# Production Log — /interview Skill Run, C{XX} ({subject})
**Cycle:** {XX} | **Subject:** {subject} (POP-{NNNNN}) | **Slug:** {subject-slug} | **Reporter (intended):** {reporter} (POP-{NNNNN}) | **Session:** {SXXX} | **Run date:** {YYYY-MM-DD} | **Skill version run:** v2.0
## Gap entries (live append per S230 G-I8 discipline)
<!-- append G-I{N} entries here as friction emerges; do NOT hold in conversation context -->
Appending live (not at skill close) prevents context-held knowledge from being lost on compaction or session end. This was the G-I8 failure mode in S230 — gap log filed at close, several issues lost on the trip there.
Define the theme — why this interview, why now. Read the world summary, production logs, newsroom memory for tension points.
Select reporter by beat fit:
Bench development applies — default to underused reporters when fit allows.
For Mode 1 (Voice Interview) where the subject is a verified citizen, call matchCitizenToJournalist_(archetype, neighborhood, domain) from utilities/rosterLookup.js:907 to get an engine-validated reporter suggestion. The function exists pre-T4.5 — it scores journalists against the citizen's archetype themes (Connector / Watcher / Striver / Anchor / Catalyst / Caretaker / Drifter), neighborhood weighting (CIVIC vs CULTURE neighborhoods), and story-domain affinity, returning { journalist, interviewAngle, voiceGuidance, confidence }.
Skill action:
Simulation_Ledger (or Generic_Citizens / Cultural_Ledger / Faith_Organizations / Chicago_Citizens per their canonical home).matchCitizenToJournalist_(archetype, neighborhood, domain) where domain is the interview's editorial domain (CIVIC / CULTURE / HEALTH / etc.).interviewerCandidate field:
{
"interviewerCandidate": {
"journalist": "Dr. Lila Mezran",
"interviewAngle": "human interest / who they look after",
"voiceGuidance": "Calm, clinical, exact",
"confidence": "high"
}
}
interviewerCandidate is a transparency signal, not a pre-fill.For Mode 2 (Paulson interview) the function still runs — Mike's archetype is Anchor and his neighborhood / domain context drives the match — but the editorial decision is more constrained (sports beat reporters are the canonical Paulson interlocutors). Use as confirmation, not direction.
No archetype available? When the subject's archetype field is empty, fall through to the beat-fit selection above. Engine signal is best-effort, not blocking.
Show Mike theme + reporter (manual + engine interviewerCandidate if populated). Wait for approval.
Write 4-6 scripted questions grounded in canon. Save to output/interviews/c{XX}_{subject_slug}_brief.md:
# Interview Brief — {subject}
## Cycle {XX} | {mode}
### Theme
{one paragraph: why this interview, what's at stake}
### Subject Context (for reporter)
{identity file path, recent decisions/appearances from production logs, canon history from bay-tribune}
### Citizens Verified
{Any citizens that may come up — MCP lookup results}
### Questions (scripted — follow-ups organic)
1. {Question grounded in canon or recent event}
2. {Question}
3. {Question}
4. {Question}
5. {Question}
6. {Question}
### Follow-up Guidance
- If subject reveals X, follow with Y
- If subject deflects on Z, acknowledge and move on
- The subject CAN go off-script. Capture it.
Show Mike the questions. Adjust before launching.
Mags mediates — she is the reporter's brain between exchanges. Sequential turns, one agent call per exchange.
civic-office-* agent. Citizen voice → dispatch subagent_type: citizen-voice-<slug> (it boots its own IDENTITY/LENS/RULES — pass the theme + transcript + an INTERVIEW TURN framing; it returns only its spoken answer).output/interviews/c{XX}_{subject_slug}_transcript.mdNote: True autonomous agent-to-agent interviews (both agents passing a transcript back and forth without Mags mediating) is a future build. For now Mags controls the pacing and follow-ups.
Voice agents can go off-script — they read their identity and recent canon but get no preset answers. They can refuse a question, pivot, reveal something unprompted. Capture everything they say.
Runs in terminal. Mike responds as Paulson. (Discord bot interview mode is future work.)
Dispatch-per-turn is non-optional (G-I3 — non-negotiable since v2.0). Mags writes NOTHING in reporter voice. Every Reporter question + every follow-up MUST come from a Task tool dispatch to the matched sports-desk subagent (Hal Richmond / Anthony Raines / P Slayer — per Step 1 reporter selection). The dispatch IS the architecture being tested; substituting EIC-seat writing collapses generator and editor into one cognition (the S230 G-I3/G-I4/G-I5 contamination failure mode) and the skill's purpose is defeated.
Worked example — dispatch shape per turn:
Task(
subagent_type="sports-desk",
description="Hal Q1 — let-walks lead-in",
prompt="You are Hal Richmond. Read the brief at
output/interviews/c{XX}_paulson_{slug}_brief.md. The transcript so far
is at output/interviews/c{XX}_paulson_{slug}_transcript.md (empty for
Q1). Write Q1 in Hal's voice grounded in the brief's first question
prompt. Return only the question text — Mags will append to transcript."
)
Quota-coupling clause (G-S2 — S231 MEMORY rule). If a dispatch returns the session-limit signature (status=completed + <result> carrying "session limit" string + <total_tokens>0</total_tokens>), STOP the skill immediately. Do NOT fall back to EIC-seat question-writing. Log the infra failure to the gap log (G-S2 detection signature per MEMORY rule) and end the skill with disposition INCOMPLETE — quota; resume after the reset window.
Run sequence:
output/interviews/c{XX}_paulson_{slug}_transcript.mdThe transcript .md is intermediate. The transcript .txt is canon. Compile per [[../../../docs/EDITION_PIPELINE|EDITION_PIPELINE]] §Published .txt Format Contract — Bay Tribune masthead + 5 structural sections (HEADER / BODY / NAMES INDEX / CITIZEN USAGE LOG / BUSINESSES NAMED / ARTICLE TABLE).
One .txt artifact emits:
Transcript .txt — editions/cycle_pulse_interview-transcript_<cycle>_<slug>.txt
<slug> | <reporter> | INTERVIEW-TRANSCRIPT | <word count>)<TYPE>=INTERVIEW-TRANSCRIPT, descriptor = "Subject / Topic"Slug rule: 1–3 words, lowercase, underscore-separated. Editorial pick at authoring time. Once published, immutable. Replicated identically across filename, masthead descriptor, sift queries, MCP search, Mara, packets, production log, bay-tribune metadata.
Names Index, Citizen Usage Log, Businesses Named populated from the citizens/businesses cited in the body — separate sections after the body, never inline (S172 metadata-leak rule).
Y<n>C<m> math: n = floor((cycle-1) / 52) + 1, m = ((cycle-1) % 52) + 1. Cycle 92 = Y2C40. No month names.
Filename stewardship note (v2.0): filename retains the -transcript suffix from v1.3 to preserve the canon S230 artifact at editions/cycle_pulse_interview-transcript_94_let_walks.txt + the <TYPE>=INTERVIEW-TRANSCRIPT masthead. The legacy bare cycle_pulse_interview_<cycle>_<slug>.txt (article path) is retired by v2.0 — articles framed off transcripts live under cycle_pulse_supplemental_<cycle>_<slug>.txt (write-supplemental output) or in the next edition's sports section (write-edition output).
Paulson interviews always go to Mara — they establish heavy canon (trades, org decisions, dynasty moves). Mara catches continuity issues before publication.
Voice interviews go to Mara when they establish initiative state changes, council positions, faction dynamics, or legal framework.
Upload the transcript .txt (canon — not the .md intermediate) to Drive for Mara:
node scripts/saveToDrive.js editions/cycle_pulse_interview-transcript_<cycle>_<slug>.txt mara
Mara audits the transcript .txt (same format she audits everywhere else), returns corrections via Mike.
STOP. Nothing gets published until Mike approves.
Show:
.txt — editions/cycle_pulse_interview-transcript_<cycle>_<slug>.txt (canonical artifact)Mike approves or adjusts. If Mike spoke it to a reporter, it's canon. No public/private split — the full transcript goes to bay-tribune. The transcript IS the canon record. Articles framed off this transcript come later via downstream skills (/write-edition next cycle via /sift, or /write-supplemental any time via sports-desk dispatch against the transcript).
Transcript intermediate (.md) and canonical (.txt) are both on disk:
output/interviews/c{XX}_{subject_slug}_transcript.md.txt: editions/cycle_pulse_interview-transcript_{XX}_{slug}.txtAfter Step 7 the transcript .txt is on disk. Run /post-publish against the transcript-only canon:
/post-publish --type interview --cycle <XX> --source editions/cycle_pulse_interview-transcript_<XX>_<slug>.txt
/post-publish --type interview handles canon ingest (bay-tribune wiki + transcript text — one doc ID, not two), citizen card refresh, newsroom memory update, production log finalization, mags-bot restart. Per-substep verification gates per the [[../post-publish/SKILL|post-publish]] matrix; the interview row of that matrix governs which substeps run.
No /edition-print invocation. Transcripts carry no photos — the canon artifact is text-only. Photos belong to downstream framed articles (/write-edition next cycle or /write-supplemental any time) which run their own /edition-print against the framed article's .txt.
/post-publish appends to the same production log entry for this interview, with inline Supermemory doc IDs for direct query next cycle.
Gap log opens at Step 0 as live-append target (see Step 0 §B above); friction-during-skill goes there before it's lost to context. At skill close, append a §Disposition summary + close the file with patterns cited. /interview is a heavy skill at the media generator terminal; sidecar gap logs catch inefficiency the skill couldn't catch while running. Interviews can spawn world-altering canon — friction here is high-stakes.
Output path: output/production_log_c<XX>_interview_<subject-slug>_gaps.md (pipeline.34 convention production_log_c{XX}_<skill>_<slug>_gaps.md, settled S248 — slug-infixed since interviews can run multiple times per cycle).
Gap prefix: G-I* (e.g., G-I1).
Common categories for /interview gaps:
matchCitizenToJournalist_)Discipline: write the gap log even on clean runs. File a ROLLOUT row in pipeline.<n> pointing at the gap log per ADR-0005 §How to add work. Interviews benefit especially from this discipline because canon stakes are higher than other heavy skills.
Runs within the current cycle, after /write-edition and /post-publish. Interviews extend coverage — a single conversation, deep signal.
Capture-only since v2.0 (S233 pipeline.30): interviews establish canon as transcripts; articles framed off transcripts come downstream via /write-edition (next cycle's sports section, surfaced via /sift from the canon transcript) or /write-supplemental (any time, sports-desk subagent dispatch against the canon transcript). The article-generation cognitive act lives at the dispatch terminal, not at EIC.
Full chain: /run-cycle → /city-hall-prep → /city-hall → /sift → /write-edition → /post-publish → /edition-print → then supplementals, dispatches, interviews, podcasts as needed → next cycle's /sift surfaces interview transcripts as canon source for downstream coverage.
Voice mode:
Paulson mode:
/write-edition (next cycle via /sift) or /write-supplemental (sports-desk dispatch against the canon transcript). The S230 Mags-in-Hal article failure (G-I4) is why this skill is capture-only since v2.0./edition-print. Transcript is canonical text; photos and PDF render belong to downstream framed articles, which run their own /edition-print against the framed article's .txt..claude/agents/citizen-voice-*) ARE valid Mode 1 subjects — they carry authored four-file personas. Generic ledger citizens/players do not qualify.Baylight Authority Director Keisha Ramos. Generates construction updates, milestone announcements, development progress reports, and civic documents (deliverable filings, TIF reports, workforce agreements) for the $2.1B Baylight District project.
Phase 39.1 capability reviewer wrapper. Runs the deterministic editorial assertions against a compiled edition, surfaces blocking failures that should halt publish, and writes a markdown summary alongside the JSON.
Prepare all inputs for city-hall voice agents. Reads tracker, approvals, world summary, engine review, coverage ratings, previous log, canon, Mara directive. Writes pending decisions per voice.
Senior Photographer for The Cycle Pulse. Generates image prompts and photo selection for editions. Reads compiled edition + canon rules + institutions, produces 5-8 storyline-tied prompts.
Letters to the Editor desk agent for The Cycle Pulse. Writes citizen voice letters responding to cycle events. Use when producing letters section of an edition.
Close the feedback loop. Canonize to Supermemory, update world-data, write ratings to sheets, grade reporters, update criteria files, update newsroom memory. Type-aware — edition, interview, supplemental, dispatch all converge here.