| name | rpg-summarizer |
| description | Generates a narrative session recap from a transcript, using campaign context. Processes the transcript in sequential subagent chunks to preserve chronological order. |
RPG Summarizer
Convert a raw transcript into a narrative story using the campaign's writing style. The transcript is split into ~800-line chunks processed by dedicated subagents in sequence. Each subagent receives only its chunk plus a rolling summary of prior events, so events can never be reordered or confused across the session.
Step 1: Context Loading
You do NOT need to pre-load the full transcript into the orchestrator's context — the chunk subagents read their own slices.
Step 2: Load Style Prompt
Read .agent/skills/rpg-summarizer/resources/summary_prompt.txt in full. You will embed its entire content into every subagent prompt below — it defines tone, character names, section formatting, and language style, plus anti-hallucination guardrails and the IC/OOC distinction.
Step 3: Chunk the Transcript
Determine the total line count of the transcript via wc -l. Divide into sequential chunks of ~800 lines. Record each chunk's start line and end line.
Step 4: Sequential Subagent Processing
Process chunks strictly one at a time — never in parallel, since each depends on the previous result.
Start with rolling_summary = "".
For each chunk, spawn a subagent with this prompt:
You are a narrative summarizer for a Polish D&D campaign session. This is chunk [N of TOTAL].
Style & Rules
[paste full content of summary_prompt.txt here]
Canonical Names Glossary
[paste full stdout of extract_glossary.py here]
Phonetic Corrections
[paste full content of phonetic_corrections.md here]
Context from previous chunks
[If rolling_summary is empty:]
This is the beginning of the session. No prior context.
[If rolling_summary is not empty:]
Dotychczasowe wydarzenia sesji:
[rolling_summary]
Continue the narrative from where these events left off. Do not re-introduce characters or locations already established. Do not recap what has already happened.
Transcript excerpt
Read lines [START]–[END] of content/assets/sessions/{NNN}/transcript.txt (use Read tool with offset: [START-1] and limit: [END-START+1]). Read ONLY this slice.
Your Output
Return exactly two labeled blocks and nothing else:
NARRATIVE
[1–4 narrative ### sections in Polish covering this chunk's events, following the style rules above. Each section has a descriptive heading and 1–3 rich paragraphs.]
ROLLING_SUMMARY
[3–5 bullet points in Polish: key events, decisions, revelations, and the party's location/situation at the end of this chunk. Be specific — this is the only context the next chunk will receive.]
After each subagent completes:
- Append its
NARRATIVE block to your running output.
- Replace
rolling_summary with its ROLLING_SUMMARY block for the next iteration.
Step 5: Validation Pass
Before returning the assembled draft, run one final subagent as a fact-checker. Its job is to spot fabrications and propagated errors that slipped past the chunk subagents.
Spawn a subagent with this prompt:
You are a fact-checker for a Polish D&D session recap draft. Your job is to find specific factual claims that are NOT supported by the transcript, and propagated hallucinations.
Inputs
- Draft to verify: [paste the assembled NARRATIVE here]
- Transcript:
content/assets/sessions/{NNN}/transcript.txt ([N] lines)
- Canonical glossary: [paste glossary]
- Phonetic corrections: [paste corrections]
Method
For each claim in the draft that names a specific spell, item, NPC, number, or "first/last/never" superlative:
- Find a transcript line that supports it (use Grep/Bash to search by keyword).
- If you cannot find supporting evidence, flag it.
Also flag:
- Names not in the canonical glossary (likely ASR misspellings the chunk subagent didn't correct)
- Items/buffs/effects that come from a previous session and may have expired
- Direction-of-spell errors (who cast on whom)
- Mechanical numbers narrated as in-fiction events (e.g. "advantage k8")
Output
Return a list, one finding per line:
LINE_OR_QUOTE | what's wrong | suggested fix (or "remove")
If nothing is wrong, return exactly: NO ISSUES.
Do not rewrite the draft. Do not be conservative — flag anything you cannot verify.
The orchestrator then applies (or asks the user about) each finding before saving the file.
Step 6: Output
Concatenate all NARRATIVE blocks in chunk order. Do not reorder sections. Apply validation findings.
- Constraint: No
[[wiki links]] yet — plain text only.
- Constraint: Follow the structure from
templates/Session.md.
- Return the complete Markdown. Do not save to a file (the workflow handles that).
