npbuilds Synthesis Translator — The Format Multiplier
Spelunker produces auditable research briefs. But auditable briefs are rarely the right artifact for the destination — VC interviews want 3-bullet conviction statements, portfolio writeups want executive memos, social posts want tweet threads, decision memos want a recommendation with a counter-position. This skill takes one brief and produces the right shape without stripping the epistemic discipline.
Guiding Principles
- Citations survive translation. A claim in the original brief carries
[N]; that[N]survives in every output format. The References section comes along. - Confidence tags are not optional. A tweet thread can compress them to icons (✓ Confirmed, ◐ Likely, ? Speculative) but cannot drop them.
- No new claims. The translator restructures and compresses; it does not introduce assertions absent from the source brief.
- Honest about format limits. A 280-character tweet cannot capture nuance — flag what was cut.
How to Run
Input
- A Spelunker research brief (Markdown, in the format defined by
../evidence-synthesizer/references/synthesis-templates.md) - A target format:
memo|pitch|thread|decision - Optional: target audience descriptor (e.g., "VC partner with biotech focus", "personal Twitter followers", "internal product team") — affects tone but not content
Steps
Step 1 — Parse the Source Brief
Extract:
- The original question and depth mode
- Key Findings (with their confidence tags)
- The full citation list
- Any Contested or Speculative findings (these get special handling per format)
- The Brief ID
Step 2 — Apply Format-Specific Compression
Format: memo (executive memo, ~1 page)
Memo: <restated question, framed as an answer>
Brief: <SPK-id> · Depth: <mode> · Date: <date>
TL;DR: <2-3 sentences capturing the highest-confidence answer with an inline confidence indicator>
Key findings:
1. [Confidence] <Finding> [N]
2. [Confidence] <Finding> [N]
3. [Confidence] <Finding> [N]
What's not settled:
- <Contested finding> [N] — both positions and what would resolve
Recommended action: <if and only if the original brief had a Next Steps section, surface its top item>
References:
[N] <full citations from the source brief>
Format: pitch (3-bullet VC-style conviction statements)
On <topic>:
• <Confidence-tagged claim, ≤25 words> [N]
• <Confidence-tagged claim, ≤25 words> [N]
• <Confidence-tagged claim, ≤25 words> [N]
Where I'm uncertain: <one sentence naming the highest-priority Speculative or Contested element>
Sources: <inline citation list, can be condensed>
Each bullet should make a falsifiable claim, not a vague summary. Tag-icons (✓ for Confirmed, ◐ for Likely, ? for Speculative, ⚠ for Contested) are acceptable for compression.
Format: thread (tweet thread, 5-9 tweets)
Tweet 1 (hook): The answer in one sentence with the highest-confidence finding. Tweets 2-N (one per finding, 220 char max each):
- Lead with the confidence icon
- State the finding
- One link or citation handle (
(see [1])) — full citation in the final tweet Final tweet: Methodology note + brief ID + link to full brief.
Cap at 9 tweets. If the brief has more findings, pick the 7 highest-confidence ones and note: "Cut 3 lower-confidence findings — see brief for full picture."
Format: decision (decision memo with recommendation)
Decision Memo: <decision question — must be a question with a discrete set of answers>
Brief: <SPK-id> · Date: <date>
Recommendation: <option> — confidence: <tier>
Why:
1. [<Confidence>] <Finding supporting the recommendation> [N]
2. [<Confidence>] <Finding supporting the recommendation> [N]
Counter-position:
<The strongest case for an alternative option, drawn from any Contested or Speculative findings in the brief>
Sensitivity:
This recommendation would change if: <specific finding that, if it flipped, would change the answer> [N]
Open questions before deciding:
- <From the brief's Gaps section, the highest-priority unresolved item>
References:
[N] <full citations>
The "Sensitivity" section is mandatory for decision — it forces the brief's load-bearing claim into focus.
Step 3 — Honesty-Preserving Compression Rules
- Never upgrade a confidence tag in translation. Likely stays Likely.
- Never drop a Contested finding silently — it must appear in either the body or a "What's not settled" section.
- If the brief's overall confidence was "Insufficient evidence," the output must state this prominently in the first 50 words. Do not produce an executive memo or pitch from an insufficient-evidence brief without that warning.
- If compression forces dropping findings, list what was cut at the bottom of the output.
Output
The translated artifact in the requested format, plus a footer:
─────────────────────────────────────
Translated from: <SPK-id> (<original depth mode>)
Format: <memo|pitch|thread|decision>
Findings included: <N> of <M> from source brief
Findings cut: <list, if any>
Error Handling
Source brief lacks citations: Refuse to translate. Output: "Source brief has no inline [N] citations — translation would propagate unsourced claims. Re-run the brief with citation discipline first." This is the citation-discipline contract from P0; the translator enforces it at egress.
Source brief is missing depth mode: Default to displaying "depth: unknown" — do not assume.
Format choice doesn't match the brief shape: E.g., asking for decision format on a brief that doesn't address a discrete decision. Output a warning and ask the user whether to (a) reformulate as the closest valid format, or (b) abort.
Scope Boundaries
Does NOT: Add new findings, sources, or analysis. Does NOT: Change confidence tags. Does NOT: Translate between languages — output is in the source brief's language. Does NOT: Optimize for engagement metrics on social formats — honesty over virality.