| name | extraction-prompt-architect |
| description | Turn any dense source (article, research note, academic paper, SEC filing, earnings-call transcript, long report, spec) into a distilled MINIMAL set of MAXIMAL-impact extraction prompts — derived from orthogonal analysis dimensions, gap-checked for blind spots, and accompanied by a full rationale for every include/exclude decision. Use this whenever the user wants prompts to "extract", "analyze", "pull signal from", "mine", or "get the most out of" a document; whenever they ask which angles/dimensions to analyze a source along; whenever they want a "prompt set", "prompt pack", or to distill many prompts into a small high-leverage subset; or when they hand over a meaty document and ask how to thoroughly interrogate it. Trigger even if the word "prompt" is never said — e.g. "how should I break down this whitepaper", "what's the smallest set of questions that gets 80% of the value from this report", or "give me angles to analyze this filing from". |
Extraction Prompt Architect
What this skill is for
Given a dense source, produce the smallest set of prompts that extracts the
largest share of its strategically significant signal — plus the orthogonal
dimensions the prompts are built from, and a transparent justification for what
made the cut and what didn't.
The output is a distillation report (a Markdown file), not the answers. You
are designing the instrument, not running it. Do not execute the prompts unless
the user explicitly asks — designing and running are separate jobs, and mixing
them buries the design rationale under a wall of extracted content.
Why a fixed template fails
The naive approach ("here are 15 questions to ask any document") produces
generic sludge because it ignores two things that change everything: what
kind of source it is (a vendor research note has author incentives a peer
-reviewed paper doesn't) and what the reader will do with the extraction
(someone deciding whether to trade needs a different set than someone briefing
their boss). Dimensions and the minimal set are both functions of those two
inputs. Pin them first, derive everything else.
The workflow
Six steps. Each produces a labeled section of the final report so the user can
see the reasoning, not just the result.
Step 0 — Profile the source and bind the decision
Before any prompt-writing, establish:
- Source type — vendor/sell-side research, peer-reviewed paper, news,
regulatory filing, earnings call, technical doc, internal memo, etc. This
determines which blind spots (Step 3) are mandatory. Vendor and opinion
sources always need a provenance prompt; the author profits from you
believing them.
- Author incentive & accessibility — who wrote it, what do they gain from
the reader's belief, and is any of it paywalled/truncated (a prompt aimed at
a missing section just produces guesswork — flag it).
- The reader's decision — what will the output be used for? Understand /
trade / operate / build / monitor / brief. The minimal set in Step 4 is tuned
to this. State it explicitly; if the user didn't say, ask or assume the most
general ("understand the source deeply") and note the assumption.
Skipping Step 0 is the single most common way this workflow degrades into a
generic question list.
Step 1 — Derive orthogonal dimensions
Generate the analysis dimensions from source-type × decision, not a
template. A dimension is a distinct axis of strategic signal (e.g. for a power
-market note: capacity accounting, timing mismatch, regulatory design,
supply-chain, competitive advantage…).
Then run an explicit orthogonality pass — this is not optional, it is what
keeps the downstream prompt set from being half-redundant:
- Lay the candidate dimensions in a list and check each pair for conceptual
overlap. Merge any pair that overlaps more than ~30% (e.g. "competitive
advantage", "winners/losers", and "investment signals" usually collapse to
one or two axes).
- Drop any dimension that doesn't actually bind on this specific source.
- Target ~8–15 genuinely independent axes. More than that and you're slicing
the same signal twice; fewer and you're probably missing a layer.
Step 2 — Generate the maximal high-impact prompt set
Write one or more prompts per dimension. Each prompt should be reusable,
high-leverage, and force structure (tables, classifications, ranked lists) so it
extracts rather than summarizes. Tag every prompt with the dimension(s) it
serves and build a small coverage matrix (dimension × prompt) so over- and
under-covered axes are visible at a glance.
House style for a prompt: a directive framing paragraph that states the lens and
why it matters for this source, then a structured extraction body (numbered
asks / bullets), then a required output artifact (a table or ranked list), then
Article: / [PASTE SOURCE].
Step 3 — Blind-spot / gap pass (the step a dimension-only workflow lacks)
A dimension-derived set can only surface what the dimensions named. "Maximal" is
not maximal until you check what no dimension covered. Test the source
against this standing checklist and add a prompt for every real gap:
- Provenance / verifiability — does any prompt force a split between
externally checkable facts and the author's unverifiable model outputs or
opinion? Mandatory for vendor/opinion sources. (This is usually the single
highest-value addition for sell-side research.)
- Author incentive & reflexivity — is there a cui bono lens (what does the
author gain from belief?) and a reflexivity lens (if the thesis becomes
consensus, does it self-defeat — gluts, backlash, edge competed away)?
- Definitional / ontology — is the source's load-bearing vocabulary
(the 5–10 terms everything else rests on) extracted and defined? Misread the
vocabulary and every number is misread.
- Recall safety net — is there a maximal-recall sweep (e.g. paragraph-by
-paragraph) to catch concrete detail the thematic prompts silently drop?
- Base-rate / comparative — vs history, vs the author's own prior track
record, vs an independent benchmark. If the source makes datable predictions
and part of their horizon has already elapsed, exploit the current date: grade
the now-resolved checkpoints against what actually happened. A forecast's own
scored near-term predictions beat any generic base-rate.
Not every source needs all five. Add only the gaps that genuinely bind, and say
why the others were skipped.
Step 4 — Distill to the minimal set
Apply an explicit, auditable criterion. A prompt earns a minimal-set slot
only if it meets all three tests:
- Yield — extracts a large, decision-relevant share of the signal.
- Orthogonality — that share is not substantially recoverable from
another prompt already in the set.
- Load-bearing — the signal underpins the thesis or one's trust in it,
rather than being role-specific or speculation-prone.
Distillation has three moves, not one: drop the redundant, merge
overlapping prompts into one, and keep what's load-bearing. The minimal set is
decision-relative — a trader's set ≠ an operator's set ≠ a briefer's set;
tune it to the decision from Step 0.
Emit a run-order, because sequence matters. The default spine:
orient (master extraction) → trust-gate (provenance, if present) → rebuild the
core model → rank the constraints → read the concrete near-term evidence →
adversarial check → synthesize.
A good minimal set is usually 5–8 prompts for a single dense source. If it's
larger, suspect you've failed test 2 somewhere.
Step 5 — Explain everything (exclusion ledger)
The rationale is a first-class deliverable, not a footnote. Include:
- Per-step decisions (why these dimensions, why these blind-spot adds).
- The coverage matrix.
- An exclusion ledger: every prompt left out of the minimal set, the exact
signal forgone, whether it's recoverable from a set member, and which 80/20
test it failed.
- First adds beyond minimal, in priority order — what to add first if the
reader's decision shifts (e.g. "add the investment-hypothesis prompt once the
paywalled winners section is obtained").
Output format
Write a single Markdown report. Use this skeleton:
# Prompt Distillation Report — <source short name>
**Source / type / author-incentive / accessibility:**
**Reader decision this set serves:**
**Status:** Distillation only. No prompt has been run.
## 1. The inclusion criterion (the 3 tests, stated so decisions are auditable)
## 2. Dimensions (the orthogonal axes + the merge/drop log)
## 3. The minimal set — N prompts (table: # | run-order | prompt | why it clears 80/20)
## 4. New / blind-spot prompts (full text of any prompt added in Step 3, each with an include/exclude verdict)
## 5. Exclusion ledger (every omitted prompt → signal lost → recoverable? → which test it failed)
## 6. Bottom line (the set in run-order; first-adds beyond minimal)
Full prompt text belongs in the report (so it's copy-pasteable). Keep the
framing prose dense — the reader asked for an instrument plus its design notes,
not an essay.
Scope discipline
- Design, don't run. Producing the report is the job. Offer to run the set
as a separate follow-up.
- Be honest about gaps. If part of the source is missing/paywalled, say
which prompts are starved by it rather than pretending coverage.
- Right-size. A short or simple source may only warrant 3–4 prompts and two
dimensions; don't manufacture fifteen. The minimal set scales with the
source's actual signal density.
Upgrade path (mention, don't build unless asked)
If the user later wants the minimal set executed rather than just designed,
that's a fan-out job: one subagent per prompt in run-order, results merged into
a synthesis — a Workflow/plugin, not this skill. This skill stops at the
instrument.