setup-new-firm

name

setup-new-firm

description

End-to-end procedure for onboarding a new VC firm into the memopop investment-memo orchestrator. Five phases — (1) brand acquisition via Firecrawl with analyst-override on the primary-color heuristic, (2) firm-directory scaffolding under io/<firm>/ with the canonical configs/templates/assets/deals layout and the strict brand-config YAML schema, (3) investment-framework outline selection (typically forking an existing outline like alpha-partners-7-Cs and customizing — section reordering, preambles, preferred-source swaps, stage-aware fallbacks), (4) first deal setup with deck-input handling (DocSend capture via cli/capture_docsend.py, ImageMagick screenshot stitching for NDA-gated decks, deal-config JSON, codified Sources.md), and (5) optional private-repo split via gh CLI with submodule re-add for confidential firm data. Use whenever the user says "I have a new client/firm/VC," "set up <firm name>," "add <firm> to memopop," "we need brand configs for <firm>," or pastes a VC firm's URL with implicit intent to generate memos for them. Encodes lessons from the alpha-jwc onboarding on 2026-06-07 — don't trust Firecrawl's primary-color field blindly, verify against the logo palette; the deck analyst will extract DocSend watermarks as the company name if you feed it watermarked screenshots; agents leak firm-geography context into company-analysis prose (Alpha JWC = Indonesia VC → Portland-OR company memo gained Indonesia/rupiah risks); citation-enrichment fabricates example.com URLs even in codified-source mode — codified Sources.md only constrains the upstream researcher, not the downstream enricher. Cross-references [[crawl-fetch-ingest]] for portfolio/team data, [[sources-md-curation]] for the Sources.md file mechanics, and the open issues in context-v/issue-resolution/ ([[Limiting-or-Omitting-Investor-Judgement]], [[Competitive-Research-Generated-But-Not-In-Prose]], [[Including-Comparable-Exits-Valuations-IPOs]]) for known failure modes the new firm's first run will hit.

Setting Up a New Firm in MemoPop

The analyst has a new VC client. The orchestrator needs to know who they are (brand), how they think (outline), and what the first deal looks like (deck + sources). This skill is the procedure that gets a new firm from "we have their URL" to "ready to run python -m src.main for their first deal," ideally in one focused session.

When this skill activates

The signal pattern is new firm + intent to generate memos:

• "I have a new client — <firm-url>"
• "Set up <firm name> in memopop"
• "We don't yet have brand configs for <firm>"
• "Add <firm> and let's run their first deal through"
• Plain URL paste of a VC homepage in a conversation about memo generation

Do NOT activate this skill for:

• Adding a new deal to an existing firm (no skill needed — copy the deal-JSON pattern from a sibling deal directory).
• Modifying an existing firm's outline (use [[sources-md-curation]] mechanics for source-file edits; outline edits are direct YAML work).
• Per-deal source curation (delegate to [[sources-md-curation]] or [[user-adds-adhoc-sources]]).

The five phases

Phase 1 — Brand acquisition       (~10 min, Firecrawl + analyst override)
Phase 2 — Directory scaffolding   (~5 min, fixed layout)
Phase 3 — Outline customization   (~15-60 min, analyst-driven)
Phase 4 — First deal setup        (~15-30 min, deck handling varies)
Phase 5 — Private-repo split      (~5 min, optional)

Run them in order. Phases 1–4 are usually one session; Phase 5 is often a follow-up after the first run looks reasonable.

---

Phase 1 — Brand acquisition

The firm's homepage is the source of truth for colors, fonts, and logos. Firecrawl's branding format does most of the work but its primary: field is unreliable for VC sites — it tends to pick CTA-button colors over brand-identity colors. Always sanity-check against the logo palette.

Step 1.1 — Fetch the homepage with the branding extractor

# Via the firecrawl MCP server
firecrawl_scrape(
    url="https://<firm-homepage>",
    formats=["branding"]
)

Read carefully:

• branding.colors.primary / secondary / accent — candidates, not truth.
• branding.fonts[] — body + heading families; usually correct.
• branding.images.logo — often a data:image/...;base64,... data URI for WordPress/Divi sites. Decode and save.
• branding.images.favicon — usually a direct URL; curl -o it.

Step 1.2 — Reconcile primary color against the logo palette

Look at the actual logo image. If the dominant color in the logo is different from Firecrawl's primary:, trust the logo, not Firecrawl. Concrete example:

Alpha JWC's homepage has amber (#FFAC00) CTA buttons and a deep-navy (#13294B) wordmark logo. Firecrawl picked primary: #FFAC00 because button frequency dominates the CSS. The brand-identity color is the navy. Final mapping: primary: #13294B, secondary: #FFAC00, accent: #8B1B40 (their burgundy 10th-anniversary palette accent).

Step 1.3 — Save logo + favicon

# For base64 data URIs:
python3 -c "
import base64, urllib.parse
uri = '<data:image/png;base64,...>'  # or image/svg+xml;utf8,<urlencoded>
# For SVG-utf8: urllib.parse.unquote(uri.split(',',1)[1])
# For base64:   base64.b64decode(uri.split(',',1)[1])
open('assets/trademark__<Firm>--Brand.png', 'wb').write(decoded_bytes)
"
curl -sL -o assets/favicon__<Firm>--Brand.png "<favicon-url>"

Naming convention is load-bearing for the downstream brand-application step. Match the pattern from io/alpha-partners/assets/:

trademark__<Firm>--Brand.png         ← light-mode default
trademark__<Firm>--Dark-Mode.svg     ← dark-mode (when available)
favicon__<Firm>--Brand.png

Phase 1 anti-patterns

• Don't accept Firecrawl's primary blindly. Always look at the logo.
• Don't scrape colors from inline-base64 logos themselves — the palette is encoded but parsing PNGs to extract dominant colors adds complexity for little gain. Eyeballing the logo + comparing to Firecrawl's candidates is faster and more honest.
• Don't fetch the homepage with firecrawl_scrape(formats=["markdown"]) hoping to find the logo URL — most modern VC sites inline the logo as base64 in the rendered HTML, and the markdown rendering loses the data URI. Use formats=["branding"] specifically.

---

Phase 2 — Directory scaffolding

The orchestrator expects every firm under io/<firm>/ to follow a fixed layout. Deviating breaks downstream pipeline assumptions.

Canonical layout

io/<firm-slug>/
├── README.md                          ← what's in here, sensitivity notes
├── .gitignore                         ← canonical entries below
├── versions.json                      ← starts as `{}`; pipeline writes here
├── configs/
│   └── brand-<firm-slug>-config.yaml  ← THE brand-application source
├── assets/
│   ├── trademark__<Firm>--Brand.png
│   ├── favicon__<Firm>--Brand.png
│   └── fonts/                         ← (empty unless custom fonts dropped)
├── templates/
│   ├── outlines/                      ← firm-specific memo outlines
│   └── scorecards/                    ← firm-specific scorecards
└── deals/                             ← per-deal subdirectories (Phase 4)

Brand config YAML schema (strict)

Mirror io/alpha-partners/configs/brand-alpha-partners-config.yaml. The schema is consumed by src/branding.py — extra fields are tolerated but the following keys are required exactly:

company:
  name: <Firm Full Legal Name>
  tagline: <one-sentence positioning>
  confidential_footer: This document is confidential and proprietary to {company_name}.
colors:
  primary: '#XXXXXX'        # the navy/blue/dark from the LOGO
  secondary: '#XXXXXX'      # the accent/CTA color
  accent: '#XXXXXX'
  text_dark: '#1A2332'      # body text
  text_light: '#6B7280'     # muted text
  background: '#FFFFFF'
  background_alt: '#F7F8FA'
colors_dark:                # dark-mode variants (invert primary/secondary)
  primary: '#XXXXXX'
  secondary: '#XXXXXX'
  accent: '#XXXXXX'
  text_dark: '#F7F7F7'
  text_light: '#9CA3AF'
  background: '#0E1A33'
  background_alt: '#16294B'
fonts:
  family: <body font name>
  fallback: -apple-system, BlinkMacSystemFont, 'Segoe UI', Arial, sans-serif
  google_fonts_url: https://fonts.googleapis.com/css2?family=...
  weight: 400
  header_family: <heading font name>
  header_fallback: -apple-system, BlinkMacSystemFont, 'Segoe UI', Arial, sans-serif
  header_google_fonts_url: https://fonts.googleapis.com/css2?family=...
  header_weight: 700
logo:
  light_mode: io/<firm-slug>/assets/trademark__<Firm>--Brand.png
  dark_mode: io/<firm-slug>/assets/trademark__<Firm>--Brand.png
  width: 180px
  height: 60px
  alt: <Firm> Logo

Canonical .gitignore

__pycache__/
*.pyc
.venv/
.DS_Store
Thumbs.db

# Raw deck captures — the canonical input is the stitched PDF in inputs/.
# Loose screenshots at the deal root are redundant and bloat the repo.
deals/*/Screenshot *.png
deals/*/Screen Shot *.png

# Intermediate working files (uncomment if you don't want to track them)
# deals/*/outputs/
# deals/*/inputs/dataroom/

The Screenshot *.png pattern uses the Unicode narrow no-break space (U+202F) that macOS screenshots use by default — but standard filename-glob in .gitignore matches it as a regular * wildcard, so the literal pattern with a regular space works.

versions.json

Start as {}. The pipeline mutates it on every run.

README.md

Use io/alpha-jwc/README.md as a structural template. Sections to include:

• One-line confidentiality header
• "How it's consumed" — submodule mount point under apps/memopop-orchestrator/io/<firm>/
• "Directory layout" — the canonical structure with notes
• "Investment framework" — the firm's outline summary
• "Branding" — color/font choices with rationale
• "Deal status" table — populated as deals come in
• "Submodule lifecycle (notes for re-cloning)" — the git submodule add ... recipe
• "Sensitivity notes" — what's confidential, what's not

---

Phase 3 — Outline customization

VC firms each have a preferred frame for evaluating opportunities. The orchestrator's templates/outlines/ system supports per-firm outlines that override the global default. Three paths:

1. Use the global default (templates/outlines/direct-investment.yaml) — fastest, but the firm reads the memo and immediately notices the framing isn't theirs.
2. Fork an existing firm outline (most common) — copy alpha-partners-7-Cs-memo-template.yaml or another firm's, rename, and customize.
3. Author from scratch — only if the firm has a published framework that doesn't compose with any existing outline.

The fork-and-customize pattern (typical case)

cp io/alpha-partners/templates/outlines/alpha-partners-7-Cs-memo-template.yaml \
   io/<firm>/templates/outlines/<firm>-7Cs-customized.yaml

Then iterate with the analyst on what to change. Common customizations, from the alpha-jwc case:

• Section reordering — Alpha JWC wanted Risks + Diligence front-loaded as sections 1–2, with the 7 Cs sliding to 3–9. Renumbering touches:
  • sections[].number
  • sections[].filename (the NN- prefix)
  • cross_section_requirements.narrative_flow.sequence
• Verbatim preambles — Alpha JWC supplied analyst-authored preamble text for Risks and Diligence sections to appear verbatim at the top. Implement by adding a preamble: field on the section + embedding the preamble text in the description: (so the LLM sees it) + listing "Preamble paragraph rendered verbatim at top of section" in section_vocabulary.required_elements + adding a validation criterion.
• Stage-aware fallbacks — For early-stage firms, sections like CAGR and Cash on Cash Return Probability won't have own-company data. Add an explicit "Stage-aware fallback" paragraph in the section's description: telling the writer to pivot to category-level evidence (public + growth- stage comps in the same sector/geography) rather than fabricate company numbers.
• Preferred-source swaps — Replace the source list with firm-relevant databases. SEA-focused firms get @dealstreetasia, @techinasia, @e27; US growth-stage firms get @pitchbook, @sacra, @cbinsights.
• Recommendation lexicon — Default is PASS / CONSIDER / COMMIT. Some firms use different verdict terms. Update vocabulary.recommendation.required_terms.
• Mode emphasis — Default supports consider (prospective) and justify (retrospective). If a firm only ever runs one mode, mark the other in metadata.compatible_modes for clarity (does not enforce).

Renumbering sed pattern (when reordering sections)

For renumbering N existing sections to start at offset K:

# Reverse order to avoid double-renumbering (e.g., 1→3, 2→4 — if you ran 1→3
# first, then 3→5 would catch your renamed 1.)
for old in 7 6 5 4 3 2 1; do
  new=$((old + 2))
  sed -i '' "s/^  - number: ${old}$/  - number: ${new}/" "$OUTLINE"
  sed -i '' "s/filename: \"0${old}-/filename: \"0${new}-/" "$OUTLINE"
done

Phase 3 anti-patterns

• Don't reshape the section taxonomy mid-run. Per AGENTS.md §1 and CLAUDE.md architectural direction #3, the outline is the section taxonomy contract. Finalize it in Phase 3, before Phase 4.
• Don't add a Team section to a 7Cs derivative without thinking about where it folds. The 7 Cs treat team quality implicitly (Capital Syndicate for investors, Category Leadership for founder-market-fit). Alpha JWC decided team affiliations land in Category Leadership.
• Don't trust the writer to render an analyst-supplied preamble verbatim without three reinforcements (description + structure_template + validation_criteria) — see Phase 3's "Verbatim preambles" note above. Even with all three, LLM rendering of "verbatim" is probabilistic; for hard guarantees, add a small preamble-injection step in src/agents/writer.py that prepends section.preamble + "\n\n" to the saved output.

---

Phase 4 — First deal setup

Each deal gets its own subdirectory under io/<firm>/deals/<DealName>/. The directory name matters operationally — it's parsed for display, and renaming it later requires sweeping multiple files (see "Operational rename checklist" below).

Naming convention

<CompanyName>-Deck-<Stage> — example: Panthalassa-Deck-Series-B. Confirm the stage with the analyst before creating; getting it wrong forces a rename.

Per-deal layout

deals/<DealName>/
├── <DealName>.json                    ← deal config (schema below)
├── inputs/
│   ├── <deck>.pdf                     ← canonical deck input
│   └── Sources.md                     ← codified analyst-curated sources
├── assets/
│   ├── trademark__<Company>--Light.svg
│   └── trademark__<Company>--Dark.svg
└── outputs/                           ← pipeline writes here per version

Deal config JSON schema

{
  "type": "direct",                                 // "direct" | "fund"
  "mode": "consider",                               // "consider" | "justify"
  "description": "<one-paragraph company summary for agent context>",
  "url": "<company-homepage>",
  "stage": "<Seed | Series A | Series B | ...>",
  "deck": "io/<firm>/deals/<DealName>/inputs/<deck>.pdf",
  "outline": "<firm>-7Cs-customized",               // resolves to io/<firm>/templates/outlines/<name>.yaml
  "trademark_light": "io/<firm>/deals/<DealName>/assets/trademark__<Company>--Light.svg",
  "trademark_dark":  "io/<firm>/deals/<DealName>/assets/trademark__<Company>--Dark.svg",
  "socials": {
    "x": "<https://x.com/...>",
    "linkedin": "<https://www.linkedin.com/company/...>"
  },
  "notes": "<analyst research-focus notes; flag firm-vs-company-geography mismatches here>"
}

The notes: field is where the analyst should call out anything that might otherwise leak from firm-context into company-analysis prose. Example for Panthalassa (Portland-OR company in an Alpha JWC = Indonesia VC firm directory):

"Geography: US-headquartered; for Alpha JWC this is a cross-regional play rather than an Indonesia/SEA-domiciled deal — flag this in Risks (regulatory regime is US/maritime, not OJK/SEA)."

Even with this note, expect the writer to leak firm-geography into company-risk reasoning. See [[Limiting-or-Omitting-Investor-Judgement]] and the broader pattern in context-v/issue-resolution/.

Company-level brand acquisition

For the company logo + tagline, compose with the [[crawl-fetch-ingest]] skill — same Firecrawl branding extraction as Phase 1, but applied to the company URL rather than the firm URL. Save extracted SVGs to deals/<DealName>/assets/ with the trademark__<Company>--Light.svg / --Dark.svg naming convention.

Deck input — three paths

The deck is the single most load-bearing input. Three paths depending on what the analyst received:

Path A — Founder sent a PDF

Drop it in inputs/<deck>.pdf. Done.

Path B — DocSend without an NDA gate

Use cli/capture_docsend.py (built 2026-06-07; written for the alpha-jwc Panthalassa deal):

source .venv/bin/activate
python cli/capture_docsend.py \
  --url "https://docsend.com/view/<deck-token>/d/<page-token>" \
  --out "io/<firm>/deals/<DealName>/inputs/<deck>.pdf"
# Optional flags: --email <addr> --max-pages N --keep-pngs

The script drives headless Chromium via Playwright, loops through slides, screenshots each as PNG, stitches to PDF via PyMuPDF. Saves directly to the pipeline's expected input location.

Path C — DocSend with an NDA gate (or analyst-captured screenshots)

The capture CLI cannot bypass NDA eSignature gates. The analyst captures manually inside an authenticated browser, drops the screenshots in the deal directory, and the skill stitches them:

cd io/<firm>/deals/<DealName>
ls Screenshot*.png | sort > /tmp/png-list.txt
magick @/tmp/png-list.txt -resize 1920x -quality 85 -compress JPEG \
  inputs/<deck>.pdf

Critical: verify the slide order before trusting it. macOS screenshot timestamps sort lexicographically into capture order (which equals deck order if the analyst paged forward through the deck). If the analyst paged backwards, ls -r is correct. Spot-check the first and last slides against the deck's slide indicator.

Sizing: -resize 1920x -quality 85 produces ~10MB PDFs from ~150MB of retina screenshots without losing analytical fidelity (Claude Vision re-renders at 150 DPI internally either way).

Sources.md initial seed

Set mode: codified in the frontmatter. Codified mode confines the research-phase agents to the listed URLs — broad search is disabled.

The frontmatter schema, with concrete defaults:

---
mode: codified
deal: <DealName>
firm: <firm-slug>
date_curated_initial: YYYY-MM-DD
date_curated_current: YYYY-MM-DD
at_semantic_version: 0.0.0.1
curated_by:
  - <Analyst Name>
augmented_with: Claude Code (Opus 4.7)
sources:
  - url: https://...
    title: "..."
    publisher: "..."
    published_date: YYYY-MM-DD
    sections: [<tag>, <tag>]
    rank: 1                       # 1 = primary; 2 = secondary
    sensitivity: citable_externally
    note: "Analyst-added <date>. <why this source matters>"
---

Section tags are the filename slug between the NN- prefix and -research.md (e.g., 01-risks-research.md → tag risks). Tag a source for every section where it's load-bearing; over-tagging is fine.

See [[sources-md-curation]] for the full file-mechanics discipline.

Phase 4 anti-patterns

• Don't feed watermarked screenshots to the deck analyst. It will extract the watermark as the company name. Alpha-jwc's first run produced "company_name": "Dropbox (Content)" because every captured slide had a Dropbox DocSend watermark bar at top. If the source is DocSend-watermarked, either (a) crop the watermark from each screenshot before stitching, or (b) live with the issue and manually fix the extracted company name in the pipeline state before the writer runs.
• Don't expect codified Sources.md to prevent citation hallucination downstream. Codified mode only constrains the upstream researcher. citation_enrichment.py still calls Perplexity Sonar Pro post-writing to add citations to drafted prose — and that pass fabricates example.com URLs at scale (alpha-jwc's v0.0.2 had 65 of them). See [[Faked-Sources-from-Perplexity]].
• Don't run with a deal directory whose name doesn't match the actual stage. Renaming after the first run requires updating: versions.json key, deal JSON internal paths (deck, trademark_light, trademark_dark), inputs/Sources.md deal: field, outputs/<v>/Sources-aggregated.md deal: field + the H1 header. See "Operational rename checklist" below.

Operational rename checklist

If the deal name has to change post-creation (e.g., Series-C → Series-B because the actual round is different):

1. mv deals/<OldName> deals/<NewName>
2. mv deals/<NewName>/<OldName>.json deals/<NewName>/<NewName>.json
3. In that JSON, rewrite the deck, trademark_light, trademark_dark paths (any string containing <OldName> → <NewName>)
4. In inputs/Sources.md frontmatter, update deal: <NewName>
5. In every outputs/<v>/Sources-aggregated.md, update deal: <NewName> in frontmatter + the # Aggregated Sources — <NewName> H1
6. In io/<firm>/versions.json, rekey the "<OldName>" entry to "<NewName>" and update the file_path field in each history entry
7. Don't bother scrubbing <OldName> from generated prose inside 2-sections/*.md, 1-research/*.md, etc. — those are versioned outputs that affect re-runs, not future runs.

---

Phase 5 — Private-repo split (optional but recommended for confidential firms)

If the firm directory contains NDA material, deal-stage confidential analysis, or any data that shouldn't sit in the orchestrator's main repo history, split it into its own private repo under lossless-group and re-add as a submodule.

Step 5.1 — Audit before pushing

cd io/<firm>
test -d .git && echo "ALREADY A REPO — investigate" || echo "clean, proceed"
find . -type f \( -name '.env*' -o -name '*.secrets*' \) 2>&1 | head
du -sh .
gh auth status
gh api user/memberships/orgs/lossless-group | head -3   # confirm admin/member

Step 5.2 — Create the remote

gh repo create lossless-group/<firm>-secure \
  --private \
  --description "<Firm> firm data for the MemoPop investment-memo orchestrator. Consumed as a git submodule by apps/memopop-orchestrator/io/<firm>/."

Step 5.3 — Init, commit, push

cd io/<firm>
git init -b main
git add .
git commit -m "initial commit: <firm>-secure firm data layer

<descriptive commit body>

Co-Authored-By: Claude Opus 4.7 (1M context) <[e-mail]>"

git remote add origin https://github.com/lossless-group/<firm>-secure.git
git push -u origin main

Step 5.4 — Verify on GitHub before continuing

Have the analyst load https://github.com/lossless-group/<firm>-secure in a browser and confirm:

• Privacy: private (not public)
• Tree: README, configs/, templates/, assets/, deals/, versions.json all present
• No accidentally-included secrets

If anything's wrong (e.g., screenshots committed that shouldn't have been), the cleanest fix on a fresh repo with one commit is squash + force-push:

git reset --soft $(git rev-list --max-parents=0 HEAD)
git commit --amend -m "<new clean message>"
git push --force-with-lease origin main

Force-push to main is normally risky; on a private brand-new repo with zero collaborators and one commit, it's the right tool. Confirm with the user before force-pushing if any doubt exists.

Step 5.5 — Re-add as submodule

cd apps/memopop-orchestrator
rm -rf io/<firm>
git submodule add https://github.com/lossless-group/<firm>-secure.git io/<firm>
git commit -m "feat(io): add <firm> as private submodule"
git submodule update --init --recursive

Collaborators without read access to the private repo will see io/<firm> empty on clone — that's the intended access boundary.

Phase 5 anti-patterns

• Don't push raw deck screenshots. They're already in the stitched PDF and add ~100MB per deal. The canonical .gitignore from Phase 2 catches them automatically.
• Don't force-push to main on a non-fresh repo. This is only safe on a brand-new private repo with one commit and no collaborators. After that, use a deletion commit + regular push, or rewrite with git filter-repo and coordinate with collaborators.
• Don't forget to update the parent repo's .gitmodules and submodule pointer commit. git submodule add does this automatically; just don't skip the commit step in 5.5.

---

Known failure modes to flag during onboarding

The first run on a new firm will hit several known issues that are documented but not yet fixed. Warn the analyst proactively:

• example.com fabricated URLs — citation_enrichment.py invents citations at scale. Even with codified Sources.md. Track via [[Faked-Sources-from-Perplexity]] and the broader [[Separating-Retrieval-from-Generation-in-Agent-Pipelines]] exploration.
• Firm-geography leaks into company analysis — Alpha JWC's Indonesia/SEA context leaked into Panthalassa's (Portland-OR) Risks section: rupiah FX, OJK regulatory references, "Indonesian island markets." See [[Limiting-or-Omitting-Investor-Judgement]].
• Unsolicited PASS verdicts — the writer renders PASS/CONSIDER/COMMIT in mode: consider even when the analyst is preparing the memo for IC advocacy. See [[Limiting-or-Omitting-Investor-Judgement]] for the resolution direction (intent: field on per-deal focal points).
• Competitive landscape evaluator output orphaned from writer — the competitive-landscape agent runs cleanly and produces named competitors, but the writer never sees its output. Category Leadership section ends up with <needs-source> markers despite having competitor data on disk. See [[Competitive-Research-Generated-But-Not-In-Prose]].
• No comparable-exits harvester — Cash on Cash Return Probability section needs IPO/M&A comps to anchor scenarios; no dedicated agent produces them. See [[Including-Comparable-Exits-Valuations-IPOs]].

---

Composes with

• [[crawl-fetch-ingest]] — Phase 1 firm brand acquisition + Phase 4 company brand acquisition both delegate to this skill's Firecrawl patterns.
• [[sources-md-curation]] — Phase 4 inputs/Sources.md authoring mechanics live there.
• [[user-adds-adhoc-sources]] — the post-run workflow when the analyst wants to weave new URLs into existing memo research without a full re-run.
• [[Per-Deal-Focal-Points]] (spec) — the proposed analyst-supplied emphasis layer that the orchestrator should consume per deal; will eventually compose with this skill when implemented.
• changelog-conventions — when a new firm is meaningful enough to warrant a changelog entry, follow that skill's format.

See also

• apps/memopop-orchestrator/io/README.md — the firm-scoped IO pattern documentation
• apps/memopop-orchestrator/io/alpha-partners/ — the reference fully-wired firm directory (use as visual template)
• apps/memopop-orchestrator/io/alpha-jwc/ — the most recent onboarding (the worked example this skill was extracted from)
• apps/memopop-orchestrator/cli/capture_docsend.py — the DocSend → PDF capture CLI (written 2026-06-07 during the alpha-jwc onboarding)