Strategy Memo Skill Skill

Strategy Memo Skill

Produces Amazon 6-pager-style strategy memos as Quarto documents that render to publication-quality PDF (+ HTML). The canonical reference is pre-read.qmd in the intelligence-strategy-sessions project — everything in this skill is distilled from that document.

Workflow

Step 1: Interview the user

Before generating anything, ask these questions (all at once, numbered):

Company & domain — What company/team is this for? What domain (product strategy, platform investment, org change, technical direction)?
Audience — Who reads this? (e.g., Engineering + Product Directors, Board, VP-level)
Leadership framework — How should intent be framed? (Intent-Based Leadership = "I intend to…"; OKR-driven; DACI; or neutral)
Strategic thesis — In one sentence: what is the central argument or direction you are asserting?
Key questions — What are the 3–5 strategic choices the document must help the audience decide or validate?
Data sources — BigQuery project + key table names? Or will data be provided manually / hardcoded?
Target length — 6 pages is the Amazon canon. More than 8 pages is too long regardless of figures.
Tone — Urgent inflection point? Confident direction-setting? Diagnostic + open questions?

Step 2: Generate the scaffold

Once you have answers, generate a new .qmd file by adapting assets/template.qmd:

Replace all {{PLACEHOLDER}} tokens with content derived from the interview
For each key question, create a ## Part N: ... section with prose and a figure cell stub
For each data source, create a cache block in the key-numbers cell (cache-read-or-query pattern)
Set \fancyhead values (company name, document title, author)
Populate the intent declarations from the thesis and questions
Leave figure cells as stubs with # TODO: implement visualization comments

Step 3: Set up the project

Copy the supporting assets into the project root:

cp assets/Justfile ./Justfile
cp assets/scripts/preview-fig.py ./scripts/preview-fig.py
cp assets/scripts/validate-cache.py ./scripts/validate-cache.py
mkdir -p data/cache scripts

Add to .gitignore:

data/cache/
.drive-file-id
.jupyter_cache/

For Google Drive sync, create .drive-file-id containing the Drive file ID of the target PDF.

Step 4: Build loop

just render          # Full render: PDF + HTML + check + sync to Drive
just preview         # HTML only, opens in browser (fast)
just preview-fig fig1  # Screenshot a single figure (~15s, no full render)
just check           # Scan PDF for unrendered Python literals
just delete-cache    # Wipe data/cache/ — forces live BigQuery re-query
just validate        # Compare cache to live BigQuery (fails if drift > 5%)

Core Authoring Rules

See references/6pager-spec.md for document structure and writing style. See references/visualization-standards.md for figure patterns, color palette, and anti-patterns.

No hardcoded numbers in prose (strict)

Every factual number must flow from a Python variable:

Analysis of `{python} f"{total_deals:,}"` lost opportunities...

Never: Analysis of 847 lost opportunities...

Numbers that CAN be hardcoded: stylistic/non-factual values ("25 minutes of silent reading").

Cache-read-or-query pattern (mandatory for BigQuery)

_c = _read_cache('my_dataset')
if _c:
    df = pd.DataFrame(_c['records'])
    my_scalar = float(_c['scalars']['my_scalar'])
else:
    df = run_bq_query(my_query)
    my_scalar = float(df['col'].values[0])
    _write_cache('my_dataset', df.to_dict(orient='records'), {
        'my_scalar': my_scalar,
    })

Never: except Exception: my_scalar = 42 — silent fallback masks broken queries.

Figure captions: insight, not description

#| fig-cap: "Lost deal flow (2024+). The dominant flow: customers who evaluated and stayed put."

Not: #| fig-cap: "Sankey diagram of deal flow by loss reason."

LaTeX in PDF output

\begin{quote}
\small\itshape
"Customer quote here."

\upshape---Attribution, Date
\end{quote}

Escape dollars: \$110M. Em-dashes: --- (renders as —). Page breaks: \newpage.

Figure captions cannot use `{python}` expressions

Quarto limitation. Add a sync comment adjacent to any hardcoded caption value:

# CAPTION-SYNC: ratio={ratio:.0f} — update caption if ratio changes
#| fig-cap: "...approximately 4×..."

Document Structure

The proven 4-part structure from pre-read.qmd:

Introduction — Intent declarations (3 statements of "I intend to…") + "How These Sessions Work"
Part 1: The Situation — Diagnosis, primary visualization, customer/evidence quote
Part 2: Why This Intent — Strategic rationale, data moat, revenue/product evidence
Part 3: What This Requires — Technical, organizational, and execution dependencies
Part 4: Questions to Refine Intent — Exactly 3–5 strategic choices, not action items

Never add: self-blame tone, tactical product details, "Risks and Mitigations" framing, top-down action items, more than 5 strategic choices, appendices, bullet points in narrative prose.

Pre-flight Audit Checklist

Before just publish, scan for:

[ ] int(N) literals inside {python} inline expressions (literal masquerading as variable)
[ ] Numbers assigned in if _c: blocks without a preceding run_bq_query() call
[ ] Dollar amounts in LaTeX attribution blocks with no Python variable reference
[ ] Percentages written as words ("Seventy-three percent") with no {python} backing
[ ] Figure captions containing bare % characters (LaTeX compile failure)
[ ] text labels beyond xlim in matplotlib (causes comically small PDF figures)
[ ] Missing plt.close('all') after each figure cell

Agent Skills: Strategy Memo Skill

Install this agent skill to your local

Skill Files