Company Profile

Company Profile

A single slide that answers "what is this company" for a reader with thirty seconds. Four quadrants, all four full: overview, financial summary, share price or valuation chart, key facts. This is the front page of a pitch book and the first page of a diligence pack, and it is the one slide in a deck allowed to be a document, because nobody presents from it, they read it off the page.

Build it with the pptx skill. Its geometry, palette, build-script discipline and verification scripts all apply unchanged, and this file only describes what goes inside that frame. Reach for html-report instead when the answer wants two pages of prose, and comps-analysis when the question is how the company prices against its peers rather than what it is.

Evidence labels, source tiers, staleness, the readiness posture and the intake limits: .agents/skills/research-conventions/SKILL.md, read before the first deliverable.

Scope

The profile is a baseline: what is true about this company, sourced, on one page. It is not a recommendation, and it holds no view. The moment the work turns into a thesis, a scenario set, a valuation argument or a call on the stock, it has stopped being a profile: finish the baseline, then hand off and say so. .agents/skills/initiating-coverage/SKILL.md owns the thesis, .agents/skills/comps-analysis/SKILL.md owns relative valuation, .agents/skills/earnings-preview/SKILL.md owns the setup into a print. Expanding the slide to carry a view is the failure this rule exists to catch, because a one-pager that argues reads as fact to everyone downstream of it.

Workflow

1. Scope it before researching

Confirm the entity first, then ask what the slide is for.

Entity identity. Establish, before any pull: the exact issuer (a similarly named parent, subsidiary or unrelated company is the classic wrong answer), the listing line (ticker and exchange, and which share class where more than one trades), and the fiscal year end. Everything downstream carries this, and a profile of the wrong listing line is wrong in every quadrant at once.

Then run intake per .agents/skills/research-conventions/references/intake.md: read the mandate memory, ask only about the forks it leaves open, recommended option first, and start on the recommendation when no answer comes back, disclosing it as a default. Two forks are worth a question here:

• Format. One slide, recommended, or a profile deck: this slide plus two or three elaboration slides.
• Baseline, which decides what survives the character budget below: long-only (durability, capital allocation), hedge-fund (positioning, float, borrow, days to exit), coverage starter (what a new analyst needs first), index constituent (weight, passive ownership, flow), diligence (counterparty, contracts, liquidity). Offer the two the mandate makes plausible rather than the whole menu, and recommend the one it implies.

Whether the company is public or private is settled from the pull rather than asked: a private company has no share price, so quadrant 4 becomes holders, funding history and recent developments.

Asking after the research is doing the research twice.

2. Pull the data

Normalise before writing anything down: one currency, one scale ($mm or $bn, never mixed on one slide), one fiscal-year convention, and every figure labelled A for actual or E for estimate. Mark unaudited, preliminary, pro-forma, adjusted and company-defined figures as such in the bullet itself, since a pro-forma revenue figure sitting unlabelled beside a reported one is two different metrics in one column.

Provenance travels with every metric, even though the slide shows only the number. Keep period, units, source and confidence beside each figure while researching: the slide prints the number, the source line prints the tools and documents behind it, and the outline you show the user carries all four so a wrong figure is traceable in seconds rather than re-pulled.

Flag a conflict rather than picking silently. Two sources disagreeing on revenue, EBITDA, debt, share count or market cap; a data tool differing from the filing; the company's deck differing from its filing; a newer filing superseding the figure already written down. Resolve down the tier list and record the selection per .agents/skills/research-conventions/references/evidence.md, and where the conflict is material to the slide, say which figure was used in the source line.

3. Outline before code

Write the four quadrants out as plain text with the real numbers already in place. No placeholders: a bullet you cannot fill is a bullet the slide does not get. Show the outline and the accent colour, then build.

4. Build, render, look, shrink

Follow the pptx loop exactly:

1. Write <task>/build_<name>.js and run it with NODE_PATH=$(npm root -g) node <task>/build_<name>.js.
2. python .agents/skills/pptx/scripts/check.py <task>/<name>.pptx --strict, which reads the written file and reports anything off the slide, overlapping or overflowing.
3. python .agents/skills/pptx/scripts/render.py <task>/<name>.pptx --montage, then open the PNG and read it. Look for a bullet wrapping into the quadrant below it, a table row crossing the footer, an axis label clipped at the bottom, and a title sitting on top of the first quadrant header.
4. When anything overflows, fix it in this order: drop the body font by 1 to 2pt (12 to 11, then 11 to 10, which is the floor), then shorten the bullet, then move the boundary between the two rows. Re-render and look again.

Steps 3 and 4 are not optional and neither is the looking. check.py estimates overflow from font size and character count, so a pass is a reason to look, not a substitute for looking. A profile slide is the densest thing this repo produces, and density is exactly where the estimate and the renderer disagree.

Layout

16:9, LAYOUT_WIDE, 13.333 x 7.5 in, with the 0.6 in side margins and the title band from the pptx skill left alone. The content band, y 1.75 to 6.60, splits into four equal quadrants:

x=0.600                          x=6.833                       x=12.733
+--------------------------------+------------------------------+  y=1.75
| 1  Company Overview            | 2  Business and Positioning  |
| w=5.90  h=2.30                 | w=5.90  h=2.30               |
+--------------------------------+------------------------------+  y=4.05
| 3  Financial Summary           | 4  Share Price / Key Facts   |
| w=5.90  h=2.35                 | w=5.90  h=2.35               |
+--------------------------------+------------------------------+  y=6.60

Column x is 0.600 for quadrants 1 and 3, 6.833 for quadrants 2 and 4. The 0.333 in gutter is the only thing between the columns: white background, no fills, no shading, no boxes drawn around the quadrants.

Header boxes are 0.40 in and not a hair less. check.py measures a box against its text after taking off the 0.05 in inset top and bottom, so a 16pt line needs 19.2pt of clear height, and a 0.32 in box fails the overflow check even though the render looks fine.

12pt body is a deliberate exception to the deck-wide 16 to 18pt in pptx, and it is the only slide that gets it. check.py holds the hard floors either way: 10pt for body text, 14pt for table text.

Density budget

A quadrant that looks sparse has not been researched. Fill all four, then measure each against the column width.

At 12pt across a 5.90 in column, check.py fits roughly 70 characters on a line (width_in * 72 / (pt * 0.5)), so a bullet over 70 characters takes two lines and costs the quadrant a fact. Counting paragraph spacing, each body box holds 7 to 8 one-line bullets. Write to that budget: pack facts, never pad them.

• Combine related facts: HQ Austin, TX; founded 2003; 14,200 employees is one bullet, not three.
• Always carry the number: $4.0bn revenue beats large revenue, and +28% y/y beats growing fast.
• Add the comparison that makes the number mean something: EBITDA margin 25.4% (peer median 18%).
• Bold the lead term as its own run, { text: "Market position: ", options: { bold: true } }, so the quadrant scans as a list of labels.
• Set bullet: { indent: 12 }. The pptxgenjs default leaves about a third of an inch of white between the glyph and the text, which is four characters of a 70-character line spent on nothing.

Never silently omitted. These are either on the slide with their as-of, or named in the evidence gaps: market cap, float, average daily volume and days to exit, index membership, passive and ETF ownership, top holders, short interest, borrow, analyst coverage count, and the consensus setup. A missing one of these changes what a reader can do with the page, so its absence is information.

Pair the fact with the decision wherever the decision is not obvious from the number: float speaks to capacity and squeeze risk, days to exit answers whether a position can be built or unwound at all, borrow prices the other side of the trade. Float 62%; 14 days to exit at 20% ADV is one bullet doing two jobs, which is how a data table becomes a decision aid inside a 70-character line.

Evidence gaps. One compact block, never a table of empty fields. Where there are one or two gaps, they ride at the end of the source line: Not sourced: short interest, borrow (as of unavailable). Where there are more, they take the last bullet of quadrant 4 as a single line starting Not sourced:. Each gap names what is missing and, where it is not obvious, what would supply it. A gap that costs the reader a decision is worth a bullet more than the least surprising fact on the slide.

If a quadrant still runs short, the facts usually missing are segment percentages, customer concentration, guidance against consensus, and insider ownership. If it runs long, cut the least surprising fact before you cut the font.