Customize Your Skill to Enable Automatic IM Screening
Package the thesis into the skill itself, and no file has to be attached again.
Session 1 of the Circle produces myfund.md, a fund’s thesis written down as screening criteria a machine can act on. What happens next splits into two paths: attach it inside a Claude Project, Path One, or build it into a skill that carries its own copy of the thesis and fires in any chat on its own, Path Two. This page is Path Two.
A Claude Project, Path One, keeps myfund.md and the screening prompt in one place, but a Project is still one place. A deal shared outside it, in a different chat, a different device, a colleague’s own conversation, doesn’t get screened against the thesis unless someone remembers to open the right Project first.
A skill removes that dependency. It carries its own instructions and, built the way this one is, its own reference file, so the fund’s thesis travels with it instead of sitting in a Project only one person opens. Saved to an account, or turned on for a whole team, it applies itself the moment a CIM, teaser, or memo appears in any conversation.
This is more setup than Path One, and it’s worth doing once screening stops being occasional and becomes routine. Budget twenty minutes. Nothing here requires code, only a text file, a folder, and a zip.
Five steps, done once per fund. The skill only needs rebuilding when the thesis itself changes.
Enable Code execution
Skills sit behind Claude’s Code execution toggle. In Settings, under Capabilities, turn it on. Team and Enterprise admins also need to enable Skills for the workspace before anyone on the team can add one.
Draft the skill’s instructions
The block below is the same screening logic as Path One, rewritten as a skill: check disqualifiers before scoring, pick the rubric that matches the fund’s vehicle type, close with a verdict and a reason code. Copy it as written, or adjust the wording to match how the fund actually talks about a deal.
Bundle the thesis
Save myfund.md into a folder named references, sitting next to SKILL.md. The skill reads its thesis from that bundled file by default, so nobody has to attach myfund.md to a conversation again. Update the reference file and re-upload the skill whenever the thesis changes.
Package and upload
Zip the folder containing SKILL.md and the references folder together. In Settings, under Capabilities, open Customize, then Skills, then Add, and upload the zip. Give it a name the team will recognize, and turn it on.
Test it
Open a brand-new chat, outside any Project, and drop in a real or sample CIM. If the skill is working, it screens the deal against the bundled thesis without being asked to, checks disqualifiers first, and closes with a verdict and a reason code, not a request for criteria.
The skill’s instructions. Save this as SKILL.md.
Copy the block below into a plain text file named exactly SKILL.md. Most text editors save plain text without trouble; a word processor can silently add formatting or change the file extension, so check the saved file is still plain text before zipping it.
— name: deal-screening-assistant description: Screen inbound investment deals (CIMs, teasers, or memos) against a fund’s saved investment thesis. Reads the fund’s boundaries, scoring rubric, disqualifiers, and pass taxonomy from the bundled reference file references/myfund.md, extracts deal facts, and produces a one-page pass or fail screening memo. Use when a CIM, teaser, or deal memo needs a first-look screen, or when asked whether a deal fits the fund’s thesis. Triggers on “screen this deal”, “review this CIM”, “does this fit our thesis”, “should we take this meeting”, or “screen this against myfund”. —
You are the Deal Screening Assistant for this fund. Screen inbound deals against the fund’s saved thesis, not generic judgment.
STEP 1: LOCATE THE THESIS
Read the fund’s thesis from the bundled reference file at references/myfund.md, packaged with this skill. If a newer myfund.md, or a JSON thesis spec, is also attached directly to this conversation, prefer the attached version, since it may reflect a more recent quarterly review that hasn’t been repackaged into the skill yet. Use whichever version is in force as the source of truth for boundaries, the scoring rubric, the disqualifiers, and the pass taxonomy. Do not ask the team to restate anything already defined there.
STEP 2: EXTRACT DEAL FACTS
From the CIM, teaser, or memo provided, extract: company name, location, sector; a two-sentence description of what they do; financials (revenue, EBITDA, margins, growth rate, or for early-stage and blended finance deals: burn, runway, unit economics); deal type and instrument; asking price, valuation, or pricing basis (for debt, coupon, tenor, security); seller or fundraising motivation; management continuity or founder background; customer concentration; obvious red flags, including currency or country risk on cross-border deals.
STEP 3: CHECK DISQUALIFIERS FIRST
Test the deal against every rule in the thesis’s disqualifiers, including the sanctions and restricted-jurisdiction screen. If any disqualifier is triggered, stop. Report an immediate hard pass, citing the matching reason code from the pass taxonomy, and state whether reengage_on_change is true for that code. Do not proceed to Step 4.
STEP 4: SCORE AGAINST THE RUBRIC
Apply the fund’s weighted scoring rubric. Use the vehicle type from the thesis to pick the right lens: a buyout mandate is scored on EBITDA, margin, growth profile, deal size, and valuation as a multiple of EBITDA; a growth equity, venture, or blended finance mandate is scored on revenue growth, unit economics, pricing basis, capital stack fit, and currency exposure instead. Judge every criterion against its 1, 3, and 5 anchors from the rubric, not a blank yes or no.
STEP 5: OUTPUT A ONE-PAGE SCREENING MEMO
Verdict: Pass, Further Diligence, or Hard Pass, with the pass taxonomy code and the reengage_on_change flag if declining. Bull case, two to three bullets. Bear case, two to three bullets. Key questions to answer on a first call.
RULES
Do not soften red flags. Do not score a deal the thesis already disqualifies. If financials are inconsistent or incomplete, say so directly rather than filling the gap with an assumption. If neither the bundled reference file nor an attached file has a thesis in it, ask for one, or ask for the fund’s criteria directly, before screening anything.
A skill without its bundled reference file still works, it will just ask for myfund.md the first time in every new chat, the same as Path One does today. Bundling it, per Step 3, is what removes that step for good.
A skill is instructions. It still needs the thesis to be current.
Saving the skill doesn’t freeze the fund’s thesis in place, it freezes whatever version of myfund.md was inside the zip at upload time. When the thesis changes, quarterly review, a new disqualifier, a revised check size, update myfund.md, rebuild the zip with the same references folder, and upload it again under the same skill name. The old version is replaced, not stacked alongside the new one.
A personal skill is visible only to the person who added it. A Team or Enterprise admin can instead publish it for the whole workspace, so every member’s Claude account screens against the same thesis without each person building their own copy. That’s the better setting for a fund with more than one person taking first-look meetings.
A skill’s instructions are wording, not code. Anthropic doesn’t publish a guarantee that the same SKILL.md behaves identically after a model update. Re-run Step 5 with a known test CIM every so often, the same way you’d sanity-check any recurring judgment call rather than assume it still holds.
Path One and Path Two both feed Session 2 directly.
This is the second half of Session 1’s take-home, and it feeds Session 2, Screening, Thesis Fit and Preliminary IC, the same way Path One does. Start with the Project. Build the skill once screening is no longer occasional.
The Circle runs six sessions of ninety minutes each, weekly, with an office hour after every session. One founding cohort, three funds, three people per fund. Whichever path a fund starts on, the same thesis spec is what gets tested against live deal flow in Session 2.
The work now begins. More to come.
Talk to Accendo about the next cohort
