DC CAP AI Pilot · Week 4

Three Builds You Can Steal

Three real anatomies. A team Project, a team Skill, and a deeply customized Skill — each one written so you can lift the structure into your own work. The depth of customization is the difference between a build that saves five minutes and one that replaces a whole layer of tribal knowledge.

Each build below shows the same anatomy in four parts: setup (what gets configured), customization vectors (org context, outside authority, workflow shape), governance (who inherits what tier of data), and steal this (the structural move you can lift). Read in order — depth grows from the first build to the third.

Build 01 · Project · Reuses context

Partner Engagement Workspace

A Claude Project loaded with everything Student Success needs to draft, summarize, compare, and brief on the 13 university partner relationships. One workspace, many asks, for months.

Container: Project Owner: Stephanie · Student Success Inheritance: Coaches + leadership Highest tier: Tier 3
Setup · Custom instructions
# Custom instructions (paste into Claude Project) You help DC CAP's Student Success team manage year-round relationships with 13 university partner admissions and financial aid contacts. Tone: collegial, professional, warm. These are people we will work with for years. Always: name the relationship history when relevant; include specific dates and named next steps; check the verified DC CAP numbers in the reference doc before quoting impact data. Never: include individual student names, financial aid amounts, or anything beyond Tier 3 governance unless I explicitly load it for a single conversation.
Setup · Reference files loaded
  • partners_2026.md — 13 partners, primary contact, MOU renewal date, last 3 touchpoints Tier 3
  • dccap_voice_oneliner.md — voice and tone rules from the brand guide Tier 4
  • verified_numbers_FY25.md — completion %, retention %, scholar count, $ awarded Tier 3
  • moil_calendar_2026.md — when each partner expects what kind of touchpoint Tier 3
Customization · Org context
DC CAP voice rules. Verified FY25 numbers. The 13-partner roster with relationship history. Internal calendar of who expects what when.
Customization · Outside authority
Two peer-org partner-engagement playbooks, translated from their published PDFs into peer_playbooks_summarized.md. National FGLI completion baselines transcribed from federal data into fgli_baselines.md. References live as Markdown, never as raw PDFs.
Customization · Workflow shape
Project-level, not per-task. The instructions encode the standing rules; each conversation supplies the new ask. The shape lives in the kinds of asks the Project handles cleanly: drafts, summaries, comparisons, briefings.
What it answers
  • "Draft a Q3 update note to the Howard contact."
  • "What did Georgetown say about timing in February?"
  • "Compare how three partners responded to the cycle delay ask."
  • "Give me a 4-bullet brief for Eric on the Trinity relationship before Friday's call."
Steal this
A Project earns its keep when one workspace handles four shapes of ask without re-loading context.
If your candidate answers just one kind of question — say, only "draft an email" — it's probably a Skill. A Project is justified when the same body of knowledge supports drafts, summaries, comparisons, and briefings. Watch what kinds of asks you're naturally typing into Chat against your team's work over a week. If three or more shapes show up, build the Project.
Build 02 · Skill · Reuses the move

Counselor Cycle Update Email

A Skill the Student Success team can run any time DC CAP needs to send a high school counselor a cycle update — application open, deadline preview, decision notification. Same procedure, different cycle moment.

Container: Skill Owner: Stephanie Inheritance: Any pilot member writing to counselors Highest tier: Tier 4
The SKILL.md (paste this — adapt the body)
--- name: counselor-cycle-update-email description: Drafts a DC CAP cycle update email to a high school counselor — application open, deadline preview, or decision notification. Use whenever you need to write a cycle email any counselor across DC's high schools. --- ## When to trigger Any of: application open, T-2 weeks deadline, T-3 days deadline, decision notification, partnership thank-you. Counselor audience. ## What good looks like - Opens with the relationship — "thanks for the partnership," not "I am writing to inform you." - Names the cycle moment in the first sentence (open, deadline, decision) so a busy counselor can triage in 5 seconds. - One specific date. One specific next action. No "soon." - 4-6 sentences. Always. - Closes with a way to ask follow-up that doesn't require a meeting. ## What to avoid - Equity framing. We frame as accessibility and impact. - Words: unlock, leverage, seamless, robust, journey, elevate. - Plural CTAs. One ask per email. - "Per our records" / "I am reaching out to" / boilerplate openers. ## Inputs the user supplies - Cycle moment (open / deadline preview / decision / thank-you) - Counselor name + high school - Anything specific to this school or counselor (optional) ## Reference DC CAP voice rules: `dccap_voice_oneliner.md` Counselor list + tone notes: `counselor_directory_2026.md`
Customization · Org context
DC CAP voice rules baked into the "what good looks like" and "what to avoid" sections. The counselor directory loaded as a reference, so the Skill knows which schools have multi-year history.
Customization · Outside authority
One source: a published nonprofit communications style guide on counselor-facing tone. Cited in the procedure rather than loaded as a file — it is reference material, not active context.
Customization · Workflow shape
The procedural spine. Five cycle moments. Three rules each for "good" and "avoid." A specific length. One human checkpoint (the sender reviews before send). Anyone can run it; the move stays consistent.
What it produces
A 4-6 sentence email that sounds like DC CAP, names the cycle moment in the opener, gives one date and one action, and closes warm. The team gets identical structure across all 50+ counselors and across the entire cycle.
Steal this
A Skill earns its keep when "what to avoid" is as specific as "what good looks like."
Most weak Skills define the positive ("warm, professional, clear") without defining the failure modes ("never use 'unlock,' never plural CTAs, never 'per our records'"). The negative space is what makes the move repeatable. Write the avoid list before you write the good list — that is where your tribal knowledge actually lives.
Build 03 · Deeply customized Skill · Reuses the move with depth

Preston Writing Voice

A Skill that drafts and edits anything Preston will sign — board memos, LinkedIn posts, donor briefs, policy comments, conference remarks. BLUF structure, banned-word list, banned constructions, voice rules, and a curated archive of Preston's own published writing as outside reference. Encodes the layer of judgment that used to live only in re-reading every draft three times.

Container: Skill Owner: Preston Inheritance: Anyone drafting in Preston's voice; pattern usable by any DC CAP author Highest tier: Tier 3
The SKILL.md (the procedural spine)
--- name: preston-writing description: Drafts and edits any written content in Preston Magouirk's voice — LinkedIn posts, grant narratives, board memos, donor briefs, policy comments, conference remarks. Use whenever Preston will sign or speak the words. --- ## When to trigger Anything Preston will sign. Memos to Eric. Board materials. Funder briefs. LinkedIn drafts. Conference abstracts. Policy comments. Donor thank-yous. ## Structure — BLUF, every time 1. Direct answer first. 2. Strategic context second. 3. Next move third. That order. Even in a three-sentence email. ## Voice rules - Active voice. Make claims. - Evidence-grounded. Every claim that supports the argument names a source. - Short sentences. Period. - The reader can stop after sentence one and still have the point. ## Banned words (no exceptions) unlock, leverage, seamless, robust, tapestry, journey, elevate, streamline, optimize, holistic, delve, foster, vibrant, navigate, empower, utilize, realm, landscape, pivotal, paramount, cutting-edge, synergy, ecosystem, load-bearing. ## Banned constructions - "This isn't X — it's Y." Forbidden. - "Not X, but Y." Forbidden. - Em-dash chains. One per paragraph, max. - Bullet-point prose pretending to be a sentence. - "Genuinely," "honestly," "straightforward." - Equity framing. Frame as accessibility and impact instead. ## What good looks like - The reader can stop after sentence 1 and know the point. - Every number has a source within reach. - The voice is recognizable as Preston's with the byline removed. - No AI tell-tale openers ("Here's a...", "I'd be happy to..."). - No hedging close ("Hope this helps", "Let me know"). ## What to avoid - Paragraphs that explain what the next paragraph will say. - Throat-clearing before the actual answer. - Filler adverbs and hedges that soften a real claim. - Borrowed framings from other people's voices. ## References loaded - `preston_published_writing.md` (50 paragraphs from prior work, Tier 3) - `dccap_voice_oneliner.md` (org voice rules, Tier 4) - `banned_words.md` (the list, kept current, Tier 4) - `BLUF_examples.md` (5 worked examples of the structure, Tier 4)
Customization · Org context (depth)
DC CAP voice rules. Banned-word list maintained as Preston catches new ones in the wild. The accessibility-and-impact framing rule that replaces equity language. The hedging-and-hedging-close failure modes specific to how AI drafts come back.
Customization · Outside authority (depth)
A curated archive of fifty paragraphs from Preston's own published writing — the corpus the voice was actually built on. Strunk & White's economy rules. The BLUF structural pattern adapted from executive-communication standards. Every reference traces to a real source you can read.
Customization · Workflow shape (depth)
BLUF order is enforced at the structural layer — direct answer, then strategic context, then next move, every time. Banned constructions are named precisely (the "not X but Y" rhetorical move that AI drafts love is forbidden by name). The avoid list catches both content failures (filler, hedging) and AI-tell failures (boilerplate openers, throat-clearing closes). This is what depth looks like — failure modes that took a year to notice each get a named home in the file.
What the team inherits
Preston runs the Skill on every draft Claude returns. The voice arrives consistent on the first pass. Anyone Preston routes the Skill to (Cowork, a teammate co-drafting a memo) produces output that survives Preston's review with light edits instead of structural rewrites. The structural pattern — BLUF, banned-words list, banned-constructions list, archive of own work as reference — is portable. Any DC CAP author can lift the shape and fill it with their own voice.
Steal this
Depth means every failure mode you have ever fixed by hand has a named home in the file.
Banned words go in their own list. Banned constructions go in their own list — separate, because they fail differently. The structural rule (BLUF) is named at the top, not buried in prose. The reference archive is curated, not dumped. Walk through your last ten edits to AI drafts and ask: which of those corrections is in the file, and which is still in your head? The corrections still in your head are the next layer to encode.

Side by side

The same anatomy, three depths. Use the table to figure out which build your candidate becomes — and where you should invest the customization effort.

Project
Skill
Deep Skill
What's reused
Context — many shapes of ask against one body of knowledge
The move — a single procedure with varying inputs
The move + every learned layer encoded
Customization depth
Org context heavy. Outside authority light. Workflow at the project level.
Org context modest. One outside source. Workflow tightly defined.
All three vectors deep. Multiple outside sources. Diligence step embedded.
Build time
~1 hour to set up + ongoing reference-file maintenance
~1-2 hours to write the SKILL.md cleanly
~half-day. Often a multi-iteration build.
When it pays off
When you'd otherwise paste the same context into Chat 5+ times a week
When the same move runs across many inputs and people
When the move has high stakes and the team currently leans on one or two experts
Failure mode
Becomes a junk drawer of files no one trusts
Avoid list is too short — drift creeps in
Maintenance lapses; references go stale; outputs quietly drift
Governance lens
Tier of the highest reference file = tier of the Project
Procedure usually Tier 4; references can pull it up
Audit references quarterly. Tier creeps as the build matures.

The point of the gallery

You are not looking at three exemplars to admire. You are looking at three shapes you can adapt. Pick the one that matches your candidate's depth and lift the structure. Then run it through the Build Filter and bring the v1 to AI Friday.