back to slides
Workshop reference · session 1 · slide 20 deep-dive

Inject your team — the 3-layer context card

A working RCTCO prompt scores ~18/25. The same prompt with your team injected scores ~22/25 — without rewriting a single line of the prompt itself. The leverage isn't a better prompt; it's externalizing the tacit knowledge you already carry. Three layers: vocab, standards, examples. Fill once. Paste forever.

Sources cited inline. Last updated 2026-05-02 · WebMD HS workshop · author: Tony Bùi
jump: Layer 1 · Vocab Layer 2 · Standards Layer 3 · Examples Assemble Sources
Layer 1
Vocab · how you talk
5 user roles. Product names. Banned phrases. Tooling names. Cross-language pairs. 5 min to extract.
Layer 2
Standards · the bar
DoD checklist. AC format. PRD section order. Compliance flags. 10 min to extract.
Layer 3
Examples · the magic ingredient
ONE sanitized past artifact. Shape, not content. Beats every rule you could write. 10 min to sanitize.
Layer 1 — Vocab
how your team actually talks

Members not users. WebMD ONE not "the platform." YouTrack ticket not Jira issue. The lexicon your team enforces in YouTrack search, in Confluence, in handoffs — externalized so AI stops drifting toward generic SaaS vocabulary.

Inherits from terminology management in technical writing and localization — the discipline behind ISO 30042 (TBX), the OASIS DITA standard, and DocBook glossary discipline. Every mature documentation team maintains a controlled vocabulary because inconsistent terms break search, traceability, and translation. The same mechanic now applies to AI output.

The AI-specific failure: even with a Role specified, models default to vocabulary that's overrepresented in training data — generic SaaS words like "user", "customer", "platform", "ticket". A senior PM specifying "wellness coach" gets back "clinician" or "therapist" because that's what most healthcare text on the open web says. Specifying the role does not override the prior; explicit vocab does.

The compounding cost: every wrong term in output is a manual rewrite, a lost YouTrack search hit, a Confluence link that doesn't resolve, and slow vocabulary drift across the team. Over a quarter, the cost of NOT having a vocab card is measured in re-edited tickets and member-facing copy that fails brand review.

  • List the 5–7 user roles your team names Open your last shipped PRD. Underline every persona word: member, health coach, HR program manager, employer, account owner. Write down what your team calls them — not what the textbook calls them.
  • List your products and acronyms Full names + the short forms you actually use. WebMD ONE, TINYpulse, embody, BAA (Business Associate Agreement), HRA (Health Risk Assessment). 3–5 acronyms is enough; pick the ones a new hire would Google.
  • List 3–5 banned phrases The "never say this" set. Examples: never imply medical advice in member-facing copy; never use "Oops!" in error messages; never call members "patients"; never use "powered by" in co-branded surfaces. The negative space is more useful than the positive.
  • List your tooling vocab Most people remember domain words and forget tools. Be specific: YouTrack ticket (not "Jira issue"), Confluence space (not "Notion page"), Figma frame (not "screen"), sprint cadence: 2 weeks. AI defaults to whatever's most common in its training set — usually Jira.
  • Add cross-language pairs if you work in 2+ languages EN ↔ VN: member = thành viên, health coach = huấn luyện viên, HR program manager = quản lý chương trình HR. Critical for YouTrack search where bilingual tickets need consistent terms.
### Layer 1 — How my team talks Products: WebMD ONE (web + mobile member app), TINYpulse (engagement pulse surveys), embody (coaching workflow tool). User roles (use exactly): - member — the employee enrolled in the program. Never "user", "patient", "customer". - health coach — the licensed coach assigned to a member. Never "clinician", "therapist", "advisor". - HR program manager — the buyer-side admin. Never "HR admin", "client". - employer — the contracting company. Never "customer", "tenant". - account owner — internal CSM. Never "rep". Acronyms: - BAA = Business Associate Agreement (HIPAA contract) - HRA = Health Risk Assessment - EAP = Employee Assistance Program - PHI = Protected Health Information Tooling: YouTrack tickets (not Jira). Confluence spaces (not Notion). Figma frames (not screens). 2-week sprints. PRDs live in Confluence under Product / WebMD-ONE / PRDs. Banned phrases: - Never imply medical advice in member-facing copy. - Never use "Oops!" or "Sorry!" in error messages — use neutral, instructional tone. - Never use "powered by" on co-branded employer surfaces. - Never write "patient" — we serve members, not patients. Cross-language pairs (EN / VN): - member / thành viên - health coach / huấn luyện viên sức khỏe - HR program manager / quản lý chương trình HR - assessment / đánh giá (never "kiểm tra")
Notice: this is ~30 lines, not 300. The discipline is curation. Your future self will paste this hundreds of times — every line that doesn't earn its place adds friction.
  • Including obvious terms No need to teach AI "user story", "sprint", "backlog". Stick to terms that distinguish your team from a generic SaaS team. If it's in the Scrum Guide, skip it.
  • Missing cross-team aliases Sales calls them "clients". Product calls them "members". Engineering calls them "users" in code. Pick ONE for AI output and explicitly ban the others — otherwise AI mirrors whichever appears most in your context.
  • Forgetting tooling vocab Most people remember domain terms (member, coach) and forget tools. AI will write "create a Jira ticket" because Jira dominates the training data. Explicitly say YouTrack or you'll fix it in every output.
  • Not updating quarterly Vocab evolves faster than people notice. New product names ship; acronyms get retired; banned phrases shift after a brand review. A vocab card from 6 months ago drifts silently — schedule a quarterly 15-min review.
Layer 2 — Standards
the bar your tech lead enforces

The Definition of Done your tech lead enforces. The AC format your team writes in. The PRD section order your reviewers expect. The compliance flags your auditors look for. The standards you follow without thinking — externalized so AI follows them too.

Inherits from the Definition of Done as a Scrum core practice (Schwaber & Sutherland, Scrum Guide) and from the Given-When-Then formalism that Dan North introduced with BDD in 2006. Both exist for the same reason: tacit standards become visible and contestable when written down. AI brings the same need into sharper focus, because AI doesn't carry your team's tacit knowledge.

Specific evidence on AI-generated specs: arxiv 2507.15157 shows LLM-generated user stories meet AC quality criteria less often than human-written ones, regardless of model scale. ACs from AI tend to be vaguer and miss edge conditions. The fix isn't a smarter model — it's an explicit DoD passed in context.

The tacit standards trap: senior PMs follow standards they never wrote down. They reject specs without articulating the rule. AI cannot inherit a rule that doesn't exist on paper. Externalizing the DoD is the same exercise that helps a new hire ramp — AI is a permanent new hire that resets every chat.

  • Open your last 3 shipped PRDs your tech lead approved without changes The "approved without changes" filter is the trick. These are your team's actual standard, not the aspirational one. If the lead asked for changes, those changes ARE the standard.
  • List the structural sections that appear in all 3 Likely: Problem · Goals · Non-goals · User stories · AC · Edge cases · Open questions. The intersection across 3 PRDs is your real PRD shape. Sections that appear in only 1 are situational, not standard.
  • What does your tech lead ALWAYS reject specs for? That's your DoD. Examples: "no AC in G/W/T → reject", "no edge case list → reject", "PHI touched but no compliance flag → reject", "no QA test plan → reject". The rejection patterns ARE the DoD. Write them down.
  • What's your AC format? Pick one: Given / When / Then (BDD), plain bullets, table form. Don't list options — pick the one your team actually uses. Format consistency is what makes AI output droppable into your tooling without rewrites.
  • What compliance flags do you carry? For WebMD HS: [COMPLIANCE — HIPAA], [COMPLIANCE — PDPL], [COMPLIANCE — SOC2], [A11Y — WCAG 2.1 AA]. List the exact tag form your auditor recognizes. AI will use whatever you specify.
### Layer 2 — Our quality bar Definition of Done (every spec MUST satisfy): - AC in Given/When/Then form, one row per scenario - 1 mockup link (Figma frame URL, not screenshot) - Edge case list: empty / failure / permission / concurrency / offline - Compliance flag if PHI or PII touched: [COMPLIANCE — HIPAA] or [COMPLIANCE — PDPL] - A11y flag if member-facing UI: [A11Y — WCAG 2.1 AA] - QA test plan section (3+ scenarios) - Every assumption tagged [ASSUMPTION: ...] - Every open question tagged [TBC: name/role to confirm] AC format (use exactly): Given When Then And PRD section order (do not reorder): 1. Problem statement (3-4 sentences, member impact first) 2. Goals + Non-goals (bullets, non-goals are required) 3. Target users (primary + secondary, role names from vocab) 4. User stories (INVEST, table form) 5. Acceptance criteria (G/W/T, table per story) 6. Edge cases (empty / failure / permission / concurrency / offline) 7. Compliance + accessibility flags 8. Open questions 9. QA test plan Compliance flag taxonomy: - [COMPLIANCE — HIPAA] — any PHI read/write/transmit - [COMPLIANCE — PDPL] — VN personal data, especially cross-border transfer - [COMPLIANCE — SOC2] — audit log requirements - [A11Y — WCAG 2.1 AA] — member-facing UI components - [BAA-required] — third-party integration touching PHI Tech lead's auto-reject patterns (do NOT ship with these): - AC written as prose instead of G/W/T → reject - PHI mentioned without compliance flag → reject - Edge case section missing or empty → reject - "TBD" without owner name → reject (use [TBC: name] instead)
The "auto-reject patterns" block is the negative space. It's the most useful part — it tells AI what NOT to do, which is harder to infer than what to do.
  • Generic DoD instead of specific "Must be tested" is meaningless to AI (and to humans). "Must have G/W/T AC + 1 mockup link + edge case list + compliance flag if PHI touched" is enforceable. Specificity is the difference between a card that works and one that doesn't.
  • Listing aspirational standards If your team doesn't actually enforce a rule, don't put it in the card. Aspirational standards make AI output that no human follows — the gap erodes trust in the card itself. List only what your tech lead truly rejects PRs for.
  • Mixing PM standards with engineering standards Engineering's DoD ("80% test coverage", "no console.log") belongs in the eng team's card, not yours. Stay in your scope. The card you write is the card you maintain — keep it tight.
  • Not capturing rejection patterns The negative space — what your tech lead REJECTS — is more useful than the positive list. "AC as prose → reject" tells AI more than "AC should be G/W/T". Both belong, but the rejection patterns are usually missing.
Layer 3 — Examples
the magic ingredient

ONE sanitized past artifact you'd send to a new hire as "this is what good looks like." Concrete examples beat any rule. The shape — section order, edge case patterns, compliance tags, [TBC] markers — sets AI's output distribution more reliably than any abstract instruction.

Inherits from few-shot prompting — Brown et al. 2020, the GPT-3 paper "Language Models are Few-Shot Learners". The empirical finding has held across model generations: a small number of well-chosen examples beats lengthy abstract instructions at conditioning model output. Examples set the distribution; rules nudge it.

Specifically on in-context learning mechanics: Min et al. 2022, "Rethinking the Role of Demonstrations", shows that the SHAPE of demonstrations (label space, format, input distribution) matters more than the content correctness. Translation: AI learns "PRDs in this team look like THIS" from one example faster than it learns from 10 rules describing what a PRD should look like.

Why ONE good example beats 5 generic ones: overspecification dilutes signal. Each additional example competes for attention with the others; the model averages across them. ONE strong, typical example creates a clearer target than five mediocre ones. The sanitization discipline — keep the SHAPE (section order, AC form, flag taxonomy, [TBC] markers), replace identifiers — preserves the signal that matters.

  • Pick 1 past artifact you'd send to a new hire as "this is good" PRD section, YouTrack ticket your dev called "perfect handoff", research synthesis your PM thanked you for, member-facing copy that passed brand review on first try. ONE artifact. The selection itself is the work.
  • Sanitize: replace real names with bracketed placeholders Real members → [Member A]. Real coaches → [Coach 1]. Real customers → [Employer Co]. Real product feature names from unreleased work → [Feature X]. Strip identifiers; keep structure.
  • Keep all structural elements intact Section order, formatting, compliance tags, edge case patterns, [ASSUMPTION] markers, [TBC] markers, table shapes, header levels. The structure IS the signal AI learns from.
  • Trim to 1 page max More content = AI distracted by content vs structure. If your example is 5 pages, pick the most representative 1-page section. The model picks up the SHAPE faster from a focused example than from a sprawling one.
  • Mark it explicitly: "[EXAMPLE — for shape only, not for reuse]" Prevents AI from copying example content into new outputs (a real failure mode — see context-window contamination in the Drafter pattern). The marker tells AI: imitate the form, not the content.
### Layer 3 — One example of past good work [EXAMPLE — for shape only, not for reuse] ## Problem [Member A]-type members enrolled in [Program X] drop off after their first coach session at 38% (last 90 days, n=412). The drop coincides with the 7-day gap between session 1 and session 2 — no touchpoint in between. ## Goals - Reduce session-1 → session-2 drop from 38% → 25% in Q3 - Maintain coach utilization within current band ## Non-goals - Automated coaching content (out of scope for this sprint) - Member-to-member messaging (separate roadmap item) ## User stories | # | As a | I can | So that | AC (G/W/T) | Flags | |---|------|-------|---------|------------|-------| | 1 | [member] | receive a check-in nudge between sessions | I stay engaged | see AC-1 | [COMPLIANCE — HIPAA] [A11Y — WCAG 2.1 AA] | | 2 | [health coach] | see member's nudge engagement before session 2 | I tailor the conversation | see AC-2 | [COMPLIANCE — HIPAA] | ## Acceptance criteria AC-1 — member nudge: Given a [member] who completed session 1 in the last 5 days When 72 hours have passed since session 1 AND they have not opened the app Then send 1 push notification + 1 in-app card on next open And do not send if the member has opted out of program nudges [COMPLIANCE — HIPAA: notification content must not include PHI] ## Edge cases - Empty: member has no upcoming session 2 booked → suppress nudge - Failure: push delivery fails → fall back to in-app card only - Permission: member opted out of nudges → respect; do not log opt-out events as engagement - Concurrency: session 2 booked DURING the 72h window → cancel pending nudge - Offline: nudge queued; deliver on next online state within 24h or expire ## Open questions - [TBC: legal — confirm nudge text is not "treatment communication" under HIPAA] - [TBC: [Coach Lead] — confirm coaches want engagement signal in their session-prep view] ## QA test plan - 3 scenarios per AC, run on staging with synthetic [Member A] data - Verify no PHI in push notification payload (privacy QA, separate pass)
What this example teaches AI: section order, table shape for stories, G/W/T form for AC, edge case categories (empty/failure/permission/concurrency/offline), how compliance flags attach inline, what a [TBC] looks like with an owner. The model picks all of that up from one read.
  • Picking your most ambitious example AI matches AMBITION, not average. Your award-winning PRD will produce AI output that aims for award-winning every time — bloated, over-elaborated, with sections you don't normally write. Pick a TYPICAL good example, not your best one. The card should produce your team's median good work, not its peak.
  • Skipping sanitization Leaks customer data into your prompt history (and into the model provider's logs depending on plan). For HIPAA-touching examples this is a compliance event. Sanitize first, paste second — every time. No exceptions.
  • Pasting too many examples 5 examples doesn't give you 5x signal — it gives you noise. The model averages across them and produces output that resembles none of them. ONE strong example is worth 5 mediocre ones. If you must include more, scope each to a different artifact type (PRD example + ticket example), never two of the same kind.
  • Forgetting to update when team standards evolve When your team adopts a new edge case category, retires a flag, or shifts AC format, the example needs to evolve too. Stale examples teach AI old shape. Refresh the example every time you ship a new "best of breed" artifact — or quarterly, whichever comes first.
Assemble the context card
paste-once template

Stack the 3 layers into one block. Paste this entire card as your FIRST message to Coach (or any new Claude chat). Coach recognizes PERSONALIZE mode and loads your team. Every prompt that follows scores against YOUR standards, not generic ones.

TEAM CONTEXT — read this before responding: ### Layer 1 — How my team talks - Product names: [e.g., WebMD ONE / TINYpulse / embody] - We say "member" not "user" / "patient" - We say "health coach" not "clinician" / "therapist" - We say "HR program manager" not "HR admin" - Our ticket system: [YouTrack / Jira / Linear] - Our PRD home: [Confluence / YouTrack KB / Notion] - Internal acronyms: [list 3-5 your team uses, e.g., "BAA = Business Associate Agreement"] - Sprint cadence: [2-week sprints / weekly / continuous] - Banned phrases: [e.g., never use "Oops!" in member-facing copy; never imply medical advice] - Cross-language pairs (if applicable): [member = thành viên, etc.] ### Layer 2 — Our quality bar - Definition of Done (paste your team's actual DoD checklist): [- AC in G/W/T format - 1 mockup link - Compliance flag if PHI/PII touched - Edge case list (empty / failure / permission) - QA test plan] - AC format we use: [Given / When / Then] - PRD section order we follow: [Problem → Goals → Non-goals → User stories → AC → Edge cases → Open questions] - Compliance flags we use: [COMPLIANCE — HIPAA] / [COMPLIANCE — PDPL] / [COMPLIANCE — SOC2] - Always mark assumptions: [ASSUMPTION: ...] - Always mark unknowns: [TBC: question for name/role] - Tech lead auto-reject patterns: [list what your lead always rejects for] ### Layer 3 — One example of past good work [EXAMPLE — for shape only, not for reuse] [PASTE one of: - 1 past PRD section that shipped clean - 1 past YouTrack ticket your dev called "perfect handoff" - 1 past research synthesis your PM thanked you for - 1 past member-facing copy that passed brand review Sanitize: replace real member names with [Member A], real customer names with [Employer Co]. Keep the SHAPE intact. Trim to 1 page max.] ### Constraints (always apply) - Never invent product names not listed above - Never invent regulations not in this context - Mark every assumption with [ASSUMPTION: ...] - Mark every compliance touch with the appropriate flag - Use my team's vocabulary — never default to generic "user/patient/clinician" - If I ask for a PRD section, follow our PRD section order exactly - If I ask for tickets, match the shape of my example - If a fact is not in this context, mark [TBC] instead of inventing
Three places to use this: (1) paste as first message to Prompt Coach — Coach loads PERSONALIZE mode; (2) paste before any prompt in any Claude chat; (3) save as team-context.md, upload to a Claude Project — never paste again. The card grows with your team.

Maintenance — the card grows with your team

A stale context card produces stale output. Schedule 15 minutes per quarter, plus event-driven updates:

  • Update quarterly — every Q1/Q2/Q3/Q4 review week
  • Update when your DoD changes — new compliance requirement, new accessibility threshold, new flag taxonomy
  • Update when you ship a new "best example" — replace Layer 3 with the better artifact
  • Update when team vocab shifts — new product name, new role title, new banned phrase from brand review

Key sources cited

Workshop · WebMD Health Services Vietnam · 04/05/2026 tony · vibery.app