Build the Team Prompt Library and Upskilling Pathway

Tier: Foundation · Workflow area: Quality leadership & governance

This playbook sets up a git-based prompt registry with a card format that makes prompts reusable by people who did not write them, then builds the competency rubric and learning pathway that turns a mixed manual/automation team into confident AI-era testers.

When to use

Prerequisites

• •
  A git repo (or a directory in the QA repo) the whole team can PR into.
• •
  Two weeks of harvesting: ask everyone to save chat transcripts where AI genuinely helped on test work.
• •
  The team's honest skills picture — who automates, who prompts well, who has touched neither.

The workflow

1. 1.

   Define the prompt card format — the library's unit of value. A prompt without context, an example run and known failure modes is a fortune cookie. One file per card:

   Copy

   prompts/test-design/edge-case-ideation.md
   
   # Edge-case ideation from acceptance criteria
   owner: priya    version: 1.3    last-verified: 2026-08-01
   models: claude-sonnet, gpt-4-class   playbook: pb-05
   data-class: internal OK, client code -> use client-approved tool
   
   ## When to reach for this
   AC list exists, you want boundary/negative candidates before
   writing cases. Not for: exploratory charters (see pb-07 card).
   
   ## Prompt
   [the full working prompt, verbatim, with {{placeholders}}]
   
   ## Example run
   [real input excerpt] -> [real output excerpt, trimmed]
   
   ## Known failure modes
   - Invents ACs when the list is short — always ends with the
     "list assumptions separately" clause; check that section first.
   - Weak on time-zone boundaries unless you name the stack's TZ.
   
   ## Review gate
   Output enters the suite only via the pb-01 coverage rubric.
   

   The data-class line wires the library to the policy one-pager, and last-verified is the staleness alarm — prompts rot as models change.

2. 2.

   Seed with a harvest sprint, not a blank repo. Run a 90-minute session: everyone brings their best two transcripts, the group turns the best 10–15 into cards on the spot. Organize by workflow area mirroring this track's structure (test-design/, automation/, triage/, test-data/, ai-systems/) so the library and the training pathway share a map. A library seeded by the team is used by the team; a library seeded by the lead alone is the lead's blog.

3. 3.

   Make contribution cheap and review real. Contribution is a PR with the card template filled; review checks that the example run is real and the failure-modes section is honest (an empty failure-modes section means the card is untested — send it back). Add a monthly 30-minute "prompt clinic": one card demoed live, one stale card re-verified or retired. That single recurring meeting is the difference between a library and an archive.

4. 4.

   Write the AI-QA competency rubric. Four levels, observable behaviors, no vibes:

   Copy

   L1 Aware      Uses library cards as-is with the linked review
                 gates; can spot an obvious hallucination.
   L2 Practicing Adapts prompts to context; catches planted defects
                 in AI output (pb-01-style); documents failure modes.
   L3 Fluent     Designs new prompts + review gates; runs agentic
                 tools (pb-15/16) with correct gating; contributes
                 cards others reuse.
   L4 Leading    Calibrates judges (pb-40), designs eval strategy
                 for AI features, teaches clinics, evolves the rubric.
   

   Publish it with the explicit framing that it is a growth map, not a ranking — the 65.6% anxiety number means every artifact here is read twice: once for content, once for "what does this mean for me."

5. 5.

   Map the team and sequence individual pathways. A 30-minute 1:1 per person places them on the rubric per area (someone can be L3 in test design and L1 in AI-systems testing). Then assign pathways built from this track's own materials — a manual tester's first quarter: pb-01, pb-02, pb-22 (all judgment-heavy, no code), then lab-01; an SDET's: pb-08, pb-13, pb-15, then the pb-34+ AI-systems sequence toward the evaluation-engineer profile. Pair each learner with a card owner in their target area — the library doubles as the mentoring index.

6. 6.

   Instrument lightly, adjust quarterly. Signals worth tracking: cards added/re-verified per month, library citations in PRs and test plans ("used card edge-case-ideation v1.3"), rubric-level movement per quarter, and clinic attendance. Signals not worth tracking: raw prompt counts and any metric that turns cards into KPI fodder — the moment contribution is quota'd, quality dies.

7. 7.

   Close the loop with hiring and onboarding. New-joiner week one includes the library tour and the L1 behaviors; job descriptions reference the rubric levels. This is where the effort compounds — the library converts your team's accumulated AI judgment from folklore into infrastructure that survives attrition.

Variations

• Small team (≤5): skip the rubric ceremony; keep the card format, the clinic, and per-person pathway bullets in one page.
• Enterprise with a platform team: cards can compile into shared assets for approved tools (IDE agent rules files, internal GPT-style presets) — keep git as the source of truth and generate the deployed forms.
• Services org, many clients: maintain a sanitized core library plus per-engagement overlays; the overlay inherits cards and overrides data-class rules per the client's policy.

Quality check

• Every card has a real example run and a non-empty failure-modes sectionthe two fields that separate a library from a prompt dump.
• Cards link to their review-gate playbookno card implies ungated use.
• last-verified dates under 90 days on the top-used cards; stale cards retired visibly, not silently.
• Every team member has a written pathway with a next concrete step and a named mentor.
• The rubric is used in growth conversationsif it only appears in this repo, it is decoration.

Pitfalls

• The write-only library. Capture without the clinic and citation habits produces a beautiful repo nobody opens. The recurring 30 minutes is the load-bearing element.
• Prompt perfectionism. Cards are working notes with version numbers, not published literature — ship v1.0 honest, iterate in PRs.
• Rubric as stack ranking. Tie levels to opportunities (who gets the agentic-stack pilot) and never to blame, or the anxious 65.6% will read it as a layoff list and the initiative dies of silence.
• Pathways only for the already-technical. The manual testers are the population this playbook exists for; their judgment-first sequence (test design, critique, defect intelligence) is where AI leverage is highest and code prerequisites are lowest.