claude-platform-playbook

Claude Adoption Playbook — 90-day rollout

As of 2026-05. Pin to current model surface (Opus 4.7 / Sonnet 4.6 / Haiku 4.5) — refresh monthly.

A 90-day arc from “we want to use Claude” to “Claude is in production with guardrails and a Center of Excellence pattern.” Built for transformation leads, not for engineers — engineers should pair this with claude-code-adoption-guide.md.

🛑 Before Week 0: run candidate use cases through anti-use-cases.md. The “Premature” rows there are the same pre-flight gates this playbook names — but blocking, with cited frameworks. If a candidate hits any row, kill or re-route before scoring.

Week 0 — Pre-flight (before anyone writes a prompt)

Skipping Week 0 is the single most common cause of stalled Claude pilots. Spend 5 working days here.

Decisions to make

Decision Default for most orgs When to deviate
Executive sponsor Named C-level (CIO, CTO, or business unit P&L owner). Not “the AI committee.” Skip only if pilot is a single team, single use case, no cross-functional asks.
Pilot use case One internal-facing workflow with measurable cycle time and a willing team. Use pilot-selection-worksheet.html to score 2–6 candidates on 5 axes (value, speed, data, risk, sponsor) before committing. Don’t lead with customer-facing — too much governance overhead before you have evidence.
Success metrics Cycle time, accuracy/quality, user satisfaction, and cost-per-task. Numeric, not adjectival. If you can’t measure cost-per-task, you’ll lose the budget conversation in week 12.
Procurement path Direct API via console.anthropic.com for speed. Bedrock or Vertex if hyperscaler commits dictate. Bedrock/Vertex add a regional + procurement path but cost a model-version lag.
BAA / DPA Sign before any data flows. Healthcare, finance, EU operations: required before pilot.
No-train confirmation Default for API + console workloads is no-train. Document the policy version + as-of date. Re-verify quarterly; terms drift.
Data residency Choose region per workload sensitivity. EU PII, China, regulated public sector: region constraint may dictate procurement path.

Artifacts produced in Week 0

Common Week 0 mistakes

  1. Picking the use case first, sponsor second. Reverses the political work and stalls in week 4 when funding is needed.
  2. Defining success as “users like it.” Soft metrics don’t survive contact with a CFO.
  3. Letting “the AI committee” replace a single accountable owner. Committees diffuse decisions; pilots need someone whose neck is out.
  4. Skipping the BAA/DPA because “it’s just a pilot.” Legal blocks the production migration if you didn’t paper it up front.

Weeks 1–4 — Pilot (one team, one use case, governance shadow)

Goal: prove the use case works with Claude in your environment, and surface your real constraints before they become production blockers.

Build sequence

Week Workstream Owner Output
1 Use case decomposition: who uses this, when, what data, what output Pilot lead Workflow diagram + data flow map
1 Model + feature selection — start with feature-decision-matrix.html Architect Stack decision (model tier, caching, MCP, Skills, citations)
2 Eval set v0 — 30–80 representative inputs with expected outputs Pilot lead + 1 SME Versioned evalset in the team’s repo
2 First prompt + Skill — minimum viable, not pretty Engineer First passing eval run
3 Cost + latency baseline using cost-calculator.html inputs from real traffic Architect Cost-per-task estimate + monthly $ projection
3 Governance shadow — security/risk reviewer attached to standups, not blocking Risk function Issues log (not gate)
4 User trial with 5–10 pilot users, instrumented Pilot lead Usage logs + qualitative feedback + accuracy delta vs baseline

Pilot exit criteria (gate to weeks 5–8)

If any criterion fails: do not proceed to weeks 5–8. Either iterate within weeks 1–4 (extend by ≤ 2 weeks) or kill the pilot and write the postmortem.

Pilot anti-patterns

Weeks 5–8 — Guardrails (turn the pilot into something safe to scale)

Goal: codify what worked, paper over what didn’t, and make the pattern repeatable for the next team.

Workstreams

1. Eval suite hardening

For the 8-category eval scaffolding (regression, format compliance, tool-call accuracy, grounding, adversarial, cost-per-task, latency, refusal calibration) — each framed by what it catches / failure-mode / owner — see eval-starter-pack.md. Pick 2–3 categories per use case in Phase 1; expand through Phase 2.

2. Cost controls

3. Prompt + Skill versioning

4. Governance + audit

5. Production safety nets

Exit criteria (gate to weeks 9–13)

Weeks 9–13 — Scale (second use case, internal docs, COE pattern)

Goal: prove the pattern repeats, and stand up the structure so the 3rd–10th use case doesn’t need the transformation team’s hand-holding.

Activities

Activity Output
Onboard use case #2 with a different team New evalset + cost projection + governance shadow following weeks 1–4 template
Stand up Center of Excellence (3–5 people, fed by pilot team alumni) Charter, intake form, weekly office hours, templates
Internal documentation — how to start a Claude project One-pager + a build-vs-buy-worksheet.html walkthrough
Training — eng + product + risk 2-hour intro, 1-day deep-dive, ongoing slack channel
Reusable Skills + plugins library Versioned, discoverable, evalled
Shared MCP servers for org-wide systems One CRM connector, one ticketing connector — not 14
Quarterly governance review Risk register update, no-train policy verification, model version refresh, cost trend

Exit criteria (90-day mark)

Common failure modes (8 patterns)

Heatmap — scored

Each failure mode scored on probability (Low / Med / High) and cost-if-hit (★ = ~$1K, ★★ = ~$10K, ★★★ = ~$100K, ★★★★ = ~$1M, ★★★★★ = $10M+ including regulatory exposure). Early-signal column is the observable indicator that fires before the failure mode lands — wire alerting on these, not on the failure itself. Detection-latency column is the gap between early signal and visible damage.

# Pattern Prob Cost Early signal Detection latency Mitigation
1 Pilot purgatory High ★★ Week 4 retro: nobody can name the 2nd use case Weeks Pre-commit 2nd use case in Week 0 charter — see pilot-selection-worksheet.html. Score 2–6 candidates so a backup exists.
2 Eval debt High ★★★ Prompt or Skill change merged without eval run Months Block CI on missing eval pass — see eval-starter-pack.md blocking-vs-advisory matrix.
3 Cost surprise Med ★★★★ Daily $ trending up >20% week-over-week, or single workload >50% of total Days Wire 4 numeric gates ($/task, $/day, cache floor, batch floor) — see governance-overlay.md §15. Hook-enforced, not invoice-discovered.
4 Prompt sprawl High ★★ Two teams shipping similar Skills independently; no shared registry Weeks Canonical Skills library + COE registry by Week 10 — see claude-code-starter-skills.md.
5 Governance afterthought High ★★★★ Risk function not on Week 1 stand-up; no DPA/BAA log Months Embed risk reviewer in Week 1 (advisory not blocking) — see governance-overlay.md. Issues surface early, cheaply.
6 Vendor concentration panic Med ★★ CFO/board ask “what if Anthropic disappears?” in QBR Weeks governance-overlay.md §12 multi-model abstraction at the right layer. Don’t pre-build a 3-model fallback you’ll never use.
7 Model deprecation thrash Med ★★★ Anthropic announces deprecation date for a pinned model Hours COE owns model-bump runbook; pin family not point release; gate on regression eval pass — see eval-starter-pack.md.
8 The “AI committee” tax High ★★★ Decision queue >7 days; no single sponsor name on use case Weeks Single named sponsor with veto — see pilot-selection-worksheet.html sponsor-clarity axis. Committee informs, doesn’t decide.

Scoring posture. Probability and cost are calibrated against post-mortems from public AI rollout failures + the pattern frequencies named in this playbook’s own readers. Re-calibrate quarterly; if your org sees a different distribution, override the scores in your fork. The shape (early signal → detection latency → mitigation) is portable; the specific scores are not.

Detail — symptom + fix

Below: each mode in prose, with the original symptom + fix framing. Use the heatmap above to pick which to monitor first; use the prose below for runbook depth.

# Pattern Symptom Fix
1 Pilot purgatory Pilot succeeds, never scales. No second use case identified. Pre-commit to the 2nd use case in Week 0 charter — even if you swap it later.
2 Eval debt Prompts evolve faster than the evalset. Quality regresses unnoticed. Block prompt changes in CI without eval pass. Owner: pilot lead.
3 Cost surprise Month 4 bill is 5× the pilot’s monthly run rate. Cost dashboard live by Week 6, weekly review. Cap-with-alert per use case.
4 Prompt sprawl Every team writes its own copy of the same instruction set. Skills + plugins library by Week 10. COE owns the canonical versions.
5 Governance afterthought Risk function shows up in Month 5 with blocking issues. Embed risk reviewer in Week 1 (advisory, not blocking). Issues surface early, cheaply.
6 Vendor concentration panic CFO/board asks “what if Anthropic disappears?” Address in governance-overlay.md §12. Multi-model abstraction at the right layer. Don’t pre-build a 3-model fallback you’ll never use.
7 Model deprecation thrash Anthropic rev-bumps; quality moves; nobody owns re-eval. COE owns the model bump runbook. Pin model family, not point release.
8 The “AI committee” tax Decisions take weeks. Nothing ships. Single named sponsor with veto. Committee informs, doesn’t decide.

Reference team structure (3 functions)

You can run a small adoption with 4–6 people. Larger orgs will scale each function but the three roles don’t merge.

1. Build team

2. Platform team

3. Governance + COE

Companion artifacts

© gmanch94 · CC-BY-4.0 · As of 2026-05. Verify model surface + pricing at anthropic.com.

This site is open source. Improve this page.