Optional reference

Engineering reference

Use selectively

Cost derivations, a workflow sketch and the invoice-check skill.

Check before use

Product recipes reflect documentation checked 3 October 2026.

Return to the 30-slide main talk

Further exploration

Watch the mechanism, then inspect the code

Interactive explanations

Transformer Explainer
LLM Visualization

Build a small model

Neural Networks: Zero to Hero
Follow a small model from training to generation.

Optional learning links. A small teaching model does not represent every production architecture.

Context reference

Inspect what actually enters a request

Standing instructions and applicable project rules.
Tool definitions and relevant skill content.
Retained conversation, selected files and tool results.
The current user message.

Use a request inspector or provider token counters for real sizes. Avoid summing overlapping categories.

Worked arithmetic

Growing history can produce quadratic input volume

Assume n calls; initial input B; each call adds d tokens. No pruning, compaction or window cap in this toy model. Input at call k = B + (k − 1)d Total input = nB + d × n(n − 1) / 2 B = 5,000; d = 2,000; n = 10 Last call = 23,000; total input = 140,000 tokens.

This is an input-volume model. Apply actual cache, output and tool rates to calculate the bill.

Worked arithmetic

Include the write before counting savings

Same eligible prefix; all later requests hit the cache. Normalize ordinary input cost to 1 per request. Total cached cost = w + (n − 1)r Caching saves when w + (n − 1)r < n. Example r = 0.10: w = 1.25 → savings at 2 requests w = 2.00 → savings at 3 requests

Illustrative multipliers: verify model, platform, TTL and expiry rules. A cache miss changes the calculation.

Workflow reference

Make the retry edge explicit

PSEUDOCODE — not a harness API findings = investigate(rule, calculation, export) plan = reconcile(findings, acceptance_cases) latest_failures = [] for attempt in range(3): patch = implement(plan, latest_failures) result = run_acceptance_checks(patch) if result.passed: return request_review(patch, result) latest_failures = result.failures return report_blocked(latest_failures)

Use enforced tool, time and cost limits around the workflow. Assign an owner to integration.

Skill reference

An invoice-check skill for Claude Code

--- name: invoice-check description: Verify changes to invoice totals or CSV export. --- 1. Read the invoice rule and inspect the proposed diff. 2. Run the invoice and export tests; keep the actual output. 3. Check VAT, rounding and CSV escaping coverage. 4. Report findings, exit status and untested cases.

Example path: .claude/skills/invoice-check/SKILL.md. Check discovery in a fresh session.

Further exploration

Read the procedures before adopting the system

Matt Pocock’s skills

Source repository
Inspect how each skill defines its trigger, procedure and output.

Compound Engineering

Source repository
Inspect how planning, implementation and captured lessons connect.

Optional source examples. Commands and package contents change; adapt and test the parts you choose.

Publication reference

Package, test, then share

my-plugin/ .claude-plugin/plugin.json skills/invoice-check/SKILL.md 1. Give the manifest a name and version. 2. Test locally: claude --plugin-dir ./my-plugin 3. Review contents and pin the version to distribute. 4. Publish through your team’s approved channel.

Plugin overview · Marketplace setup

Sources

Questions? Check the evidence.

Mechanisms

Attention Is All You Need
Context engineering
Context-length experiments

Implementation details

Prompt caching and usage
Subagents and forks
Skills

Primary sources checked 3 October 2026. Worked examples are illustrative. All decks.