Cheat sheet

Everything that decides whether an agent system works.

Not a command reference - the official docs do that better and stay current. This is the judgement layer: how many tools, what belongs in a skill, which memory tier a fact goes in, and what has to wait for a human. One page. Print it.

9 cards · every row links to a guide · free, no sign-up

Tool surface

How an agent reaches your system.

How many tools: ~15 beats 60. Every tool is a distractor for every other one.
Granularity: Wrap the job, not the endpoint. Say it as 'I want to…'
Naming: Verb-first, unambiguous across the whole set.
Description: What it does, when to use it, and when NOT to.
Errors: Say what was wrong and what valid looks like. Enable a retry.
Error classes: Retryable / correctable / terminal. Say which.
Credentials: Server-side only. Never a tool argument.
Too granular?: If one question needs 5 calls, you mapped endpoints.

Full guide →

Memory

What survives between sessions.

Split by: Lifespan, not topic. Topic folders tell you nothing about pruning.
Tier 1: Durable facts. True until explicitly changed. One fact per file.
Tier 2: Daily log. Append-only, dated, never edited. It is evidence.
Tier 3: Working state. Disposable, cleared between tasks.
Corrections: Overwrite in tier 1. Appending leaves the reader guessing.
Loading: Index always; content on demand. 1 line per fact.
Storage: Markdown in git. Vectors only for large unstructured corpora.
Rot: Nothing removes anything by default. Schedule a pruning pass.

Full guide →

Skills

Repeatable work, encoded once.

Contains: Judgement, not documentation. What you'd tell a new hire day one.
Delete it if: Nothing in it is absent from the public docs.
Trigger: The description is the match. Use words people really type.
Test the trigger: Write 5 real sentences. Each must obviously match.
Boundaries: State the negative case: 'not for X, use Y'.
Scope: Narrow. If it opens with a mode question, it's two skills.
Ends with: Verification that runs something real. Not a claim.
Working?: Correction rate falling, not 'it ran'.

Full guide →

Orchestration

Getting work to the right place.

First step: Classify: one-liner / multi-step / unclear / bug. Then act.
Subagents buy: Context isolation. The conclusion, not the file dumps.
Delegate when: You'd have to go looking. Read it yourself if you can name it.
Don't delegate: Single known lookups, or work needing full conversation history.
On failure: Fail loudly and stop. Silent fallbacks hide the cause.
Long sessions: Degrading behaviour = accumulated contradictions. Start fresh.

Full guide →

Identity

Who is actually acting.

Principal: The agent is never the identity. It acts on behalf of someone.
Resolve where: At the edge, before any tool runs. Never model-supplied.
Enterprise set: SSO (SAML/OIDC), SCIM, roles, audit logs.
SCIM matters: Manual deprovisioning is an audit finding, every time.
Scope tools: At the tool layer. Absent beats present-and-checked.
Default: Deny. New capabilities unavailable until granted.
Spend: Hard ceilings server-side. Agents loop; loops cost money.
Audit: Log principal + agent + tool + args + result + approval ref.

Full guide →

The gate

Where a human has to say yes.

Always gated: Published externally / moves money / irreversible.
Sort by: Reversibility, not importance. What can you undo?
Everything else: Ships. A policy that gates all of it is a manual process.
Prompt rules: Not gates. They lose to long context and confident phrasing.
Pattern: prepare() returns a summary + token. execute() needs approval.
Bind token to: The exact payload. £40 approved ≠ £400 executed.
Show: The diff, not the intent. 'Send newsletter?' is not reviewable.
Fatigue: 40 approvals/day = nobody reads. Reduce volume, not scrutiny.

Full guide →

Debugging

Why the output was wrong.

Cause 1: No context. Coherent but disconnected from reality. Most common.
Cause 2: Tool surface misled it. Watch its FIRST choice.
Cause 3: Contradictory instructions. Find the conflict; don't add a third.
Cause 4: Genuine limit. Same failure repeatedly, corrections don't converge.
The test: Could a good engineer succeed given only what the agent had?
Read: The transcript, not the agent's summary. One is evidence.
Change: One thing at a time, or you learn nothing transferable.

Full guide →

Deployment

Getting it past a security review.

The reframe: New vendor = months. New workload in their cloud = a form.
Bedrock/Vertex: Same model. Different auth, billing, region, quota.
Auth win: No long-lived API key. Short-lived, IAM-governed credentials.
Model IDs: Not portable. Never hardcode; config per provider.
Quota: Cloud-side, per region. Pilot quota ≠ production quota.
Have ready: What's sent, retention, training, residency, audit, revocation.
Pilot with: The sceptical team. A glowing report from fans convinces nobody.
Say no when: No compliance driver and no cloud commitment.

Full guide →

Team rollout

Getting past three enthusiasts.

Why it stalls: Social, not technical. Nobody wrote down what good looks like.
Do first: A project instructions file that is actually correct.
Onboarding: Working session with real tickets. Not a demo.
Don't curate: Sessions where things fail teach more than cherry-picked wins.
Capture wins: Same day, as a shared skill. Otherwise nothing compounds.
Measure: Unprompted weekly use. Cycle time. Falling 'how do I…' questions.
Not: Licences activated. Lines accepted (rewards slop).

Full guide →

The six rules underneath all of it

01Context beats prompting, and it is not close.
02Write down judgement, not documentation.
03Measure before you restructure, not after.
04A rule in a prompt is a preference. A rule in the system is a gate.
05Fail loudly. A silent fallback hides the cause until it is expensive.
06Sort by what you can undo.

Every card comes out of something running in production, and every claim is argued properly in the guides. Nothing here is for sale.