The Field Guide · working artifacts
The brief. The gate.
The ledger. The trio.
Delegation, as artifacts.
High-DQ delegation runs on four working documents, and none of them requires technical skill — only management. The forms below are tool-agnostic and stable. The implementations at the bottom are tool-specific and updated as the tools change. That split is deliberate: tools change quarterly; the management layer doesn't.
1 · The stewardship brief
Prompting dictates steps. A brief defines the outcome, the boundaries, the context, and the standard — then leaves the method to the agent. Copy the template; the filled example below it shows one running on a real class of task.
TASK
One sentence. What needs to exist when this is done.
TRIGGER [when this brief deploys]
The condition that fires this brief — so anyone who
recognizes the condition can run it.
DONE LOOKS LIKE [desired results]
- The deliverable, its format, its length
- One example of good output, attached
- The one question this must answer
BOUNDARIES [guidelines]
- In scope / out of scope
- Constraints: tone, audience, deadline
CONTEXT [resources]
- The 2–3 files that matter (curated, not everything)
- What to ignore, and why
CHECKPOINTS [accountability]
- Where I look before you continue
- If blocked: propose 3 options, pick one, flag it
SHIP / SEND BACK [consequences]
- Ships if: matches DONE LOOKS LIKE
- Comes back if: … (with specific feedback)
Filled, for a lost-deal post-mortem:
TASK
A post-mortem the board can act on for the [Company X]
renewal we lost, due before Thursday's meeting.
TRIGGER
Any lost deal or lost renewal above $250k.
DONE LOOKS LIKE
- Two pages maximum: a 90-day timeline, then findings
- Exactly two process failures, each anchored to a direct
quote from the call transcripts or email thread
- One paragraph answering: "what would have had to be true
60 days earlier for this renewal to survive?"
- Last quarter's post-mortem attached as the standard
BOUNDARIES
- Sales blames engineering; engineering blames sales.
Present both claims against the record — do not
harmonize them, do not assign blame
- External factors out of scope unless a quote cites them
- No recommendations about individuals
CONTEXT
- The deal file: transcripts, the email thread from day 60
onward, the original proposal
- Ignore the pricing subfolder — this post-mortem is about
everything except pricing
CHECKPOINTS
- Show me the timeline before writing findings
SHIP / SEND BACK
- Ships if: both failures carry verbatim quotes traceable
to source, timeline reconciles with CRM dates
- Comes back if: any claim lacks a source, or it reads
like a case for either department
Notice what the brief does not contain: no instructions about tone or method. The method belongs to the agent. The standard belongs to you.
2 · The verification gate
Output from an agent arrives fluent and confident whether it is right or wrong — so review is designed, not performed. A gate is written once per task type and reused. This one runs a weekly competitor digest in about six minutes:
GATE — weekly competitor digest stakes: internal, recoverable
Runs on everything (automated / mechanical):
- All six named competitors present; flag any absence
- Every factual claim carries a source link
- No source older than 14 days unless marked [context]
Aimed sample (5 minutes):
- Pick 3 source links at random; confirm each says what
the digest claims it says
- Check the section on [current rival of concern] in full —
history says this is where errors cluster
Escalate to full review only if:
- The digest will be quoted to the board this week, or
- Any sampled citation fails (one broken citation rejects
the document — fix the brief, rerun)
The reading-everything alternative costs forty minutes — and catches less, because none of its minutes are aimed.
3 · The trust ledger
Trust in an agent should be a record, not a feeling. One row per delegated task, ten seconds per row: after twenty rows your autonomy settings stop being a mood and start being evidence. Watch the Adjustment column diverge by task type — that divergence is the entire point.
| Task type | Autonomy granted | Outcome | Adjustment |
|---|
| Competitor digest | unattended, sampled gate | passed sample | hold |
| Competitor digest | unattended, sampled gate | passed sample | hold |
| Contract first-read | checkpointed at clause list | missed a renewal clause | brief: add clause checklist |
| Competitor digest | unattended, sampled gate | passed sample | widen: sample 2 links, not 3 |
| Contract first-read | checkpointed at clause list | clean | hold — one pass is not a trend |
| Competitor digest | unattended, lighter sample | broken citation caught | narrow back to 3; brief: require source dates |
| Contract first-read | checkpointed at clause list | clean | hold |
| Contract first-read | checkpointed at clause list | clean | widen: checkpoint every other contract |
Two task types, opposite directions, every move dated and reasoned. Nothing here is a feeling.
4 · The variant-brief trio
Asking three people to do the same task is insulting and wasteful. Asking three agents is free — and the spread between their answers is information about your brief. One deliverable, briefed three ways, run in parallel:
DELIVERABLE — memo: how to respond to a competitor's price cut
VARIANT A — cost
Objective: protect margin.
Constraint: no response that reduces ARPU more than 5%.
VARIANT B — risk
Objective: protect the base.
Constraint: rank options by churn exposure, not margin.
VARIANT C — speed
Objective: time-to-response.
Constraint: only options executable in 30 days, current team.
Where the variants agree, you have convergent evidence. Where they diverge wildly, you have found — at zero diagnostic cost — exactly where your brief was ambiguous. Fan out, gate, keep the best, and write the missing constraint into the brief for next time.
Tool-specific · updated June 2026
Running this on the Claude stack
The artifacts above are stable. Here is how they map onto current tooling — this section is rewritten as the tools change, which is precisely why it lives here and not in print.
THE BRIEF → a file, not a chat message
Keep briefs as markdown files in a briefs/ folder.
In Claude Code, standing context (your company, your
standards, your bans) lives in CLAUDE.md — written once,
loaded every session. Per-task briefs reference it
instead of repeating it.
THE GATE → a checklist the agent runs before you do
End briefs with: "Before presenting, verify against
DONE LOOKS LIKE and list each criterion as met/unmet."
Then run YOUR gate on what survives. Adversarial check:
a second session briefed only to attack the first
output's claims — fluency against fluency.
THE LEDGER → a plain spreadsheet
Task type · autonomy · outcome · adjustment. No tooling
needed; the discipline is the ten seconds per row.
PARALLEL ATTEMPTS → subagents or parallel sessions
Claude Code runs parallel subagents natively — three
variant briefs, one command, compare at the gate.
At minimum: three tabs, three variants, one decision.
CHECKPOINTS → plan-then-execute
"Show me the outline/plan before proceeding" works in
any chat. In Claude Code, plan mode does this natively.
A rule of thumb for whatever tool you hold: if your instruction would survive being read by a contractor six months from now, it is a brief. If it only makes sense in today's chat window, it is a prompt — and it will evaporate.
Where do you actually stand?
The artifacts raise your DQ. The assessment measures it — twelve scenarios, four minutes, one number.
Take the Quick Score →