Preflight is a session initiation checklist for AI-assisted / agentic work.

It is the counterpart to Landing the Plane (session shutdown): where landing focuses on cleanup + handoff, preflight focuses on scoping + context selection + verification setup so the session doesn’t drift.

preflight-the-plane.svg

When to run it

Run preflight when:

  • you start a new chat (especially after a reset)
  • you resume after time away
  • you’re about to start an autonomous “loop” (background run / repeated iterations)
  • you notice context pollution and decide to restart cleanly

The Problem

Without a preflight:

  • the session starts with ambiguous intent
  • the agent invents defaults and “helpful” scope expansions
  • there’s no shared definition of done
  • verification is an afterthought (“looks right”)
  • you waste the first 10–20 minutes rebuilding context from memory

The Preflight Checklist

1) Pick one unit of work

State the task in one sentence.

Add:

  • what’s explicitly out of scope
  • what “done” means (observable outcome)

(If you can’t do this, the session will likely mix concerns.)

2) Rehydrate only the minimum context

Decide what the agent should read first:

  • the smallest set of files/docs that define intent and constraints
  • the current “state” artifact (plan/status/progress) if one exists

Rule of thumb: prefer a few high-signal artifacts over “here’s everything”.

3) Establish ground truth (verification first)

Before implementing, agree on:

  • the command(s)/checklist that will prove progress (tests, build, smoke check, lint, or a manual verification script)
  • what output counts as success vs failure

This turns “done” into something the workflow can actually detect.

4) Write a tiny plan (and stop)

Produce:

  • 3–7 steps max
  • the first step should be a small, reversible move
  • call out the first verification point (“after step N, run X”)

Then pause for confirmation before large code changes.

5) Set guardrails for this session

Pick the constraints that prevent common failure modes:

  • timebox (e.g., 25–60 minutes)
  • “ask-first” actions (renames, deletions, major refactors)
  • rollback strategy (checkpoint/commit frequency)
  • if doing autonomous work: scope limits and where the agent may write

Preflight for Autonomous Loops (extra)

If you’re starting a loop (iterative agent runs), preflight must include:

  • Exit criteria: an unambiguous condition (tests green, specific string promise, checklist complete)
  • Objective checks: the loop must run the checks, not just claim it did
  • Max iterations / circuit breaker: stop after N attempts
  • State file: a progress.md/ledger/checklist the loop reads and updates each iteration
  • One change per iteration: reduce blast radius and make rollback possible

A trigger phrase (human → agent)

“Preflight this session:

  1. ask clarifying questions until the task is one unit of work with explicit done criteria,
  2. list the minimal files/docs you need to read,
  3. propose a 3–7 step plan with verification points,
  4. propose guardrails (ask-first actions, checkpoints), then stop and wait for approval.”

References