AI Guides › Workbench
By Nigel Guy · 6 min read
You ask an AI model for a feature, a report or a script, and it hands back the happy path with a note saying error handling, tests and documentation can come "in a follow-up". That feels efficient, because the first answer arrives fast. It is a loan: you pay the missing work back later, usually at a worse moment, and you pay it with your own time.
The "boil the ocean" prompt is a reaction to that habit. Garry Tan, president and CEO of Y Combinator, put the idea into his open-source gstack toolkit for Claude Code. gstack's published builder ethos argues that when AI makes the full version cost minutes more than the shortcut, you should take the full version. It also says the ocean is the destination, reached one "lake" at a time, where each lake is a unit you can finish. That second half is the part social-media summaries drop, and it is the part that stops the prompt hurting you.
The rule: demand completeness inside a fence you drew, one finishable unit at a time, and make the model report what it left out.
gstack's README and ethos document, on GitHub, do use the "boil the ocean" language and the lake-and-ocean distinction. I could not verify the exact wording of any prompt Tan runs day to day, so treat the prompts below as my own rebuilds of the idea, not quotations. I also could not verify the productivity claims attached to gstack, and I have not repeated them.
| Prompt | What it does | Cost | Best for | Catch |
|---|---|---|---|---|
| 1. The Full-Unit Brief | Asks for the complete version of one bounded task: errors, tests, docs, edge cases | Free to use; spends more of your usage allowance per task | Coding, scripts, spreadsheets, templates | Without a fence it sprawls |
| 2. The Lake Map | Splits a large goal into finishable units before any work starts | Free; one short extra turn | Anything bigger than one sitting | You must approve the map, or it is theatre |
| 3. The Leftovers Ledger | Makes the model list what it skipped, deferred or assumed | Free; adds a short table to each answer | Reviewing finished work | Only as honest as its self-report, so spot-check it |
At time of writing none of these needs a paid add-on. They are text. What changes is usage: a fuller answer is longer and may use more tool calls, so on a metered plan it consumes more of your allowance. Check your plan's current limits rather than assuming.
Use it when one task has a clear finish line. Fill in the task, the done-definition, the fence, and the stack or format.
You are a senior practitioner doing finished, handover-quality work.
TASK: [ONE_BOUNDED_TASK]
DEFINITION OF DONE: [WHAT_A_COMPLETE_RESULT_LOOKS_LIKE]
CONTEXT: [STACK_TOOLS_FORMAT_AUDIENCE]
OUT OF SCOPE: [WORK_THAT_IS_RELATED_BUT_NOT_PART_OF_THIS_TASK]
Approach:
1. If any input above is missing or ambiguous, ask me before starting. Do not guess.
2. Where a complete solution costs only a little more effort than a shortcut, build the complete one.
3. "Complete" means all of the following, for this task only:
- failure paths handled, not only the success path
- checks or tests that prove the result works
- short notes a newcomer would need to use or maintain it
- every edge case you can foresee, handled rather than merely mentioned
4. Anything in OUT OF SCOPE stays out. If you spot other worthwhile work, list it under "Spotted, not done" with one line each. Do not do it silently and do not drop it silently.
Before you answer, check: did I cover each item in step 3? Did I touch anything outside the fence?
Output format:
- The finished work
- A "How I checked it" section
- A "Left out, and why" section (write "nothing" if so)
Use it as written for a self-contained job. The OUT OF SCOPE line is not decoration. Anthropic's own prompting guidance notes that some Claude models tend to overbuild, adding extra files, abstractions or flexibility nobody requested, and recommends explicit scope instructions to keep that in check. A completeness prompt with no fence pushes the model in exactly that direction.
Use it before a goal that is too big for one answer, such as "rebuild our onboarding emails" or "add accounts to this app". The ocean is the goal; this prompt finds the lakes.
You are a delivery lead planning finishable work.
GOAL: [THE_BIG_OUTCOME]
CURRENT STATE: [WHAT_EXISTS_TODAY]
CONSTRAINTS: [DEADLINE_BUDGET_TOOLS_PEOPLE]
Split the goal into units that each meet three tests: it can be finished completely in one working session, it can be checked on its own, and it leaves things working if I stop afterwards.
Output a table with these columns: unit number, unit name, what "complete" means for it, how I will check it, depends on, rough size (small / medium / large).
Then recommend which unit to do first and why. Mark any unit that is really two units in disguise.
Ask me about any missing input rather than inventing it. Do not start any unit until I approve the map.
Approve the map yourself. Cut anything you do not need, then paste one unit into Prompt 1 as the TASK.
Use it after any piece of work, from any model, including work produced without Prompt 1.
Act as a strict reviewer of the work above. Do not defend it.
Produce a ledger with one row per item, using these columns: item, status (done / partial / skipped / assumed), why, and what it would cost me to close.
Cover these areas: error and failure handling, tests or checks, documentation, edge cases, assumptions you made without asking me, and anything you were unable to verify.
Rules: no row may be vague ("various edge cases" is not acceptable). If you cannot tell whether something was handled, mark it "unverified". End with the single item most likely to hurt me if I ignore it.