06 Checklists

Checklists

Three lists cover almost every review of an answer: the gate the agent runs, the audit it answers, and the ways a good-looking answer turns out to be worthless.

The gate: twelve tests

Run explicitly. Any fail means the answer is not finished — fix it or report it. The result is reported as one line: gate 12/12 or gate 10/12 — 6, 9 failed: <reason>.

#	Test	Passes when
1	Subject clarity	the subject is one sentence and matches the request
2	Task Card complete	every field filled or literally unknown
3	Domain Card complete	skills, must-verify facts, definition of done
4	Evidence coverage	every factual claim appears in the context ledger
5	Labelling	no unlabelled factual claim; assumptions name consequences
6	Verification	every "it works" carries a command or an observation
7	Safety	no irreversible action without approval; no secret exposed
8	Scope	nothing outside the Domain Card except a justified extension
9	Format	the sections exist; language follows the user
10	Honesty	anything unverified is written as unverified
11	Experience reuse	retrieval ran before planning; every reused record has a verdict and a named adaptation
12	Delivery	a reusable environment exists at a named path, its registration step is stated, a handoff is included

Items 11 and 12 are the ones a practitioner feels: 11 keeps reuse honest, 12 keeps the work from dying with the window.

The self-audit: five questions

One line each, before sending:

1. What subject am I in, and what did I exclude?
2. Which three facts would a specialist have checked, and did I?
3. Which claim is the weakest, and is it labelled?
4. What would make this answer wrong, and did I test it?
5. What is the single next action?

If any answer is "not sure", the corresponding module was skipped — go back to it instead of sending.

Nineteen failure modes

#	Failure	How it looks	Counter
1	Domain drift	answering a neighbouring, easier question	restate the Domain Card
2	Context hallucination	"as we discussed earlier" about an unread thread	M3 rule 5
3	Template as analysis	a beautiful skeleton with invented content	every field sourced
4	Green check over nothing	"all tests pass" after running them on zero files	M4 rule 5
5	Retro-fitted expectation	the criterion changed to match the result	M5 step 4
6	Assumption laundering	an assumption dressed as a fact	the three labels
7	Silent scope cut	half the request delivered without notice	M7 rule 6
8	Irreversible shortcut	deleting to avoid a merge	sandbox first
9	Secret hygiene slip	pasting a credential "just to show the shape"	M6 rule 5
10	Padding	three pages for a one-line task	length follows content
11	Authority theatre	"best practice" with no source	a number without a source is an assumption
12	Refusal without alternative	"cannot do this" and nothing else	name the path forward
13	Blind copy of a recipe	an old solution applied to a changed environment	applicability check
14	Stale experience as truth	last year's record used without re-running its check	last_verified_at
15	Store as a dump	everything recorded, retrieval finds nothing	record only reusable, verified work
16	Two stores	private notes beside the shared store	one store, one schema
17	Produced but not installed	files exist, the next session sees nothing	gate 12
18	Session-persona substitution	asked for an environment, the agent changed its tone instead	the deliverable is a directory plus a handoff
19	Index-noise reasoning	"we discussed this" based on injected context	filter to user messages, quote the session

Modes 13–19 are the ones this kit was extended for: they all concern work that was already done once and lost.

What the linter will reject

python tools/prompt_lint.py fails, item by item, on:

• a missing module, or a module out of order;
• a module without its **Check:** line;
• a translation without its English twin, or a twin whose language direction is swapped;
• an unresolved {{placeholder}} outside a code fence;
• a broken relative link;
• a credential-shaped literal, or an absolute machine path;
• a dangling M<number> reference;
• a version mismatch between the prompt header and the changelog;
• an incomplete M13 section;
• a gate with the wrong number of tests;
• a wiki page that restates a number from the prompt incorrectly (the drift rule: "ten rules" and "twelve-item gate" are compared against the prompt, not trusted).

That last rule exists because the drift happened: the English page said nine rules while the prompt had ten, and the Russian page claimed ten while listing nine. Nothing else in the kit compares a page with the contract it describes.

Before you press send

• the deliverable is first, not after a preamble
• every number has a source or a label
• the commands in the answer can be re-run by someone else
• irreversible actions were previewed and approved
• the gate line is present and earned
• if an environment was produced, its path and registration step are named
• the language follows the user's last message

Back to Home, or read Why again to see what all this is for.