carloscape/octorato/investigate-before-asking

Investigate Before Asking (Mandamiento Canónico)

"primero investigo luego pregunto" — operator, 2026-05-21

Promoted to canonical brain rule.

The mandate

Before you ask the operator any clarifying question, spend 30-60 seconds doing read-only investigation that could answer the question itself.

If the answer is reachable by grep, a file read, a list-directory, an MCP read tool, or any read-only API call — you go find the answer first. You only ask when the question is genuinely unanswerable from artifacts visible to you.

Why this matters

Asking questions has a real cost:

• It interrupts the operator.
• It signals you didn't try.
• It slows the loop by minutes per round-trip.
• Most clarifying questions are answerable from artifacts already on disk

(DDs, ER snapshots, prior memories, repo state, SQL files, captures).

• Operators repeatedly granting "look at X" is them doing the investigation

the agent should have done.

How to apply

Before forming the question, ask yourself:

1. Is this answer in a file I can read? (Read, Grep, Glob)
2. Is this answer in a captured artifact? (DD JSON, ER cards, knowledge.json, memory)
3. Is this answer reachable via a read-only API call? (REST GET, MCP read)
4. Is this answer in a memory I haven't checked? (MEMORY.md)
5. Is this answer derivable from comparing two snapshots I already have?

If ANY of those is yes — do that first. Only after the read-only investigation actually fails to settle the question do you formulate the ask.

What you can ask without investigation first

The pattern in practice

Bad:

"What columns does provider.facility have now after Nick's changes?"

Good:

[re-parse SVG, compare to old cards, summarize the 9 column-level deltas]

"Nick removed is_siteofcare_exempt, renamed is_synthetic →

is_sole_proprietor, added is_us_address + county. Here is the

full diff. Should we apply all of these in PR #X, or split?"

The second version costs the agent 30 seconds and saves the operator a context switch. The operator's question is now about direction, not facts the agent could have found.

When investigation reveals the question doesn't need to be asked at all

Frequently the investigation produces a full enough picture that the agent can:

• Propose a concrete plan with manifest, and ask only for go/no-go.
• Identify the question is already settled by an earlier memory / decision.
• See that the "question" was actually a false dichotomy — both options

fail for the same reason, or one is obviously correct given the data.

In those cases, present the finding + recommendation. Don't ask unnecessarily.

Practical 30-second protocol

When you feel a clarifying question forming:

1. Pause. Don't ask yet.
2. Identify the question's data dependency.
3. Run 1-3 read-only calls that would satisfy that dependency.
4. Re-evaluate whether the question still needs to be asked.
5. If yes — ask, but now show the investigation so the operator can
   correct your facts before answering the question.
6. If no — present finding + proposed action.

Anti-patterns

• ❌ Asking "where is X?" when Glob would find it in seconds.
• ❌ Asking "what does file Y look like?" when Read is available.
• ❌ Asking "did Z happen?" when an API call would tell you.
• ❌ Asking the operator to summarize artifacts they already shared.
• ❌ Asking three questions in a row that all share the same data

dependency — investigate once, then ask the one real direction question.

Relationship to other skills

This skill is upstream of:

• progressive-code-exploration — investigation technique for large files.
• session-memory-search — investigation in prior conversation memory.
• workspace-skill-discovery — investigation across arm-level skills.
• gap-analysis-pattern — investigation that diffs documentation vs code.

The principle is the same across all of them: read first, ask only when reading is genuinely insufficient.