Blindspot Pass — Blind-Spot Investigation
Before implementation, find the gap between the map (plan/prompt) and the territory
(actual codebase, domain, constraints).
Origin: Blindspot Pass — see skills/loop/references/talk-source.md
Iron Rules
- Do not implement yet. No code changes while this skill is active.
- End goal of the investigation: help the user write a better prompt.
Procedure
- Parse task description and context sources from
$ARGUMENTS. Empty → the task
currently under discussion; if there is none, ask one question first. If sources are
given (git history, docs, specific modules, etc.), investigate those first.
- Scope codebase-wide or large → spawn the
unknowns:unknowns-scout agent with the
Agent tool. Hand it a packet: the user's original prompt verbatim, the context
sources, the target area and its entry points, and what to return (finding table +
improved prompt draft); map its rows onto the card kinds below.
Scope is a few files → investigate directly with Read/Grep/Glob.
- Compile findings, each tagged with a kind:
- Landmine — touching this breaks something non-obvious (regression risk, fragile
or missing tests, a module mid-migration)
- Convention — an unwritten rule the codebase enforces
- Missing concept — a mechanism the user's prompt has no word for
- History — an earlier or reverted attempt at this exact task
Also collect the questions to answer and the information still needed to
sharpen the prompt.
- Lead with the contrast: What you asked for (the task as the user framed it, and why
it sounds small) vs What you're actually walking into, with a tally by kind
("4 landmines, 2 conventions, 1 missing concept, 1 reverted attempt"). Then the cards —
kind, finding, why it bites, recommended action — sorted by importance × impact.
- End with an improved prompt draft reflecting the findings — the core deliverable of
this skill. It names the execution order and ends with an explicit checkpoint
("stop and show me the plan before writing code").
- Offer with one AskUserQuestion: proceed with this prompt now / edit it first / stop here.
- If undecided items that could change architecture surface, suggest continuing with
/unknowns:interview (or /interview for copied installs). If the user lacks the
unfamiliar domain's vocabulary itself, suggest /unknowns:teach-me.
Output
Artifact tool → publish the page; else .unknowns/<YYYY-MM-DD>-blindspot-<slug>.html; else markdown.
Reaction control: a copyable prompt fix per card; selected fixes assemble into the improved prompt draft.
Details: skills/loop/references/output-routing.md
Scope
Not limited to application code — specs, migrations, infra and vendor integrations all
have landmines and unwritten conventions worth mapping before you touch them.
The subject is always a specific codebase, plan or system and its blind spots.
If the gap is the user's vocabulary for an unfamiliar field, that is
/unknowns:teach-me (or /teach-me for copied installs), not this skill. Running both
is fine: teach-me first for the words, blindspot for the territory.
1---2name: blindspot3description: Blind-spot pass — investigate unknown unknowns before implementation and turn them into a better prompt. Use when the user says "blind spot pass", "what am I missing", "unknown unknowns", or before work in an unfamiliar codebase area. Not for routine, well-understood tasks.4---56# Blindspot Pass — Blind-Spot Investigation78Before implementation, find the **gap between the map (plan/prompt) and the territory9(actual codebase, domain, constraints)**.10Origin: Blindspot Pass — see skills/loop/references/talk-source.md1112## Iron Rules13141. **Do not implement yet.** No code changes while this skill is active.152. End goal of the investigation: help the user **write a better prompt**.1617## Procedure18191. Parse task description and context sources from `$ARGUMENTS`. Empty → the task20 currently under discussion; if there is none, ask one question first. If sources are21 given (git history, docs, specific modules, etc.), investigate those first.222. Scope codebase-wide or large → spawn the `unknowns:unknowns-scout` agent with the23 Agent tool. Hand it a packet: the user's original prompt verbatim, the context24 sources, the target area and its entry points, and what to return (finding table +25 improved prompt draft); map its rows onto the card kinds below.26 Scope is a few files → investigate directly with Read/Grep/Glob.273. Compile findings, each tagged with a kind:28 - **Landmine** — touching this breaks something non-obvious (regression risk, fragile29 or missing tests, a module mid-migration)30 - **Convention** — an unwritten rule the codebase enforces31 - **Missing concept** — a mechanism the user's prompt has no word for32 - **History** — an earlier or reverted attempt at this exact task33 Also collect the **questions to answer** and the **information still needed** to34 sharpen the prompt.354. Lead with the contrast: **What you asked for** (the task as the user framed it, and why36 it sounds small) vs **What you're actually walking into**, with a tally by kind37 ("4 landmines, 2 conventions, 1 missing concept, 1 reverted attempt"). Then the cards —38 kind, finding, **why it bites**, recommended action — sorted by **importance × impact**.395. End with an **improved prompt draft** reflecting the findings — the core deliverable of40 this skill. It names the execution order and ends with an explicit checkpoint41 ("stop and show me the plan before writing code").426. Offer with one AskUserQuestion: proceed with this prompt now / edit it first / stop here.437. If undecided items that could change architecture surface, suggest continuing with44 `/unknowns:interview` (or `/interview` for copied installs). If the user lacks the45 unfamiliar domain's vocabulary itself, suggest `/unknowns:teach-me`.4647## Output4849Artifact tool → publish the page; else `.unknowns/<YYYY-MM-DD>-blindspot-<slug>.html`; else markdown.50Reaction control: a copyable **prompt fix** per card; selected fixes assemble into the improved prompt draft.51Details: skills/loop/references/output-routing.md5253## Scope5455Not limited to application code — specs, migrations, infra and vendor integrations all56have landmines and unwritten conventions worth mapping before you touch them.57The subject is always a **specific codebase, plan or system** and its blind spots.58If the gap is the user's **vocabulary** for an unfamiliar field, that is59`/unknowns:teach-me` (or `/teach-me` for copied installs), not this skill. Running both60is fine: teach-me first for the words, blindspot for the territory.
ajitta