dig

Dig

Flexible skill for exploring, planning, and speccing work on the Apify MCP server. Do NOT edit source files — this skill is for understanding and planning only.

Step 0: Parse arguments and determine intent

$ARGUMENTS contains the user's request and optional repo path overrides.

Flags (optional):

Everything not matching a flag is the user's request.

Resolution order for source repos: flag path → default sibling path → node_modules/ (compiled types only) → GitHub URL (last resort). Always verify the path exists before using it.

Determine the intent

Infer the intent from the user's natural language. There are three modes:

Rules:

• Default to Explore. When in doubt, do less — the user can always ask for more.
• Only enter plan mode for Plan and Spec.
• Only create GitHub issues for Spec.

Step 1: Explore

Read the relevant source files and explain your findings. This is the baseline for all intents.

What to do:

1. Read the relevant source files in this repo
2. Check similar existing features as reference
3. Only check the internal repo, MCP SDK/spec, or MCP Apps SDK/spec if the user's question touches those areas
4. If you spot a related open issue, mention it casually — but don't go searching for issues unless it's relevant

Stop here if the intent is Explore.

Step 2: Plan (Plan and Spec only)

Use the EnterPlanMode tool, then design the approach — and commit to one. If the request is genuinely ambiguous (e.g. it maps to two different existing knobs), surface the fork up front and ask; otherwise pick the most defensible design and commit rather than listing options.

Investigate first:

1. Assess internal repo impact (check ../apify-mcp-server-internal if available)
2. Check MCP spec/SDK if the feature involves protocol behavior
3. Check MCP Apps spec/SDK if it involves widgets or interactive UIs
4. Use mcpc @stdio tools-call to probe current behavior if useful (requires pnpm run build)
5. Read the repo's convention docs — follow them, don't reinvent

Conventions live in the repo, not here. AGENTS.md, CONTRIBUTING.md, and DEVELOPMENT.md hold the naming, validation, test-layout, and public/internal-separation rules — read them. Design minimally: reuse and adjust before adding.

Design output — required sections. A plan has to be implementable by someone else, so produce:

1. Approach — the chosen design in a few sentences and the main trade-off you accepted.
2. Files to change — each file to create or modify with a one-line note on what changes (path + change; no line numbers needed).
3. Interfaces — the key function/type signatures you add or change (names, params, returns), so the implementer and reviewers know the contract.
4. Test strategy (the oracle) — which unit/integration tests prove it, where they live, and what each asserts. This is how "done" is judged — never leave it implicit.
5. Risks & impact — edge cases, and apify-mcp-server-internal impact.
6. Data flow — only when non-trivial: entry → transforms → output.

Sections 2–4 must be concrete (real files, signatures, and tests — no "TBD", no "add error handling"). Sections 1, 5, 6 stay brief. This structure is for Plan; Explore and Spec stay lean.

Stop here if the intent is Plan. Exit plan mode with ExitPlanMode.

Step 3: Spec (Spec only)

Create GitHub issues. First exit plan mode with ExitPlanMode.

Check existing issues

Search for duplicates and related issues:

gh issue list -R apify/apify-mcp-server --search "<keywords>" --json number,title,state
gh issue list -R apify/ai-team --search "<keywords>" --json number,title,state
gh issue list -R apify/apify-mcp-server-internal --search "<keywords>" --json number,title,state

If a matching issue exists, update it with gh issue edit instead of creating a new one.

Create issues

One issue per implementation phase. A phase = one PR-sized unit of work (~50-200 lines changed). Each issue should be independently implementable.

Use the repo's feature_spec.yml template. Only the Problem and Proposed solution fields are required. Include Plan and Alternatives considered only when they add real value. No fluff, no filler — straight to the point.

## Problem
[Concrete evidence: error messages, user reports, issue links. Not "users are confused" — instead "3 users reported X in #channel".]
 
## Proposed solution
[Short. Reference existing code paths. List files inline if needed.]
 
## Plan
- [ ] Step 1
- [ ] Step 2

Style (Explore explanations and Spec issues):

• Plain language, no fluff — see CLAUDE.md § Communication style
• Skip any section that would be empty or generic
• 10-30 lines, not 100
• Concrete steps > prose

(A Plan's design follows its required-sections structure above — keep each section tight, but don't truncate files / interfaces / test strategy to hit a line count.)

Self-review before presenting:

• Is this the minimal design? Could scope be smaller?
• Am I reusing existing patterns or reinventing?
• Could this adjust existing code rather than add new code?
• Does it require refactoring first? If so, that's a separate issue.

Present issue content to the user for review before creating. Use gh issue create with t-ai label.

Available resources