Agent Coding Workflows

Quick start

• Detect invocation: user asks for a workflow or uses /flow, /workflow, /andrej, or /AndrejHide.
• If /AndrejHide is used, pick the workflow automatically.
• Classify the task: standard change, algo/perf, or urgent bugfix.
• Run the matching workflow steps in order.
• Pull in other skills as needed.

Workflow selection

• Standard change: most feature or refactor tasks.
• Algo/perf: algorithms, hotspots, or scaling work.
• Urgent bugfix: regressions or production-impacting issues.

Auto mode

Use this when the user says /AndrejHide.

1. If the task mentions performance, scaling, optimization, latency, or complexity -> choose Algo/perf.
2. If the task mentions regression, outage, production bug, or urgent fix -> choose Urgent bugfix.
3. Otherwise -> choose Standard change.
4. If still ambiguous -> run assumption-clarifier first, then decide.

Thresholds (tunable)

Default triggers for plan-lite and review. The user can override by stating new thresholds.

• Plan-lite required if any are true:

  • Estimated > 30 minutes, or
  • Touches > 3 files, or
  • Diff > 150 lines, or
  • Changes a production or critical path.

• Conceptual-reviewer required if any are true:

  • Any non-trivial task, or
  • Touches > 2 files, or
  • Ambiguous requirements, or
  • Risk of regression.

If the user provides overrides (e.g., "thresholds: files=5 lines=300"), use them for this task.

Preflight (from Andrej’s post)

Do this before any workflow if the task is non-trivial:

1. Convert the request into explicit success criteria or tests.
2. Surface assumptions and ambiguities early.
3. Keep a lightweight plan (3–5 steps) if risk is non-trivial.
4. Bias toward minimal diffs and simplicity.
5. Reserve a conceptual review pass for subtle errors.

Workflow A: Standard change

1. Use assumption-clarifier to surface missing info.
2. Use plan-lite if the task is non-trivial.
3. Use minimal-diff-implementer for the smallest safe change.
4. Use criteria-test-loop if tests exist or can be added.
5. Use conceptual-reviewer before handoff.

Workflow B: Algo or performance

1. Use assumption-clarifier to lock requirements.
2. Use plan-lite for a short path to success.
3. Use criteria-test-loop to build naive-correct first.
4. Optimize only after tests pass.
5. Use conceptual-reviewer to validate correctness and tradeoffs.

Workflow C: Urgent bugfix

1. Use assumption-clarifier for a rapid repro and scope.
2. Use minimal-diff-implementer to keep the fix tight.
3. Use criteria-test-loop with a focused test.
4. Use conceptual-reviewer for regression risk.

Output format

• Chosen workflow: name and reason.
• Steps executed: short bullets.
• Open questions: if any.
• Next action: what happens next.

Guardrails

• Do not invent a new workflow unless asked.
• Do not skip conceptual-reviewer on medium+ changes.
• If information is missing, run assumption-clarifier first.
• Prefer naive-correct first, then optimize.
• Avoid extra abstraction and unrelated cleanup.