SOP: Implementation-notes pattern for sub-agent dispatches
When to apply
Any sub-agent dispatch where the agent is implementing a SPEC (not just researching). Specifically:
- REQUIRED for: build scripts (
alpaca-*-deploy.py, settings.json modifications, schema migrations), multi-step refactors, financial-execution scripts, anything with ≥5 design decisions the sub-agent will make during the build. - OPTIONAL for: well-bounded code edits with 1-2 design decisions, single-file additions following clear templates.
- SKIP for: research briefs, single-essay synthesis, content writing — the artifact itself IS the implementation-notes equivalent.
The pattern
Add to the sub-agent dispatch prompt:
Maintain a running implementation-notes file at:
~/.claude/state/implementation-notes-<task-slug>.md
(or in-vault at: ~/rdco-vault/<project>/implementation-notes/<task-slug>.md
if the artifact warrants permanent record)
Capture as you work:
- **Design decisions**: choices you made where the spec was ambiguous
- **Deviations**: places where you intentionally departed from the spec, and why
- **Tradeoffs**: alternatives you considered and why you picked what you did
- **Open questions**: anything you'd want me (parent agent / founder) to confirm or revise
This file is a side-channel for in-flight reasoning. The summary-line return is the post-hoc verdict; this file is the durable record of how you got there.
Why this matters for RDCO
Our sub-agents already return a single advisor-mode summary line ending with DECISION: surface. That's the post-hoc verdict. The implementation-notes file captures the IN-FLIGHT reasoning the verdict came from — which:
- Lets the parent agent audit specific calls rather than re-deriving them. E.g., if a deploy script makes 5 ambiguous tax-code choices, the notes show WHICH choices and WHY. Parent can spot one wrong call without re-running.
- Lets founder spot misinterpretations EARLIER in a long-running build. The skill-improvement loop (
/self-review→/improve autonomous) gets better signal because the notes file shows where the sub-agent had to invent (not just what it shipped). - Reduces re-dispatch cost when a downstream cycle needs to extend or correct the build. The next session reads the notes file instead of reverse-engineering.
Where it lives
- Transient (most common):
~/.claude/state/implementation-notes-<task-slug>.md— survives across sessions but isn't permanent vault content. Cleaned up when stale. - Permanent (warranted by the artifact): in-vault next to the spec, e.g.,
~/rdco-vault/01-projects/<project>/implementation-notes/<task-slug>.md. Use when the build is a load-bearing artifact (a paper-trade deploy, a strategy spec, a schema migration).
Example dispatch language to fold in
When writing sub-agent prompts that meet the REQUIRED criteria above, append this block verbatim:
IMPLEMENTATION-NOTES REQUIREMENT (per 2026-05-18 SOP):
Maintain a running implementation-notes file at <path>. As you work, capture:
- Design decisions (spec ambiguities resolved)
- Deviations (intentional departures from spec + reasoning)
- Tradeoffs (alternatives considered + why you picked what you did)
- Open questions (parent / founder needs to confirm)
This file is your in-flight reasoning record. The final summary-line return is the verdict; this file is how you got there.
Audit hook (future)
A future check could scan ~/.claude/state/implementation-notes-*.md files on the morning brief and surface any with ## Open questions sections that have no resolution yet. That would close the loop from "sub-agent flags a question" to "founder sees it surface organically."
For now, the parent agent handles surface-to-founder manually when the notes contain meaningful open questions.
Source
- Original Thariq tweet (paraphrased): "Implement
. As you work maintain a running implementation-notes.html file that captures anything I should know about how the implementation diverges from or interprets the spec." Founder shared via iMessage 2026-05-18 17:09 ET. - Founder approval to adopt: iMessage 2026-05-18 ~4:04pm ET — "for the Thariq pattern, let's get the improve cycle going now."
Related
- [[~/.claude/projects/-Users-ray/memory/feedback_implementation_notes_sub_agent_pattern.md]] — feedback memory (loaded into every session)
- [[~/.claude/projects/-Users-ray/memory/feedback_fresh_eyes_subagent_for_own_artifacts.md]] — adjacent pattern (review own artifacts via fresh-eyes subagent)
- [[~/.claude/skills/improve/SKILL.md]] — the skill this pattern was adopted into via /improve cycle