Skip to content

Backlog philosophy (workspace + agent capability building)

Purpose: Single reference for what “backlog” means here, how we capture work from conversation, and how to structure a backlog if the goal is compounding agent + human capability over time.

Canonical prioritized list: BACKLOG.md (repo root) — rules: .cursor/rules/backlog-prioritization.mdc, .cursor/rules/backlog-notify.mdc. Epic → Story → Description: BACKLOG_STORY_DETAILS.md, BACKLOG_HIERARCHY.md. Product backlog URLs (13+): backlog-practices-reading-list.md.

Rich tickets (optional): Field definitions, epic keys (EP-NN), templates, and TKT-NNNNN keys live under backlog/ — see backlog/TICKETING_MODEL.md. The root table stays the slim ordered index; full markdown tickets are for history, assignee, and paste-friendly descriptions.

Related: PRIMARY_OBJECTIVES.md, Jira/Sheets tooling TBD, .cursor/skills/feed-primary-objective/SKILL.md, onboarding/.


1. Traditional definitions (external anchors)

General project / agile usage

Per awork — Backlog (glossary): a backlog is an evolving list of tasks, requirements, or features not yet completed, used to plan, prioritize, and track progress — with practices like regular updates, clear prioritization, consistent item descriptions, and visibility to the right people.

Nuance: In the broad sense, “backlog” is tool-agnostic (Jira, Sheets, markdown). It is not the same as a committed timeboxed plan — it is the pool from which commitments are pulled.

Scrum (precise vocabulary)

The Scrum Guide distinguishes:

Artifact Meaning
Product backlog Emergent, ordered list of what is needed to improve the product; single source of work for the team on that product (Scrum.org — Product Backlog).
Sprint backlog Subset committed for one Sprint, plus a plan to deliver Done work.

Nuance: “Product” can be a service, internal tool, or capability substrate — not only customer-facing software. For a solo operator, “Scrum team” collapses to you + agents, but the ordering + emergence ideas still reduce chaos.


2. Working definition (this workspace)

Backlog = an ordered, living inventory of candidate work — ideas and commitments-in-waiting — used to prioritize, plan, and track what gets done next, with enough structure that future-you (and collaborators) can reason without re-deriving everything from chat memory.

Not the same as: unbounded idea dumps with no ordering, or a roadmap narrative with dates (unless we explicitly link items to horizons).


3. Capture use cases (why one process is not enough)

Victoria’s intent: many items arise in discussion; scope, time, attention, and risk differ. Treat these as different intake lanes so triggers and fields can differ.

Lane Typical trigger What “good capture” includes Process note
A. Deferred work “We’ll do that later,” out of scope this sprint/session, post-MVP, parking lot Link to source (chat/session), why deferred, rough epic, reactivate condition (e.g. “after public report stable”) Automatic assist: when user or assistant explicitly defers, draft a backlog row; user confirms or edits.
B. Recommendations not taken User ignores, declines, or forgets an assistant suggestion; or says “not now” without a defer reason What was recommended, assumed benefit, cost/risk, why skipped (unknown ok) Automatic assist: assistant ends relevant turns with optionalBacklog candidate” one-liner; only writes to canonical backlog when user says capture / yes / add to backlog (avoids nagging).
C. Horizon / invention Large bets, new systems, architecture pivots, “huge idea” Outcome, non-goals, dependencies, spike vs build, minimum slice Often stays large until refinement (Scrum: product backlog refinement); may spawn child items in lanes A/B.

Design principle: Automation suggests, human commits for anything that changes external systems (Jira, Sheets) or sensitive scope — aligned with common human-in-the-loop patterns in agent improvement discourse (e.g. proposals reviewed before “deploying” behavior changes).


4. Research touchpoints (agent capability backlogs)

These are adjacent patterns, not prescriptions:

  • Feedback loops: Systems that improve often separate proposalreviewapply (evaluation, production signals, human edit as “gold”) — see e.g. AgentC2 — self-improving agent (general architecture ideas).
  • Hierarchical goals: Manager-style decomposition (outcometasksverify) reduces coordination load — e.g. Zylos — hierarchical agent coordination (research framing).
  • Deferral / resume: Checkpointing long workflows maps to “we paused; don’t lose the why” — e.g. Zylos — checkpointing (research framing).

Use these to inform backlog states and quality gates, not to over-engineer a personal repo.


5. If building the “perfect” backlog for agent self-capability (opinionated structure)

If the product is “the collaboration substrate” (skills, rules, knowledge, evals, tools), a useful structure is five layers — from stable intent down to executable work:

Layer 0 — Capability themes (stable buckets)

Map to epics (e.g. AI skills, rules, foundational AI knowledge, traditional technical foundations, AI prime objectives — plus optional Collaboration & Cursor workflow, Shipping & spin-outs). Themes change slowly.

Layer 1 — Outcomes / bets (6–18 month horizon)

Each item names observable capability: “We can reliably X without Y failure mode.” Not a pile of tasks yet.

Layer 2 — Programs or initiatives (weeks–quarters)

Grouped work with a definition of done for the program (e.g. “website-qa skill + one live run,” “backlog Jira sync v0”).

Layer 3 — Backlog items (ordered, emergent)

Concrete PBIs / issues: sized enough to prioritize. Scrum’s ordering + refinement applies (Scrum.org — refinement).

Layer 4 — Committed slice (timeboxed)

What “this week / this session” actually takes on — the sprint backlog analogue. Small WIP to protect focus.

Cross-cutting tracks (not separate backlogs unless volume forces it):

Track Holds
Intake / triage Raw captures from lanes A–C before they become Layer 3 items
Quality & eval “How we know capability improved” (checklists, regression prompts, /website-qa-style gates)
Governance / safety Items affecting rules, secrets, visibility — higher scrutiny before merge

Anti-patterns to avoid:

  • One undifferentiated list → everything feels equally urgent.
  • No commitment boundary → backlog anxiety (“infinite list”).
  • Auto-writing Jira/Sheets on every stray sentence → noise and trust erosion.

5b. Workspace practice — Backlog candidates in daily journals

Each journal/.../daily.md includes Backlog candidates (from this session) (see journal/daily-journal-template.md). That section is assistant-assisted at /session-end or substantive wrap-ups, scoped to the current chat unless the user supplies more context.

Weekly reviews roll those lines up under Carry-over from journals (journal/weekly-review-template.md).


6. Automation boundaries (draft policy)

Behavior Policy
User says “later / defer / parking lot Assistant proposes a structured capture (lane A); user confirms once.
Assistant recommends change user doesn’t adopt Assistant may offer one-line backlog candidate (lane B); no auto-write to canonical store without confirmation.
Big idea surfaces Assistant summarizes bet + suggests first thin slice (lane C → Layer 3).
Session end Optional: activeContext.md / journal “Tasks for Later” stays short; meatier items promoted to canonical backlog when tooling exists.

Refine when Jira/Sheets integration exists.


7. Changelog

  • 2026-03-23 — Linked daily journal Backlog candidates section + weekly Carry-over from journals; session scan skill and journaling rules.
  • 2026-03-22 — Initial doc: merged traditional definitions (awork, Scrum), workspace working definition, three intake lanes, research touchpoints, five-layer capability backlog opinion, automation draft policy.