“I Let ChatGPT Manage My Workweek” - @KatieParrott
Why this is in the vault
First-person, single-operator account of building a personal AI project manager on top of ChatGPT + Notion + Google Calendar + Slack + Drive - with the actual setup recipe, prompt skeleton, weekly cadence, and a candid failure-mode list. This is the personal-scale mirror of 2026-04-09-every-four-ai-agents (Every’s company-scale Anton/Max/Strategy/Reporter stack) and the personal-scale mirror of RDCO’s own COO-agent setup. The interesting content isn’t the “AI can manage your work” claim - it’s the specific dependencies that make it work or break (database freshness, OKR document quality, single-task prompt discipline) and the explicit admission of what AI cannot fix (the operator’s own avoidance behavior). Source-fidelity caveat: Every’s email body did not render via Gmail (known empty plaintextBody gotcha); reconstructed from canonical URL via WebFetch.
Sponsorship
Mixed. Third-party paid: Intent (Augment Code), a spec-driven development platform with agents - standard Every paid placement, properly disclosed. Self-cross-promo (Every-internal products mentioned in editorial): Proof (Every’s editor, used here as the OKR document store), Plus One / Margot (Every’s personal-agent product, presented as alternative to ChatGPT for the same workflow), and Every Consulting (linked CTA). Per the sender-pattern note in the README, treat Every-internal product mentions as cross-promo regardless of disclosure framing - the editorial recommendation surface is not neutral when the alternative being recommended is owned by the publication.
The core argument
Parrott - 15 years a writer, self-described “garbage at project management” - built a ChatGPT-based agent to translate quarterly OKRs into a weekly operating rhythm. The setup is mechanical and reproducible:
Inputs the agent reads as context.
- Quarterly OKRs (4 objectives, ~12 KRs) stored in a Proof doc
- Strategy / planning docs explaining the why behind the OKRs
- Workstream documentation (editorial calendars, recurring meetings)
- A stack-ranked goal priority - “which objective would you protect if you could only protect one?”
Tools the agent connects to.
- Notion board (source of truth for project status)
- Google Calendar (Google or Outlook works)
- Slack (catches ad-hoc tasks dropped in conversation)
- Google Drive (working drafts)
Behavior contract (the load-bearing prompt rules).
- “You are my project manager. Your job is to help me organize each week and keep my quarterly objectives on track.”
- OKR document is source of truth for priorities
- Notion is source of truth for project status
- Don’t invent statuses, deadlines, or tasks
- When dates don’t exist in Notion, acknowledge the gap (don’t guess)
- Ask one clarifying question for ambiguous tasks rather than guessing
- Flag non-OKR requests before helping with them
- Protect stated priorities from daily impulses
The cadence.
- Monday: one-page plan - what’s due, what’s at risk, priority focus areas tied to OKRs
- Friday: status report - what shipped, what slipped, which goals are drifting
- On-demand: “what should I work on now?” returns a single recommendation after checking calendar gaps + open Notion tasks
Concrete decomposition example. “Stand up a reliable Vibe Check pipeline” became: Audit existing process - Draft process brief - Solicit feedback - Implement changes. The agent’s actual work was breaking quarterly chunks into weekly-sized actions with deadlines.
What worked, what broke
Worked:
- Quarterly goals reliably translated to weekly tasks
- Early warning on slipping objectives (problems surfaced before they snowballed)
- Reduced cognitive load of mental project tracking
Broke or limited:
- Database staleness was fatal. “If it’s not up-to-date, the whole system falls apart.” Forgotten Notion updates degraded recommendations to noise within days.
- Communication overhead. Parrott had to manually update Notion and tell the agent what was complete. Two-write tax.
- The “me” problem - AI can’t fix avoidance. She sat on a proposal for a week despite the agent reminding her daily. The agent flagged it. She avoided it anyway. Self-doubt is not a project-management bug.
- Strategic decisions stayed human. Believing in your own ideas is not delegable to a project manager, AI or otherwise.
Punchy framing she uses: deadlines used to feel like “weather systems: suddenly overhead, occasionally catastrophic, mostly outside my control.” The agent moves them onto the calendar. It does not move them out of your head.
Mapping against Ray Data Co
Strong mapping. This is the personal-scale recipe for the exact pattern RDCO is operating at COO scale - and the failure modes she names are the same ones we’ve already mitigated structurally.
-
Architecture parity. Parrott’s stack (OKR doc + Notion board + Calendar + Slack/Drive, with prompt rules pinning sources of truth) is structurally identical to RDCO’s vault + Notion task board + Calendar + Discord/iMessage stack. The harness pattern is convergent: a few shared databases as the substrate, an agent that reads them and produces priority + status, the operator providing judgment. This is the second independent confirmation in 30 days that the pattern is industry-default - the first being 2026-04-09-every-four-ai-agents at company scale and 2026-05-03-every-context-window-codex-goes-to-work (Moretti’s
ce:strategy+ce:product-pulsetwo-skill loop). Three convergent data points = the harness/skills pattern is the playbook, not a thesis. -
The Monday plan / Friday status cadence is the gap RDCO does not cleanly cover. We have
/morning-prep(daily, calendar-aware) and/check-board(on-demand, task-oriented). We do not have a structured weekly Monday-plan + Friday-status loop. The closest is the autonomous loop’s idle-cycle behavior. Decision surface: worth considering a/weekly-plan(Monday) +/weekly-status(Friday) skill pair, scheduled via cron, that produces the one-page plan and the slipped-vs-shipped report Parrott describes. Low-risk to prototype; the inputs already exist (Notion task board has Critical Component checkbox perproject_critical_component_field.md, vault has SOUL.md / project READMEs as the OKR analog). Verdict: queue this as a Notion task for/check-boardto pick up - do not block on founder approval, it’s reversible additive skill work. -
Database-freshness failure mode is already mitigated for RDCO and we should keep that lead. Parrott’s biggest failure was Notion staleness because she had to dual-write (do the work, then update Notion). RDCO avoids this because most Ray-owned status writes go through the agent -
/check-boardupdates Notion as it works, vault notes are written by skills, not transcribed by hand. Implication: the personal-PM pattern Parrott describes will degrade for any operator who has to dual-write. RDCO’s “agent owns the writes” architecture is the durable advantage. Worth surfacing in any future Sanity Check piece on “why personal AI PMs break.” -
The “me problem” is the unhobbling ceiling. Parrott names what’s actually irreducible: avoidance, self-doubt, choosing which idea to believe in. The agent can flag, remind, decompose - it cannot will. This connects to the founder’s
feedback_calibrate_overconfidence.md(Ray defaults to overconfident from polished docs; founder’s lived reality is the ground truth). Same shape: the agent’s job is to reduce friction on execution, not to substitute for judgment. The harness is the floor, not the ceiling. -
Prompt rules are stealable. The behavior contract she wrote - “don’t invent statuses, ask one clarifying question, flag non-OKR requests, protect stated priorities from daily impulses” - is well-formed and worth lifting verbatim into a future RDCO prompt-pattern reference. Especially the “ask one clarifying question rather than guessing” rule, which RDCO already practices but doesn’t have written down as an explicit pattern.
-
L5 north-star check. Per
project_l5_north_star_strategic_direction.md, RDCO’s front-burner work is unhobbling the COO agent (toolset + visibility). A Monday-plan + Friday-status skill pair, the database-freshness lead, and the lifted prompt rules all reinforce the front-burner thesis. None of this pulls toward operating small bets - it deepens the harness.
Decision surface (founder review): None requiring immediate judgment. The /weekly-plan + /weekly-status skill pair is a reversible additive build - queue to Notion, let the autonomous loop pick it up.
Related
- 2026-04-09-every-four-ai-agents - company-scale version of the same pattern (Anton + Max + Strategy Interviewer + Campaign Reporter on shared Notion graph)
- 2026-03-31-every-onboarding-ai-project-manager - Every Consulting’s “Claudie” PM agent; the consultancy-scale version, with the “fire and rehire” iteration story
- 2026-05-03-every-context-window-codex-goes-to-work - Moretti’s
ce:strategy+ce:product-pulsetwo-skill loop; same harness architecture at solo-PM scale, agent-native PM vocabulary - 2026-04-16-every-youre-the-manager-now - “you’re the manager now” thesis; parallel-agent supervision UX; Austin Tedesco’s 1-100 confidence check is a complementary stealable pattern
- 2026-04-15-every-claude-managed-agents-mini-vibe-check - Claude’s managed-agent surface; relevant when picking the agent runtime
- 2026-04-08-four-levels-of-ai-use - Parrott is operating at L4 personal use (custom workflow agent, not generic assistant); RDCO is L4-pushing-L5
- 2026-04-14-levie-agent-deployer-role-jd - the role-definition framing; Parrott is informally an agent-deployer for herself
- 2026-04-04-compound-engineering - Plan / Work / Review / Compound loop; Parrott’s Monday-plan + Friday-status is the Plan + Review halves
- 2026-04-01-karpathy-llm-knowledge-bases - knowledge-base-as-context thesis; Parrott’s OKR doc + Notion board is the personal-scale knowledge base
- 2026-03-23-every-four-hour-workweek-job - same author Katie Parrott on the four-hour-workweek tradeoff; voice/series continuity
Tracked-author candidates
- Katie Parrott (Every Working Overtime staff writer) - already a recurring Every byline in the vault (
2026-03-23-every-four-hour-workweek-job). Consistently writes the personal-scale operator-facing column with concrete recipes. Worth a CRM stub if not already present - this is the second high-utility piece from her in 6 weeks.
Copy-paste caution
Body reconstructed from canonical URL via WebFetch (Every’s email body did not render). All quoted phrases are ≤15 words and attributed. Source-fidelity caveat surfaced in frontmatter and in “Why this is in the vault.” Original at every.to.