# Changelog

## 0.2.0, revision 11 (November 3, 2026)

- AI Coding Memory 1.5.0 replaces 1.4.0 in `kit/ai-coding-memory.kit.json`. The 16 routine playbook templates no longer describe the author's own systems: each names example sources for several stacks (AWS, Google Cloud, Azure, Datadog, GitHub, Stripe and similar) and a step to fill in your own, and asks you to estimate the cost per run instead of quoting ours. A new guideline, Routine Playbook Customization, tells an assistant how to copy a template, adapt it with the owner, keep it current after kit upgrades and add new playbooks. Three templates are renamed: Cloud cost watch, Customer data quality, Weekly leadership digest. 24 guidelines; entity types, collections and relations are unchanged. An organization that installed 1.4.0 from this file can install 1.5.0 over it: templates it has not edited are updated, edited ones are kept. The 1.4.0 file is withdrawn.
- The kit map is regenerated for 1.5.0.

## 0.2.0, revision 10 (November 2, 2026)

- AI Coding Memory 1.4.0 is now installable from this playbook: `kit/ai-coding-memory.kit.json` is the full kit as one Personize manifest, validated against the platform's manifest schema, with internal examples replaced by placeholders. The AI starting guide installs it through the Personize API when the kit is not in an organization's catalog, instead of stopping.
- New reference, "Team Memory: governed operational memory": memory organized around entities, events, decisions and ongoing work, distinguished from session and personal memory, with two conceptual visuals and related research. Linked from the README, the AI guides, the kit map and the playbook page. Counts and performance claims are unchanged.
- October 2026 is now the complete benchmark month. The monthly activity data runs through October 31: 2,269 authored PRs, 2,155 merges, 6,597 commits. Against DX's Q1 2026 median for daily Cursor users, October reads 118.7× as counted and 16.0× at September's PR size (16.1× by code lines); the median merged PR fell from 788 lines to 139. Both are PR-volume comparisons, not productivity and not the 20× measure. Per calendar day, non-merge commits and code lines ran about 3.0× the April–September average.
- Added "October 2026, the benchmark month" to the results reference, and `evidence/october-2026.json`: finished and filed tasks, PRs, lead time, time to merge, first-attempt acceptance, API-equivalent compute by kind and two separate cost-per-task measures, each with its definition. Decision wait is kept as a snapshot at collection, not an October figure.
- Updated the introduction, short version, AI measurement guide, evidence appendix and metrics page for the complete month. The October 1–19 task audit and earlier dated cohorts keep their numbers and windows.
- Revision 8's partial-month readings (13× at September's PR size, October 1–21 activity ratios) are superseded by the complete month.
- Section 5, the memory guide and the workflows reference: open questions and answers are now separate records (inquiry and decision), with reuse levels, a pickup check, precedent search and a tiered import of older answers. The claims note now covers the plan runner's opt-in read-back claim; no claim uses the versioned check yet, so one coordinator still runs at a time.
- Section 6 and a new workflows reference section, "What changed after this snapshot": rules that came from incidents (one cap for everything, reviews as capped headless runs, task sizing, pasted contracts, thin reviews, red against the base, route instead of work, drive instead of patch), parking work that needs a person, security audits as capped dry-run-first jobs whose findings are not published, and UI testing against a test account.
- New reference, "Learning from other frameworks and platforms": what I adopted from others and how far along each adoption is, a chooser for when to use another method or tool instead of or alongside this playbook, and a first second-reviewer calibration (27 merged PRs that passed my deep review, November 2). Checked November 2, 2026.
- Section 7 and the results reference: what 1,109 of my own answers to agent questions showed (September 15 to October 30), with the changed-answer shares marked as estimates from a hand-labeled half sample. The modules carry short summaries so the core stays under its word limit.
- Workflows guide: driving a queue without doing the work, parking work that needs a person, audits and UI checks.
- Agentic Kanban: the decision classes and the dispatcher are added as `planned`, with the single hand-run round of October 31 and its misfits. The private dashboard behind the October figures is described by its definitions only.
- A stable link, `https://hamedtaheri.com/playbook/latest/`, now redirects to the current release. The "Give to your AI" buttons, the copy-paste prompts, `llms.txt` and each AI guide's header use it, so a shared link never goes stale. The AI starting guide adds a section for readers who already use Personize, and the FAQ adds the October benchmark.
- Kit map regenerated for AI Coding Memory 1.4.0 (routine records: Customer Account, Metric Snapshot, routine fields on Task). The routines are not running yet.
- Charts brought to the complete October. The year chart's caption and table read the complete activity data instead of the October 20 snapshot. The Claude Code activity chart adds the complete October (9,484 main sessions, a 737-hour session proxy, index 3.93 billion) from a rescan that reproduces the October 14 snapshot within 0.2%. The run-outcomes card adds a second window, October 26–31: 228 of 415 runs approved and waiting on me, against 120 of 206 on October 6–14 ([aggregates](https://hamedtaheri.com/playbook/v0.2.0/evidence/pipeline-outcomes-late-october.json)). The run ledger for October 15–26 was not kept.

## 0.2.0, revision 9 (October 23, 2026)

- Added an interactive map of the AI Coding Memory kit (`reference/kit-map.html`), linked from the kit architecture page. It is generated from the kit definition (now 1.3.0) and shows every entity type, the relation graph and matrix, statuses, properties, guidelines and document tags. Click an entity type on the map to show only its relations. Emails and repository names in the kit's examples are replaced with placeholders.
- Added "The kit map" to the maintainer guide: which script regenerates it and the checks before publishing.

## 0.2.0, revision 8 (October 21, 2026)

- Added "How much more work happened in October" to the results reference. Merged PRs per calendar day rose 9.99× from September, but the pipeline's one-PR-per-task packaging, task-branch repair PRs and integration merges drive most of that. On non-merge commits and code and test lines changed, October 1–21 activity is 1.53–2.03× September's per calendar day. It is reported as an activity ratio, not a productivity multiplier.
- Added the monthly activity series (October 2025 to the current partial month): commits, merge share, lines changed by category, new design documents, merged PRs by destination and per-day rates. Repository names and PR links are excluded.
- Added `scripts/playbook-activity.py` with tests so the series can be refreshed and rebuilt. Added AI measurement guidance for comparing periods across a workflow change.
- Added "A year against the DX benchmark" to the results reference, with a generated chart: merged PRs a week as a multiple of DX's Q1 2026 median for daily Cursor users (4.1). April–September ran 4.8–8.8×. October 1–21 reads 13× at September's PR size (18× by code lines; 87× as counted). Documented what DX defines (per contributing engineer, bots excluded) and what it does not publish (PR size, a time unit for the Cursor figure). Lines per merged PR are reported as an upper bound.
- The annual progress chart and the benchmark chart are now generated from the monthly activity data, so a refresh updates both. The October 20 recount stays as a frozen file.
- Added a maintainer guide in the source repository (`MAINTAINING.md`, outside the bundle): which script refreshes which data, which prose numbers are updated by hand, the checks to run and the monthly refresh routine.

## 0.2.0, revision 7 (October 20, 2026)

- Added the October 2025–October 2026 progress chart: merged-PR bars and a commit line on separate zero-based axes, with April execution and October shared-memory milestones.
- Used the common-window October 20 recount for both series, with partial October explicitly labeled and normalized calendar-day rates available in the chart data. Preserved earlier dated audits.
- Updated the homepage and About page with the author's product-building perspective, technical work, personal products and artwork. Made the homepage artwork a full-width gallery link.
- Corrected the subscription version label and synchronized release sources, generated reading views and downloads for publication.

## 0.2.0, revision 6 (October 20, 2026)

- Renamed the canonical page to `/20x-developer-playbook/`, including its full version, framework and reference pages. Old page names redirect permanently with deep paths preserved. Updated navigation, reference links and social-image tooling. Stable `/playbook/v0.2.0/` downloads remain available.
- Reorganized the results around what work got through, what verification stands behind it, and what it cost. Kept review rounds, repair rates and timings in the detailed reference.
- Made counting levels explicit: requested outcomes, implementation units and PRs are distinct. Of 275 merged units, 213 are classified as reviewed with passing tests; 62 have other evidence categories, including 14 flagged units.
- Added a checkable path from requirement through verification, review and integration to acceptance. Highlighted the existing team-memory section and connected its task/evidence links to inspectable completion claims.
- Repaired text encoding in review/cost sources and the standalone builder.

## 0.2.0, revision 5 (October 20, 2026)

- Refreshed complete October 1–19 PR counts: 1,026 authored and 972 merged; independently checked destinations. The October 12–18 week has 770 merges, including 462 internal task-branch merges.
- Added a metric availability audit, aggregate JSON and web report: 275 merged units, 185 fully merged recorded parents, repair incidence, review coverage, elapsed-time proxies and API-equivalent usage. Kept acceptance, human effort, autonomy and memory-only causal uplift explicitly unknown.
- Explained the 20× title as a repeatable accepted-delivery capacity target with a defined baseline, distinct from the approximately 188× raw weekly PR-volume comparison.
- Updated the short/full playbooks, AI measurement guidance, current chart, navigation and downloads. Preserved dated earlier case studies and the existing Agentic Kanban onboarding draft.

## 0.2.0, revision 4 (October 19, 2026)

- Adopted The 20× Developer Playbook title and agreed subtitle, centered on systems built, shared-memory collaboration and experience. 20× remains a goal.
- Preserved revision-3 task costs, review checks and unattended-run findings through October 19, plus the current homepage and artwork.
- Reframed the existing Agentic Kanban pages as a working draft for developer onboarding and community trials; retained their contracts, templates and running/piloting/planned practices.
- Synchronized site labels, metadata, AI guidance and downloads.


## 0.2.0, revision 3 (October 19, 2026)

- New task data, October 1 to 19: pipeline cost per merged task (median $1.35, mean $2.32, 90th percentile $4.87, highest $19.65 across 272 tasks), what review stood behind each merged revision, and how split features finish. Evidence in [task checks](https://hamedtaheri.com/playbook/v0.2.0/evidence/task-checks.json); coordinator excluded.
- Review: "What the records show". The first deep review stopped 1 in 5 PRs while the first ordinary review approved all but one; 146 of 272 merged tasks had a deep review and passing tests on the exact merged revision; 13 (4.8%) were not fully checked. Two gaps named: fix PRs are reviewed on their own diff, and commits that meet review conditions are not re-reviewed. The AI prompt asks for confirmation on the combined revision.
- Review: "never to round three" now reads "not to another fix round", consistent with the round-6 rules.
- Cost: the $5 hard cap is per run; nothing enforces a per-task total yet. New table "October 1 to 19, per task", including about $12.20 per PR landed in the integration branch for the October 16–19 unattended run, coordinator included. Cost per accepted task is still not established.
- The unattended run's totals are final: 81 PRs landed, about $990, about a third of it the coordinator.
- Preparation: of 41 features split into parts, 7 had every part merged.
- PR counts keep their October 15 cutoff; other numbers outside these sections keep October 14.

## 0.2.0, revision 2 (October 19, 2026)

- "Running the board unattended" now covers the whole run, October 17 to 19, about three days: one night alone, then two days with a person dropping in. New rules: a question never blocks the run (ask in plain text with a recommendation and keep working); judgement calls get a peer reviewer from another model family; landing is a serial queue that reruns tests and build on the merged result; a stacked part lands only after its parent; a part that ships a temporary stand-in waits for its successor; know which gate actually runs; plan the person's time, not only the agents'. Three new failure patterns. Still piloting.
- New in the playbook: "When the rounds don't stop", from 108 reviewed pull requests. Review rounds rarely converge on their own; past round 6 only blocking findings get fixed, no hardening the acceptance did not ask for, and size is a reason to stop. The AI prompt says the same.
- Contract: the `unattended` block gains the round-6 rules, a landing queue, the question rule and the peer rule.
- Measurements outside the unattended section unchanged.

## 0.2.0, revision 1 (October 18, 2026)

- New section in the full framework, "Running the board unattended", from one night (October 17 to 18) with a coordinator running the backlog: a never list, rules approved in advance and checked again on the exact revision, a bounded review loop, dependency-aware scheduling with a parallel cap and a budget stop, a two-tier gate for breaking epics into parts, keeping context across compactions, and five failure patterns to avoid. Marked piloting; one night proves little.
- Landing in an integration branch now moves a task to human review, never to done. Spend against a budget counts every ledger that bills, the coordinator included.
- The morning brief folds in the overnight pack: each act taken under a rule, to confirm or undo, and what waits on a person.
- Contract: an `unattended` block (off by default) and the overnight pack in the morning brief. The four-minute version, the AI guide and the FAQ say the same in short.
- Measurements unchanged.

## 0.2.0 (October 16, 2026)

- New: **Agentic Kanban**, the team layer for agent-built software, on top of per-task tools such as Superpowers and Spec Kit: what replaces sprints, standups and velocity; three lanes sized to the change; four ceremonies; four numbers; three questions to start. Four-minute and full versions, a worked example, an AI guide, a machine-readable contract, ceremony templates and a portable skill.
- Starter kit: acceptance with evidence and revision, routine occurrences, attempt records, and `tools/flow.cjs` for the morning brief and weekly report without a model.
- Version 0.1.0 stays online at /playbook/v0.1.0/ for existing links.

## 0.1.0, revision 8 (October 15, 2026)

- PR counts refreshed through October 15, 11:30 p.m. Vancouver, with the same search rules: 455 authored and 431 merged in October, against 154 merged in September. The refresh reproduced the October 14 counts exactly. Commits, sessions and costs keep their October 14 cutoffs.
- Both versions now say what changed and when: April to September brought parallel sessions, sessions that message each other, longer self-checking runs and guardrails; October added shared memory for the agents and scripted pipelines to run them. The chart labels the phases as timing, not separate effects.
- The four-minute version opens on the monthly PR chart. The run-outcomes chart moves under "How the work flows" with a plainer headline.
- Prose no longer quotes running totals that go stale within a day; exact counts stay in the dated charts and evidence tables.
- Added `refresh_prs.py` to repeat the PR refresh; the evidence charts take their October labels and scale from the data.

## 0.1.0, revision 7 (October 15, 2026)

- Reordered the four-minute version around where the work waits: 120 of 206 pipeline runs (October 6–14) ended with an approval only a person could clear. Results now lead with tasks and cost per task rather than commits; caveats sit in one place.
- Moved Personize from the opening to practice 5. Added who reviews the PRs, what this does and doesn't show, and what the fleet vision depends on.
- Full version: role routing and the symptom table come first; the 12-question assessment moves after the practices. Practices 1 and 2 open with what went wrong. Added a copy-ready decision-rounds prompt, an illustrative brief, and "Where developers fit".
- Web visuals: run outcomes, the pipeline with stage costs, four incident cards and the lean-worker comparison. The headline chart shows only my own trend; published per-developer benchmarks stay in the evidence appendix.
- Assessment: output-ambition choices stop at "more than +200%".
- AI guide: removed instructions about when to mention the author; kept the contact. FAQ answers who reviews.
- The starter bundle no longer includes image-generation prompts.
- Measurements unchanged; October 14 cutoffs.

## 0.1.0, revision 6 (October 15, 2026)

- Replaced the seven-checkbox self-check with a six-area, 12-question workflow assessment: delivery, usable capacity, economics, coordination, planning flow and quality.
- Results explain each answer, evidence gaps, a selected bottleneck and practical next tests, with copy/download for the team or its AI. No overall maturity score or promised output multiplier.
- Kept expected additional output separate from observed rates, with an optional comparable-work count ratio and explicit zero/unknown-baseline handling.
- Added the assessment method, interpretation rules and SPACE/DORA context. AI guidance and downloadable material use the same questions.

## 0.1.0, revision 5 (October 15, 2026)

- Added Hamed's journey from early LLM assistance to agent-native development, his product-manager perspective and why Personize became part of the working method.
- Described today's reported use of tens of agents separately from the bold vision of hundreds or thousands working in parallel alongside developers.
- Kept the four-minute version concise while retaining all seven practices, scoped measurements and the prompt for a reader's AI.
- Aligned the full version's preparation, execution, review, cost, memory and learning sections with that motivation.
- Explained the entity-driven memory/governance use case and the current limits that stand between today's runner and the larger fleet vision.
- Updated AI guidance, FAQ and metadata. Existing metrics retain their October 14 cutoffs.

## 0.1.0, revision 4 (October 15, 2026)

- Replaced the origin-story short version with a practical summary of all seven full-playbook sections, scoped results and one prompt to pass to an AI.
- Kept October 14 measurement cutoffs; editing the prose does not refresh the data.
- Clarified the pipeline median: 205 dollar-valued records among 206 runs, including stops; coordinator and unpriced Codex usage excluded. Not every run completed all stages.
- Removed an unsupported link between self-check scores and dollar costs, and labeled the deeper-review budget as an estimate.
- Aligned the AI starting guide, FAQ, navigation and discovery text; corrected the self-check link.
- Aligned the task brief template with the six-field contract and corrected pilot paths, output setup and whole-workflow budgeting.
- Qualified reviewer credential controls, review-subagent hook coverage and separately authorized landing/fix-branch merges.

## 0.1.0, revision 3 (October 14, 2026)

- The 4-minute version is now a story: where it started, six months of pushing, the wall, the turning point (a memory the agents share), what happened next, how the job changed, and why now.
- Headline numbers now follow the journey: commits, sessions and merged PRs, with cost as part of the story.
- Unpublished documents are mentioned only in the full playbook and the AI guides.
- "Book a call" replaces "Compare notes".

## 0.1.0, revision 2 (October 14, 2026)

- New 4-minute version ([summary.md](https://hamedtaheri.com/playbook/v0.2.0/summary.md)) with the seven rules, a real run, a two-minute self-check, a day in the loop and a one-week test.
- Headline now leads with cost per task: about $16 on October 5, a median of $0.67 per pipeline run on October 6–14, and $3.20 per merged PR with the coordinator included.
- New pipeline evidence: 206 runs, outcomes and stage costs ([pipeline-outcomes.json](https://hamedtaheri.com/playbook/v0.2.0/evidence/pipeline-outcomes.json)).
- Section 1 adds the task model, how Superpowers planning becomes tasks, decision rounds, and the shared-table lesson.
- Section 2 adds the coordinator rule, the lean worker settings, a day in the loop, recurring work and operational failures.
- Section 3 adds the nine-round PR and the one-deep-review flow.
- Section 4 adds the nine-day cost timeline, the levers and budget rules.
- Section 5 corrects the concurrency note (the backend supports `expectedVersion`; the queue script doesn't use it yet), and adds people on the board and build-or-use.
- New section 6, guardrails, security and production. Learning moves to section 7. New reference section with a glossary, limits and what's included.
- New AI files: [FAQ](https://hamedtaheri.com/playbook/v0.2.0/agent/faq.md), [pilot guide](https://hamedtaheri.com/playbook/v0.2.0/agent/pilot.md), [build or use](https://hamedtaheri.com/playbook/v0.2.0/agent/build-or-use.md). [start.md](https://hamedtaheri.com/playbook/v0.2.0/agent/start.md) now routes by role.
- Benchmark comparisons, the subscription usage target and the October projection moved from the main page to the [evidence](https://hamedtaheri.com/playbook/v0.2.0/evidence/README.md).
- The runner scripts are available on request.
- The starter bundle no longer includes PNG images; served Markdown uses absolute links.

## 0.1.0 (October 14, 2026)

First working draft: six sections, AI modules, templates, memory starter kit and evidence.