ELDORADO / WHY CHANGE

Less Prompting,
More Shipping

The destination—and the first six months toward it

Improve accepted engineering work.
Not the number of times we ask AI.

Proposed strategy · 35 minutes + 15 minutes Q&A

ELDORADO / WHY CHANGE

We are not starting from “no AI”.

≈80 → ≈100engineers, by year end

>10teams

Already using AIuneven depth and quality

Solo architect → CTO
No formal authority over teams

Stack, tools, process, baselines
Capacity and actual bottlenecks

ELDORADO / WHY CHANGE

A faster draft is not a shipped outcome.

Intent→Change + tests→Independent acceptance

Existing release authority stays with people.

Less rework. Sound quality.
Understand the total burden.

If the bottleneck is CI or review capacity, fix that—not the prompt count.

ELDORADO / TARGET

Target / Autonomous SDLC.

Solid → lifecycle · Dotted ⇢ executes and governs

Humans direct, guide and review. Not deployed or promised in six months.

ELDORADO / TARGET

People contribute throughout the cycle.

Eight separate AI Agents, each human-controlled.

AI Agent⇢Scope

AI Agent⇢Architecture

AI Agent⇢UI/UX

AI Agent⇢Planning

AI Agent⇢Delivery

AI Agent⇢UAT

AI Agent⇢Security

AI Agent⇢Releases

AI Agent = Claude, Cline, ChatGPT or a similar tool, used under human control.

Dashed ⇢ contribution from outside the cycle.

ELDORADO / ROADMAP

Roadmap / Build dependable work first.

STAGE 1 · HUMAN + AIDelivery firstStandardize connected skills
and working practices

→

STAGE 1 · EXPANDOther SDLC phasesOrder undecided
Delivery keeps improving

⇢

STAGE 2 · SEPARATE AUTHORITYAutonomous capabilitiesProven practices + controls
toward the Target

Humans direct, guide and review throughout.

A strategic progression—not a six-month schedule or automatic graduation.

ELDORADO / ROADMAP

Connected skills. A repeating learning loop.

Understand work→Measure + baseline↓Engineer topology→Use in real work↓Collect results→Reassess

↻ Repeat within each phase. Not six global rollout stages.

Triggers → skills → handoffs → acceptance.

Relationships and cooperation—not merely a catalogue.

2 minutes. Within every adopted phase repeat the loop: understand the current situation, select relevant metrics and baseline, engineer connected skills, use them in real work, collect outcomes/friction/feedback, then reassess and repeat. Topology defines what exists, how skills relate, when each triggers and how they cooperate. In Delivery trace codebase principles through system logic, tests, builds/packaging and deployment before designing that topology. A code-change skill may hand an exact patch and regression evidence to independent review; its trigger is an eligible change, not an ambient right to act. Governance and Measurement support the whole loop. Qualify revisions before use, and reconsider simpler interventions whenever the evidence says they fit better.

Sources: roadmap; stage-01-skills-adoption.

ELDORADO / 6 MONTHS TARGET

6 Months Target / One credible contribution.

LONG-TERM DESTINATIONHuman-controlled autonomous SDLC

FIRST SIX MONTHSOne workflow understood and improved—or stopped with evidenceOne willing partner · reusable practice · owned upkeep

Earn a dependable building block.
Do not promise the entire destination.

ELDORADO / 6 MONTHS ROADMAP

Influence is earned beside the work.

CTOSponsor access
Resolve priority trade-offs

ArchitectPair, build, evaluate
Bring evidence—not mandates

Willing teamAgree contribution
Own acceptance and upkeep

An invitation with an exit.
Not a rollout disguised as a favor.

ELDORADO / 6 MONTHS ROADMAP

Months 1–2 / Earn the right to test.

01 · Discover

Trace real changes.
Rank friction and evidence gaps.

Output: current-work map

02 · Agree

One willing team.
Baseline + smallest intervention.

Output: bounded experiment packet

Decision 1 → Fund one experiment—or hold.

No partner, reliable checks or lawful evidence → no live pilot.

ELDORADO / 6 MONTHS ROADMAP

Months 3–4 / Test, then decide.

03 · Work alongside

Build and independently qualify.
Pair on bounded real work.

Output: failures + reproducible examples

04 · Compare

Reconcile outcomes and burden.
Include pending and abandoned work.

Output: retain / change / hold / stop

Decision 2 → Keep only what the evidence supports.

Missing follow-up is “unknown”, not “all clear”.

ELDORADO / 6 MONTHS ROADMAP

Months 5–6 / Transfer only if earned.

05 · Selective transfer

Champion + backup.
Check a second team’s fit.

Only with capacity and funded upkeep

06 · Evidence review

Team operates without daily help.
CTO chooses the next investment.

Owned workflow—or documented stop

New scope → new baseline.
No automatic company-wide rollout.

ELDORADO / 6 MONTHS ROADMAP

One architect. One active live pilot.

Example weekly allocation for the AI architect

2 days
Pair + build

1 day
Evaluate

1 day
Relationships

1 day
Decisions + buffer

Support is work.
Put it in the budget.

Proposed capacity envelope—not known company availability.

ELDORADO / 6 MONTHS ROADMAP

Measure better work—and skill health

Is the work improving?

Baseline → pilot · like-for-like cohorts

Validated task success (%)Independently accepted tasks ÷ all matured eligible starts.

Changes with escaped quality issues (%)Changes with confirmed defects ÷ accepted changes with completed follow-up.

Cost per accepted outcome (currency/outcome)Allocated cohort cost, including failures and review/rework ÷ independently accepted tasks.

Are skills fit and maintained?

Separate qualification + catalog populations

Skill selection precision (%)Correct activations ÷ all scored activations in labeled evaluation.

Skills within substantive review window (%)Owned, reviewed, not-overdue revisions ÷ all active approved revisions.

Independent quality qualification still required.

Choose relevant measures · Unknown is not zero
Skill health is not productivity proof.

2 minutes. Choose a small relevant subset from the sixteen optional contracts, not a requirement to collect all five shown here or all sixteen. These are candidate definitions, not actual company values or installed telemetry: task success, escaped quality, cost and precision have prototype calculations without network collectors; substantive-review coverage is unimplemented. Workflow only: compare like-for-like periods/cohorts with the same eligibility, task mix, observation and follow-up horizons, acceptance rubric and source coverage; record staffing, model/client and concurrent process changes. Before/after differences are not causal ROI. Validated task success is 100 times distinct independently accepted tasks by as_of divided by all eligible starts whose preregistered observation horizon has matured. Count each task once across retries, clients and skills. Failures, abandonment and unresolved matured starts remain in the denominator; unknown outcomes are not labeled failures. Exclude pre-start cancellations, sandbox/evaluation and out-of-scope tasks; flag invalid/missing starts and show unknown counts and coverage. Escaped quality is 100 times accepted changes with at least one linked confirmed post-acceptance defect during approved follow-up divided by accepted changes whose predeclared follow-up is complete. Count affected changes, not raw defects; separate severity and absolute defect/severe-incident counts. Show immature changes and pending triage separately; missing linkage or incomplete follow-up is not a clean result. Cost divides allocated costs for every task/attempt in the matured start cohort through as_of, including failed and abandoned attempts and shared period costs, by distinct independently accepted tasks in that same cohort. Include licenses, compute, tooling, setup, qualification, onboarding, support, upkeep and human quality review/rework effort at approved rates without double counting; exclude unrelated projects and documented accounting-basis exclusions. Use one approved reporting currency or dated approved conversion. Missing cost buckets mean a partial estimate with coverage, not total cost; zero accepted outcomes means null with full cost shown. Restate late invoices. Skill selection precision is 100 times true-positive activations divided by true-positive plus false-positive activations in candidate/client evaluation trials with independently labeled relevance and known activation. Freeze the corpus, retain negative/failed trials; unknown activation/relevance is unscorable, reported with coverage and blocks qualification if planned coverage is incomplete. This does not measure artifact success or live workflow acceptance. Substantive-review coverage is 100 times active approved revisions with current owner, linked substantive review evidence and review_due_at at or after as_of divided by all active approved revisions in the catalog snapshot. Experimental/retired revisions are excluded; report deprecated/revoked-in-use and ownerless counts separately. Missing owner/date/evidence stays outside the numerator. Review evidence includes regression, security and usefulness, not a rubber-stamped date. Independent Quality qualification still tests artifact fitness, safe execution and revision regression. Keep qualification trials and live workflow starts separate; never pool their denominators. No applicable skill population is N/A, not a baseline zero; missing evidence for an existing population is unknown, not good health. Predeclared live skill/version slices may reuse the workflow contract unchanged, without summing overlapping slices. Agree access, purpose and retention; link tracker, PR/CI, defect, cost and minimal effort evidence without individual rankings. If evidence is incomplete, limit or hold the claim; critical policy effects stop work regardless of averages.

Sources: measurement; six-months-roadmap; measurement/validated-task-success; measurement/changes-with-escaped-quality-issues; measurement/cost-per-accepted-outcome; measurement/skill-selection-precision; measurement/skills-within-substantive-review-window; skills-quality.

ELDORADO / 6 MONTHS ROADMAP

A dashboard can be confidently wrong.

Tempting signal
More PRs or generated lines
Shorter PR elapsed time
Higher acceptance rate
No defects recorded

Question before believing it
Useful work—or smaller slices?
Less effort—or less waiting?
Same task mix and denominator?
Enough follow-up and linkage?

AI usage is diagnostic.
It is not the score.

ELDORADO / 6 MONTHS ROADMAP

“It worked once” is not a release gate.

Pin candidate
and environment→Independent
qualification→Separate
pilot permission

Try to break it
Wrong task · injected instruction
Missing tests · recovery

First revision
No approved prior skill = N/A
Remove + reload → current practice

Failed mandatory safety checks → HOLD. An average cannot override them.

ELDORADO / 6 MONTHS ROADMAP

EXP-01 / Better acceptance is not enough.

Fictional teaching example · not Eldorado results or approval

Missing test evidence
→ checklist first→Incomplete follow-up
→ hold expansion→Higher review burden
→ stop candidate

Keep the useful practice.
Reject the unsupported benefit claim.

Existing AI practice is the comparator. A skill is tested only if needed.

2 minutes. Keep one example from baseline to decision. Suppose changes repeatedly lack regression evidence. Compare a repository test map plus review checklist against packaging the same guidance as a skill; the cheaper checklist wins if it performs as well. A skill carries reusable instructions and verification, not authority. Freeze a small non-urgent bug-fix scope, one willing team and one qualified client/model tuple. The full fictional packet includes explicit real-use blanks for people, dates, immutable references, sources and data approval. Those blanks are not permission. The apparent nine versus eight accepted is not a demonstrated gain. Baseline has two failed, one abandoned and one unresolved; pilot has one of each. Three later starts have not matured and stay out of the matured denominator. At first review pilot effort median is 35 minutes among observed rows only, not the whole cohort. One non-severe baseline defect is known; pilot defect follow-up is incomplete. A tiny sample, different mix and learning also limit inference. The fictional owner may authorize one two-week evidence-only extension with no new starts, at most four architect hours, two reviewer hours and EUR25 additional spend. If reconciliation cannot finish, stop rather than keep extending. Assume reconciliation finds a 40-minute median against 30 baseline for comparable work, with reviewers tracing extra effort to checking generated tests. This breaches the fictional no-increase burden guardrail even with nine accepted tasks. It is sufficient for a local stop decision, not statistical proof the intervention fails everywhere. Remove/reload and verify current-practice recovery; preserve evidence and useful test map/checklist. Return failure cases to the author. Retain would require different evidence: complete comparable outcomes and costs, acceptable quality/burden, agreed precision, independent acceptance and funded upkeep. A new revision must requalify. Critical unauthorized effects stop immediately, not at the scheduled review. The sample table and extension caps are speaker-reference detail, not the main strategy story. Use only the three visible decisions in the planned narration; expand into these numbers only if asked.

Sources: six-months-roadmap.

ELDORADO / 6 MONTHS TARGET

Month six / An evidence portfolio, not a victory lap.

Visible changeEvidence to ask for

One workflow understoodMap + baseline + explicit gaps

Retained practice—or justified stopComparable outcomes + costs + limits

Less dependence on the architectOwner + backup + recovery rehearsal

Proposed achievable picture. No guaranteed savings, headcount reduction or org-wide uplift.

ELDORADO / BEYOND 6 MONTHS

Beyond 6 Months / Progress toward the Target.

Deepen

Coaching, context or deterministic tooling where friction remains.

Transfer

Only with repeatable value, local fit and supported ownership.

Toward autonomy

Separate authority case, containment and human takeover.

Established destination. Evidence-led steps.
Separate authority—not automatic graduation.

2 minutes. The next quarter may deepen the existing workflow or stop expansion entirely. Skills remain optional; no catalog or central control platform is a prerequisite. Transfer needs new-scope baseline and local capacity. Consider bounded autonomy only if it solves an evidenced bottleneck and controls are proven: exact-artifact approvals, scoped identity, idempotency, audit, independent stop and takeover. People retain product, risk and release decisions. The long-term destination remains human-directed, guided and reviewed autonomous SDLC. These six months contribute toward that vision, not replace it; they do not promise its timing or authorize its execution. The CTO chooses investment based on the evidence, not an inevitable maturity staircase.

Sources: stage-02-autonomous-agent; target-operating-model; roadmap; six-months-target.

ELDORADO / BEYOND 6 MONTHS

Toward autonomy / Change authority deliberately.

Observe +
assemble evidence→Propose +
route work⇢Separately authorize
scoped execution

Exact-artifact approval · scoped identity · audit + recovery

Independent stop. Human takeover.

Readiness failure → hold or redesign. Missing evidence is never permission.

2 minutes. Progress toward the established Target by granting only the extra operation that an evidenced business case needs. Reliable observation and evidence assembly come before proposing and routing; neither permits mutation, merge or deployment. A separate human decision names the exact resources, data, artifact and scope. Prove identity separation, checks outside candidate control, idempotency, duplicate/outage recovery and audit. Test a kill switch and human takeover outside the model before trusting scoped execution. The agent cannot appoint approvers or edit its own gates. Changed boundaries reopen review; harmful behavior revokes authority and feeds learning back to Quality and Maintenance. This is the bridge toward the autonomous Target, not an installed system or a dated commitment.

Sources: stage-02-autonomous-agent; target-operating-model.

ELDORADO / DISCUSSION

The first CTO conversation

Which outcome is worth improving first?
Who can volunteer time—and independent review?
What data, risk and effort limits apply?
When will we decide to continue or stop?

Sponsor the learning.
Do not pre-approve the answer.

ELDORADO / DISCUSSION

Q&A / What would change your mind?

15 minutes for the hard questions.

No willing team?

Too little evidence?

A cheaper non-AI fix?

The proposed plan can change. The evidence standard should not.

ELDORADO / REFERENCE

References / Follow the evidence.

AI Architect Test Task · Eldorado.gg
Context, four questions, assumptions, 50-minute format
Long-term Target · Roadmap →
6 Months Target →
6 Months Roadmap →
EXP-01 packet and decision limits →
Measurement definitions and safeguards →
Beyond 6 Months →

Less Prompting,More Shipping

We are not starting from “no AI”.

A faster draft is not a shipped outcome.

Target / Autonomous SDLC.

People contribute throughout the cycle.

Roadmap / Build dependable work first.

Connected skills. A repeating learning loop.

6 Months Target / One credible contribution.

Influence is earned beside the work.

Months 1–2 / Earn the right to test.

Months 3–4 / Test, then decide.

Months 5–6 / Transfer only if earned.

One architect. One active live pilot.

Measure better work—and skill health

Is the work improving?

Are skills fit and maintained?

A dashboard can be confidently wrong.

“It worked once” is not a release gate.

EXP-01 / Better acceptance is not enough.

Month six / An evidence portfolio, not a victory lap.

Beyond 6 Months / Progress toward the Target.

Toward autonomy / Change authority deliberately.

The first CTO conversation

Q&A / What would change your mind?

References / Follow the evidence.

Less Prompting,
More Shipping